# OpenRouter

Latest posts and updates from OpenRouter.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## OpenRouter Now Powers Inference on Hugging Face

DevFeed: [OpenRouter Now Powers Inference on Hugging Face](<https://devfeed.tech/articles/openrouter-now-powers-inference-on-hugging-face-137752.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/announcements/hugging-face-inference-providers/>)

Author: Sasha Blumenfeld

Published: 2026-11-16T00:00:00Z

Content type: release

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [inference-providers](<https://devfeed.tech/topics/inference-providers.md>), [hugging face](<https://devfeed.tech/topics/hugging-face.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [model-serving](<https://devfeed.tech/topics/model-serving.md>)

Tags: [bucket](<https://devfeed.tech/tags/bucket.md>), [chat](<https://devfeed.tech/tags/chat.md>), [hugging-face](<https://devfeed.tech/tags/hugging-face.md>), [inference](<https://devfeed.tech/tags/inference.md>), [inference-providers](<https://devfeed.tech/tags/inference-providers.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [routing](<https://devfeed.tech/tags/routing.md>), [send](<https://devfeed.tech/tags/send.md>), [storage](<https://devfeed.tech/tags/storage.md>), [traces](<https://devfeed.tech/tags/traces.md>)

### AI overview

Hugging Face Inference Providers now use OpenRouter's routing for supported chat models. Routing selects providers using live price, speed, and reliability, with automatic failover and retries; existing Hugging Face tokens and code continue to work without setup. Users can prioritize with model ID suffixes, inspect provider prices, speed, and uptime, view spend by provider, and disable providers. A reported live experiment measured 46% fewer capacity errors, 16.9% faster time to first token, and 16.3% higher throughput, at 6.0% higher cost per request. Separately, OpenRouter can broadcast request traces as JSON files to a user-owned Hugging Face Storage Bucket (beta), including messages, output, token counts, and cost, but excluding images and other media. Setup requires a bucket and write-enabled S3 credentials.

### Source excerpt

Chat models on Hugging Face Inference Providers now run through OpenRouter's routing, and OpenRouter can send your traces to a Hugging Face Storage Bucket.

## Tighter Access Controls for Management Keys

DevFeed: [Tighter Access Controls for Management Keys](<https://devfeed.tech/articles/tighter-access-controls-for-management-keys-140687.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/announcements/scoped-management-keys/>)

Author: Sasha Blumenfeld

Published: 2026-11-16T00:00:00Z

Content type: release

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [least privilege](<https://devfeed.tech/topics/least-privilege.md>), [identity and access management](<https://devfeed.tech/topics/identity-and-access-management.md>), [Secrets Management](<https://devfeed.tech/topics/secrets-management.md>), [CRUD](<https://devfeed.tech/topics/crud.md>)

Tags: [access-controls](<https://devfeed.tech/tags/access-controls.md>), [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [analytics](<https://devfeed.tech/tags/analytics.md>), [api](<https://devfeed.tech/tags/api.md>), [api-keys](<https://devfeed.tech/tags/api-keys.md>), [audit](<https://devfeed.tech/tags/audit.md>), [automate](<https://devfeed.tech/tags/automate.md>), [automation](<https://devfeed.tech/tags/automation.md>), [bearer](<https://devfeed.tech/tags/bearer.md>), [budget](<https://devfeed.tech/tags/budget.md>), [cost](<https://devfeed.tech/tags/cost.md>), [dashboard](<https://devfeed.tech/tags/dashboard.md>), [delete](<https://devfeed.tech/tags/delete.md>), [deploy](<https://devfeed.tech/tags/deploy.md>), [error](<https://devfeed.tech/tags/error.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [keys](<https://devfeed.tech/tags/keys.md>), [management](<https://devfeed.tech/tags/management.md>), [workspace](<https://devfeed.tech/tags/workspace.md>)

### AI overview

Management keys can now be restricted by scope (one workspace or the organization) and access level (read-only or read/write). Workspace-scoped keys cannot access other workspaces; within their own workspace they can manage API keys, members, guardrails, provider keys, observability destinations, and presets, and read budgets, activity, and analytics. Budget changes, credits, organization settings, workspace creation or deletion, organization membership, and private endpoints still require an organization key. Read-only keys can use permitted read routes, including analytics, while create, update, and delete requests fail. Scope and access level are fixed at creation; changing them requires creating a replacement key and deleting the old one. Workspace Admins can create keys for their workspace, while Organization Admins can create workspace keys or organization-wide keys.

### Source excerpt

Make a management key read-only, limit it to one workspace, or both, so each key can only do the job it was created for.

## Reranking in RAG: How to Fix Bad Retrieval with a Reranker

DevFeed: [Reranking in RAG: How to Fix Bad Retrieval with a Reranker](<https://devfeed.tech/articles/reranking-in-rag-how-to-fix-bad-retrieval-with-a-reranker-131155.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/reranking-rag/>)

Author: OpenRouter

Published: 2026-11-13T00:00:00Z

Content type: article

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Retrieval Augmented Generation (RAG)](<https://devfeed.tech/topics/retrieval-augmented-generation-rag.md>), [information retrieval](<https://devfeed.tech/topics/information-retrieval.md>), [Embeddings](<https://devfeed.tech/topics/embeddings.md>), [hybrid-search](<https://devfeed.tech/topics/hybrid-search.md>), [voyage-ai](<https://devfeed.tech/topics/voyage-ai.md>)

Tags: [rag](<https://devfeed.tech/tags/rag.md>), [reranker](<https://devfeed.tech/tags/reranker.md>), [reranking](<https://devfeed.tech/tags/reranking.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>)

### AI overview

Reranking can move the correct passage above near-miss results by scoring each retrieved chunk together with the query. The guide recommends retrieving about 30 candidates, reranking them, and sending the top three to the LLM, then measuring precision@1 and end-to-end answer quality on your own data. In its 274-question synthetic benchmark, Gemini vector search put the exact answer first 81% of the time; adding voyageai/rerank-3-lite raised this to about 97% for roughly $0.10 per 1,000 queries. Voyage's dedicated rerankers were faster and cheaper than tested LLM rerankers, while openai/gpt-6-luna reached 99.6% precision@1 at about four seconds per call. Results may differ on real documents. Reranking cannot recover answers absent from its candidate set; candidate count trades cost for possible gains. Set embedding input_type correctly, and keep chunks within reranker context limits: cohere/rerank-v3.5 silently considers only the first 4,096 tokens.

### Source excerpt

When a RAG (Retrieval-Augmented Generation) app gives wrong answers even though the right information is in your documents, retrieval is often the cause. Vector search is great at finding chunks on the right topic, but it often falls for near misses, like pulling the bucket limit for the Enterprise plan when the user asked about the Growth plan. Reranking in RAG fixes this by taking your vector search's top 30 candidates and passing them through a dedicated cross-encoder model that re-scores each chunk directly against the question.

## How We Built a Self-Correcting Localization System for OpenRouter

DevFeed: [How We Built a Self-Correcting Localization System for OpenRouter](<https://devfeed.tech/articles/how-we-built-a-self-correcting-localization-system-for-openrouter-130215.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/announcements/openrouter-in-8-languages/>)

Author: Robert Sun

Published: 2026-11-12T00:00:00Z

Content type: article

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Localization (l10n)](<https://devfeed.tech/topics/localization.md>), [Internationalization (i18n)](<https://devfeed.tech/topics/i18n.md>), [human review](<https://devfeed.tech/topics/human-review.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [api](<https://devfeed.tech/tags/api.md>), [code](<https://devfeed.tech/tags/code.md>), [code-checks](<https://devfeed.tech/tags/code-checks.md>), [costs](<https://devfeed.tech/tags/costs.md>), [languages](<https://devfeed.tech/tags/languages.md>), [localization](<https://devfeed.tech/tags/localization.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [pipeline](<https://devfeed.tech/tags/pipeline.md>), [reports](<https://devfeed.tech/tags/reports.md>), [system](<https://devfeed.tech/tags/system.md>), [translation](<https://devfeed.tech/tags/translation.md>)

### AI overview

OpenRouter's localization pipeline keeps translations current by fingerprinting source strings and processing changed entries daily. It sends batches of up to 40 to GPT-6.1 Sol with style notes, glossary rules, and UI context; code checks placeholders, markup, whitespace, and terms, then a second call verifies meaning. Rejections return with reasons for retries; strings that still fail remain in English and are flagged. Daily pull requests receive AI review and a human merge. In a 454-case comparison, the full pipeline reduced authored check failures from 147 to 4 and outside judges counted 109 critical errors versus 203 for a bare prompt, though overall preference favored the pipeline narrowly (139 to 117), and back-translation scores were treated as a tie. The cheaper GPT-6 Luna model had lower first-try accuracy (91.7% versus 99.2%) at about one-third the cost on the 254-case comparison. Reader reports and reviewer-found defects become glossary rules, style notes, overrides, or regression cases. A full translation and verification of today's 16,918 strings into seven languages is estimated at about $460; daily costs vary with changed strings.

### Source excerpt

OpenRouter now speaks 8 languages. Here is the translation pipeline behind it, what it costs, and how reader reports become fixes.

## How to Benchmark LLMs

DevFeed: [How to Benchmark LLMs](<https://devfeed.tech/articles/how-to-benchmark-llms-129838.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/how-to-benchmark-llms/>)

Author: OpenRouter

Published: 2026-11-12T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Benchmark](<https://devfeed.tech/topics/benchmark.md>), [test](<https://devfeed.tech/topics/test.md>), [Data Quality](<https://devfeed.tech/topics/data-quality.md>), [reliability-engineering](<https://devfeed.tech/topics/reliability-engineering.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [compare](<https://devfeed.tech/tags/compare.md>), [cost](<https://devfeed.tech/tags/cost.md>), [gpt](<https://devfeed.tech/tags/gpt.md>), [gpt-6-1](<https://devfeed.tech/tags/gpt-6-1.md>), [gpt-6-1-sol](<https://devfeed.tech/tags/gpt-6-1-sol.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [llms](<https://devfeed.tech/tags/llms.md>), [models](<https://devfeed.tech/tags/models.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [synthetic](<https://devfeed.tech/tags/synthetic.md>)

### AI overview

Benchmark candidate models on representative tasks with independently computed answers and code that checks both correctness and output format. Validate the grader with boundary, malformed, and known-wrong cases; run identical tasks under fixed settings; track response coverage, verified cost, and request time; and retain repeats and unresolved attempts. In this 60-task synthetic study, GPT-6.1 Sol passed 180/180 trials at about $0.2449 per 100 responses, while DeepSeek had the lowest cost (about $0.00394) with 157/180 passes and 14 truncations. Sonnet's many failures included malformed outputs, so scores reflect contract compliance as well as correctness. Public ranking agreement was weak in this ten-model sample (Spearman 0.109). These results apply only to the tested tasks and settings; the text recommends held-out cases and more repeats for close rankings.

### Source excerpt

Choose models using evidence from your application's inputs and output formats. Run identical tasks through OpenRouter and compare graded accuracy, billed cost, and request time. On our 60 synthetic tasks, GPT-6.1 Sol passed 180/180 trials across three repeats under the tested contracts and settings.

## How to Stop an AI Agent Stuck in an Infinite Loop

DevFeed: [How to Stop an AI Agent Stuck in an Infinite Loop](<https://devfeed.tech/articles/how-to-stop-an-ai-agent-stuck-in-an-infinite-loop-129837.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/ai-agent-infinite-loop/>)

Author: OpenRouter

Published: 2026-11-12T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [function calling](<https://devfeed.tech/topics/function-calling.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-sdk](<https://devfeed.tech/tags/agent-sdk.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [api](<https://devfeed.tech/tags/api.md>), [api-key](<https://devfeed.tech/tags/api-key.md>), [arguments](<https://devfeed.tech/tags/arguments.md>), [billing](<https://devfeed.tech/tags/billing.md>), [budget](<https://devfeed.tech/tags/budget.md>), [bug](<https://devfeed.tech/tags/bug.md>), [cost](<https://devfeed.tech/tags/cost.md>), [costs](<https://devfeed.tech/tags/costs.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [infinite](<https://devfeed.tech/tags/infinite.md>), [model](<https://devfeed.tech/tags/model.md>), [retry](<https://devfeed.tech/tags/retry.md>), [stop](<https://devfeed.tech/tags/stop.md>), [tool](<https://devfeed.tech/tags/tool.md>)

### AI overview

The guide recommends four layers to stop runaway agent retries: enable doom-loop detection, normalize equivalent tool inputs with a loopKey, cap each run by steps and cost, and set an API-key credit limit. In tests across six models, detection reduced calls from 314 to 64 and spend from $0.29 to $0.05 across 18 runs, but exact-argument matching could be evaded by small input changes. Per-run caps cover novel retries; the key limit is enforced before provider routing and protects against bugs or many concurrent sessions. With BYOK, include_byok_in_limit must be enabled. A request whose worst-case input plus max_tokens exceeds the remaining limit fails without charge; in-flight requests can leave spend slightly above the cap. The guide also notes that step caps may allow pending tool calls and a final turn to run, so actual calls can exceed the configured step count.

### Source excerpt

An AI agent gets stuck in an infinite loop when a model keeps calling a tool that fails, usually with the same arguments, because nothing in the agent loop tells it to stop. Each retry resends the whole conversation, so every turn costs more than the last, and a loop left running overnight can cost thousands of dollars.

## Self-Hosted LLM vs API: When Does Self-Hosting Pay Off

DevFeed: [Self-Hosted LLM vs API: When Does Self-Hosting Pay Off](<https://devfeed.tech/articles/self-hosted-llm-vs-api-when-does-self-hosting-pay-off-129836.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/insights/self-hosted-llm-vs-api/>)

Author: OpenRouter

Published: 2026-11-12T00:00:00Z

Content type: article

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Self Hosted AI Tools](<https://devfeed.tech/topics/self-hosted-ai-tools.md>), [Self-hosted](<https://devfeed.tech/topics/self-hosted.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [vllm](<https://devfeed.tech/topics/vllm.md>), [API](<https://devfeed.tech/topics/api.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>)

Tags: [000](<https://devfeed.tech/tags/000.md>), [2](<https://devfeed.tech/tags/2.md>), [2026](<https://devfeed.tech/tags/2026.md>), [3](<https://devfeed.tech/tags/3.md>), [api](<https://devfeed.tech/tags/api.md>), [bill](<https://devfeed.tech/tags/bill.md>), [charged](<https://devfeed.tech/tags/charged.md>), [charges](<https://devfeed.tech/tags/charges.md>), [cost](<https://devfeed.tech/tags/cost.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [hosting](<https://devfeed.tech/tags/hosting.md>), [less](<https://devfeed.tech/tags/less.md>), [llm](<https://devfeed.tech/tags/llm.md>), [model](<https://devfeed.tech/tags/model.md>), [open](<https://devfeed.tech/tags/open.md>), [self-hosted](<https://devfeed.tech/tags/self-hosted.md>), [self-hosting](<https://devfeed.tech/tags/self-hosting.md>), [tokens](<https://devfeed.tech/tags/tokens.md>), [vs](<https://devfeed.tech/tags/vs.md>), [weight](<https://devfeed.tech/tags/weight.md>)

### AI overview

Self-hosting pays off when a rented GPU stays busy enough that its hourly cost is below the API cost for the same token mix. At the article's October 8, 2026 prices, Qwen3.8 27B on a $0.50/hour RTX 3090 breaks even at roughly 6-12% load, but delivers only about 15 tokens per second per user; DeepSeek V4.1 Flash on H200s never beats the cheapest ZDR endpoint, while two B200s require 76% load. These are lower-bound estimates based on published benchmarks and rents, excluding engineering and peak capacity; prices change. The article suggests keeping a GPU busy on steady traffic and sending peaks to the API, with ZDR filtering for overflow where policy allows. ZDR and EU/US routing can address some retention and residency needs, but requests still run on provider hardware. Requirements to control the deployment or keep data on hardware you operate call for private endpoints or self-hosting.

### Source excerpt

If you pay an API bill for an open-weight model, you may wonder whether your own GPU would cost less. GPU rent is charged every hour, busy or idle, while an API charges only for the tokens you use.

## MCP vs CLI vs Skills: What Each Costs in Context

DevFeed: [MCP vs CLI vs Skills: What Each Costs in Context](<https://devfeed.tech/articles/mcp-vs-cli-vs-skills-what-each-costs-in-context-129100.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/insights/mcp-vs-cli-vs-skills/>)

Author: OpenRouter

Published: 2026-11-11T00:00:00Z

Content type: comparison

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [context window](<https://devfeed.tech/topics/context-window.md>), [MSP MCP](<https://devfeed.tech/topics/msp-mcp.md>), [Agent Skills](<https://devfeed.tech/topics/agent-skills.md>), [Command-line interface](<https://devfeed.tech/topics/cli.md>), [Model Routing](<https://devfeed.tech/topics/model-routing.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-sdk](<https://devfeed.tech/tags/agent-sdk.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agent-context](<https://devfeed.tech/tags/ai-agent-context.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [anthropic-claude](<https://devfeed.tech/tags/anthropic-claude.md>), [anthropic-claude-haiku-5-5](<https://devfeed.tech/tags/anthropic-claude-haiku-5-5.md>), [api](<https://devfeed.tech/tags/api.md>), [caching](<https://devfeed.tech/tags/caching.md>), [cli](<https://devfeed.tech/tags/cli.md>), [context](<https://devfeed.tech/tags/context.md>), [context-window](<https://devfeed.tech/tags/context-window.md>), [costs](<https://devfeed.tech/tags/costs.md>), [mcp](<https://devfeed.tech/tags/mcp.md>), [mcp-server](<https://devfeed.tech/tags/mcp-server.md>), [model-context-protocol](<https://devfeed.tech/tags/model-context-protocol.md>), [skills](<https://devfeed.tech/tags/skills.md>), [tokens](<https://devfeed.tech/tags/tokens.md>), [tool](<https://devfeed.tech/tags/tool.md>)

### AI overview

In one GitHub task, the CLI used far fewer input tokens than the GitHub MCP server: on Sonnet 5.5, 1,421 versus 59,417 with 46 default tools; the skill used 3,315. Tool definitions and returned data both add cost: a full release object added about 20,400 tokens, while requesting only needed fields reduced issue results. Scoping MCP to two tools cut its first request, and a 4,000-character result cap reduced total task tokens, though that cap relied on fields appearing near the result's top. OpenRouter Tool Search reduced up-front definition costs to under 600 tokens, but whole-task results varied by model: it raised Haiku's cost and produced search retries or tool-call errors in some runs. The article recommends using a CLI when available, or scoped MCP with small results, and measuring on the intended model and provider. These are results from one server and one task.

### Source excerpt

Every tool you give an agent sends its definition to the model on every request, and you pay for it as input tokens. A Model Context Protocol (MCP) server can add tens of thousands of tokens to the context window. The usual alternatives are a command-line tool (CLI) or a skill.

## LLM Quantization: What fp8 and fp4 Do to Quality and Price

DevFeed: [LLM Quantization: What fp8 and fp4 Do to Quality and Price](<https://devfeed.tech/articles/llm-quantization-what-fp8-and-fp4-do-to-quality-and-price-129099.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/insights/llm-quantization/>)

Author: OpenRouter

Published: 2026-11-11T00:00:00Z

Content type: article

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [quantization](<https://devfeed.tech/topics/quantization.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>)

Tags: [api](<https://devfeed.tech/tags/api.md>), [bf16](<https://devfeed.tech/tags/bf16.md>), [bits](<https://devfeed.tech/tags/bits.md>), [fp4](<https://devfeed.tech/tags/fp4.md>), [fp8](<https://devfeed.tech/tags/fp8.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-quantization](<https://devfeed.tech/tags/llm-quantization.md>), [quality](<https://devfeed.tech/tags/quality.md>), [quantization](<https://devfeed.tech/tags/quantization.md>)

### AI overview

In an OpenRouter test of Qwen3.8 27B across six endpoints, scores on 240 exact-answer tasks ranged from 98.3% to 99.2% for fp16, fp8, and fp4, with overlapping confidence intervals; the test cannot rule out small gaps or generalize to other models and workloads. One fp4 endpoint scored lower on a separate 40-product set, but the difference was too uncertain to attribute to precision. Lower precision was often cheaper, though not consistently, and provider choice affected price and speed. To control precision, inspect endpoint offerings and set `provider.quantizations`; list every acceptable precision to give routing more fallback choices. Filters can fail if no endpoint matches, and pinning a sole endpoint can expose requests to its rate limits. The reported measurements are specific to October 20, 2026.

### Source excerpt

When you call an open-weight model through an API, several providers may serve it at different precisions, such as 16-bit bf16, 8-bit fp8 or 4-bit fp4. The precision can affect both the answers and the price. LLM quantization creates these versions by storing a model's weights in fewer bits, so the model needs less memory and hardware, at some cost in accuracy.

## LLM API Rate Limits: How to Fix OpenAI and Claude 429 Errors

DevFeed: [LLM API Rate Limits: How to Fix OpenAI and Claude 429 Errors](<https://devfeed.tech/articles/llm-api-rate-limits-how-to-fix-openai-and-claude-429-errors-129102.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/llm-api-rate-limits/>)

Author: OpenRouter

Published: 2026-11-11T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [rate-limiting](<https://devfeed.tech/topics/rate-limiting.md>), [Error Handling](<https://devfeed.tech/topics/error-handling.md>), [OpenAI](<https://devfeed.tech/topics/openai.md>), [Claude](<https://devfeed.tech/topics/claude.md>), [anthropic](<https://devfeed.tech/topics/anthropic.md>)

Tags: [anthropic](<https://devfeed.tech/tags/anthropic.md>), [api](<https://devfeed.tech/tags/api.md>), [batch](<https://devfeed.tech/tags/batch.md>), [claude](<https://devfeed.tech/tags/claude.md>), [error](<https://devfeed.tech/tags/error.md>), [errors](<https://devfeed.tech/tags/errors.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [limits](<https://devfeed.tech/tags/limits.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-api](<https://devfeed.tech/tags/llm-api.md>), [openai](<https://devfeed.tech/tags/openai.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>)

### AI overview

A 429 can come from a provider capacity limit, OpenRouter's shared free model cap, or a spend cap; identify the cause from limit_source and rate limit headers before retrying. For scheduled limits, wait for Retry-After or X-RateLimit-Reset, adding jitter to avoid a synchronized burst. In the reported 60 prompt batch, this completed every request in 89-92 seconds in Python and 106-218 seconds in TypeScript, while default SDK retries left 23 failed. When waiting is unsuitable, OpenRouter's ordered models fallback completed the batch in 13 seconds, routing eight requests to a paid model at $0.000066; this does not bypass the shared free tier cap. For provider capacity limits, OpenRouter can route to another provider serving the same model: with failover enabled, all 150 pinned-model requests succeeded across reported runs, while up to 32 failed when it was disabled. The article also recommends limiting concurrency, keeping max_tokens near expected output, checking daily quota, and using paid variants in production.

### Source excerpt

A 429 Too Many Requests error means you sent more requests or tokens than your limit allows in the current time window. OpenAI, Anthropic and OpenRouter all return it, and it usually shows up the first time an app runs under real load or a batch job fans out across several workers.

## How to Build a Voice AI Agent

DevFeed: [How to Build a Voice AI Agent](<https://devfeed.tech/articles/how-to-build-a-voice-ai-agent-129101.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/build-voice-ai-agent/>)

Author: OpenRouter

Published: 2026-11-10T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [Conversational AI](<https://devfeed.tech/topics/conversational-ai.md>), [function calling](<https://devfeed.tech/topics/function-calling.md>), [livekit](<https://devfeed.tech/topics/livekit.md>), [Tutorial](<https://devfeed.tech/topics/tutorial.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [build](<https://devfeed.tech/tags/build.md>), [call](<https://devfeed.tech/tags/call.md>), [generate](<https://devfeed.tech/tags/generate.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [tools](<https://devfeed.tech/tags/tools.md>), [voice](<https://devfeed.tech/tags/voice.md>), [voice-ai](<https://devfeed.tech/tags/voice-ai.md>)

### AI overview

The tutorial builds a Python restaurant booking voice agent using OpenRouter for transcription (ElevenLabs Scribe v2), tool-calling and replies (GPT-6 Luna), and speech generation (ElevenLabs Flash v2.5). It transcribes callers, checks availability and books through mock tools, then streams completed reply sentences to TTS; LiveKit Agents adds live audio, voice activity detection, playback, and interruption handling. In tests, server-side time to first audio had a 1.9-second median overall (1.4 seconds without tools, 2.4 with tools), excluding network and playback delays; the example's extra LLM requests can make it slower. TTS made up about 90% of measured model cost. Scribe v2 was chosen for its accuracy-speed balance, though recognizers often miss uncommon names, so callers should confirm key details. The tutorial stresses replacing mock tools with a backend that verifies successful actions, testing with representative speech, and measuring end-to-end latency. Separate speech-to-speech tests confirmed tool selection but did not measure full conversation speed or quality.

### Source excerpt

A voice agent needs to do more than generate a good answer. It must understand what someone said, decide what to do, call any needed tools, and start speaking again before the pause gets awkward.

## Introducing the Extractions API

DevFeed: [Introducing the Extractions API](<https://devfeed.tech/articles/introducing-the-extractions-api-129097.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/announcements/extractions-api/>)

Author: William Kofski

Published: 2026-11-10T00:00:00Z

Content type: release

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [API](<https://devfeed.tech/topics/api.md>), [data type](<https://devfeed.tech/topics/data-type.md>), [Image](<https://devfeed.tech/topics/image.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [alpha](<https://devfeed.tech/tags/alpha.md>), [api](<https://devfeed.tech/tags/api.md>), [api-key](<https://devfeed.tech/tags/api-key.md>), [apis](<https://devfeed.tech/tags/apis.md>), [array](<https://devfeed.tech/tags/array.md>), [asset](<https://devfeed.tech/tags/asset.md>), [assets](<https://devfeed.tech/tags/assets.md>), [cloudflare](<https://devfeed.tech/tags/cloudflare.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [fields](<https://devfeed.tech/tags/fields.md>), [images](<https://devfeed.tech/tags/images.md>), [launch](<https://devfeed.tech/tags/launch.md>), [models](<https://devfeed.tech/tags/models.md>), [name](<https://devfeed.tech/tags/name.md>), [numbers](<https://devfeed.tech/tags/numbers.md>), [request](<https://devfeed.tech/tags/request.md>), [spans](<https://devfeed.tech/tags/spans.md>), [text](<https://devfeed.tech/tags/text.md>), [values](<https://devfeed.tech/tags/values.md>)

### AI overview

The alpha Extractions API lets you name typed fields and receive values from text, images, or PDFs. Cloudflare's Clairic supports all field types and up to four PNG/JPEG images; it can return spans, null probabilities and estimate distributions. ScaleDown Extract handles one PDF or image per request, but supports only text and list fields and returns confidence without spans or null probabilities. Mercury Decide reads text and returns text and list fields. A receipt example shows Clairic flagging a misspelled merchant name with lower spelling confidence, while other fields scored highly; a PDF example shows Extract returning a lower-confidence total. Route low-confidence values for review, and treat estimates by their distributions. The API is in alpha, so its shape may change.

### Source excerpt

A new API for pulling typed fields out of documents and images: name the fields, get values, spans and confidence back. Three models serve it at launch; on Cloudflare's Clairic the same request reads text and images and returns numbers, dates, lists, estimates and forecasts.

## How to Get Structured Output from Any LLM: JSON Schema, Tool Calls, or a Decision Model

DevFeed: [How to Get Structured Output from Any LLM: JSON Schema, Tool Calls, or a Decision Model](<https://devfeed.tech/articles/how-to-get-structured-output-from-any-llm-json-schema-tool-calls-or-a-decision-model-129103.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/structured-output-llm/>)

Author: OpenRouter

Published: 2026-11-10T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [JSON Schema](<https://devfeed.tech/topics/json-schema.md>), [JSON](<https://devfeed.tech/topics/json.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Tool](<https://devfeed.tech/topics/tool.md>)

Tags: [decision](<https://devfeed.tech/tags/decision.md>), [decision-model](<https://devfeed.tech/tags/decision-model.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [json](<https://devfeed.tech/tags/json.md>), [json-schema](<https://devfeed.tech/tags/json-schema.md>), [llm](<https://devfeed.tech/tags/llm.md>), [model](<https://devfeed.tech/tags/model.md>), [schema](<https://devfeed.tech/tags/schema.md>), [structured](<https://devfeed.tech/tags/structured.md>), [structured-output](<https://devfeed.tech/tags/structured-output.md>), [tool](<https://devfeed.tech/tags/tool.md>)

### AI overview

For reliable structured LLM output, use a strict JSON Schema when an endpoint enforces it, force a single schema-backed tool call for tool-only models, or use a decision model when you need only labels. In the reported tests, 209 of 216 schema replies were valid, with all failures from one provider that ignored the schema; setting `require_parameters: true` makes unsupported requests fail instead of silently ignoring it. Forced tool calls can also fail validation: GLM 5.3 FlashX produced 22 valid outputs out of 24, and tool-choice options vary by provider. Keep a validator: Response Healing can repair JSON syntax but does not enforce schemas, fix tool arguments, streamed output, or truncated replies. Decision model Jev returned probabilities for fixed choices and matched chat models on the 24-message test at about half GPT-6 Luna's cost, but it does not extract IDs or items and was weak at arithmetic. Findings are limited to the tests described.

### Source excerpt

When your code reads fields from a model's reply, like an order ID or a refund amount, it expects clean JSON every time. Most of the time, it gets it. But now and then the model wraps the JSON in a code fence, adds a field you didn't ask for, or sends a number as text, and your code breaks. If you use several models, one fix won't cover them all, because each model and provider handles structured output in its own way.

## Introducing the New Auto and Auto Code Routers

DevFeed: [Introducing the New Auto and Auto Code Routers](<https://devfeed.tech/articles/introducing-the-new-auto-and-auto-code-routers-122068.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/announcements/auto-and-auto-code-routers/>)

Author: Brian Thomas

Published: 2026-11-06T00:00:00Z

Content type: release

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Model Routing](<https://devfeed.tech/topics/model-routing.md>), [ai-governance](<https://devfeed.tech/topics/ai-governance.md>)

Tags: [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cache](<https://devfeed.tech/tags/cache.md>), [code](<https://devfeed.tech/tags/code.md>), [coding](<https://devfeed.tech/tags/coding.md>), [effort](<https://devfeed.tech/tags/effort.md>), [foundation](<https://devfeed.tech/tags/foundation.md>), [generation](<https://devfeed.tech/tags/generation.md>), [governance](<https://devfeed.tech/tags/governance.md>), [health](<https://devfeed.tech/tags/health.md>), [live](<https://devfeed.tech/tags/live.md>), [model](<https://devfeed.tech/tags/model.md>), [next](<https://devfeed.tech/tags/next.md>), [policies](<https://devfeed.tech/tags/policies.md>), [pool](<https://devfeed.tech/tags/pool.md>), [pricing](<https://devfeed.tech/tags/pricing.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [request](<https://devfeed.tech/tags/request.md>), [router](<https://devfeed.tech/tags/router.md>), [routers](<https://devfeed.tech/tags/routers.md>), [state](<https://devfeed.tech/tags/state.md>)

### AI overview

The new Auto Router and coding-focused Auto Code select a model and reasoning effort from the models allowed by governance policies, weighing eleven request factors alongside live prices, provider health, benchmark performance, and cache state. Auto Code favors a higher tier of coding models, prioritizes latency, and responds more strongly to failures; Auto Router targets mixed traffic and cost efficiency. The router can adjust reasoning effort before switching models to preserve cached input, and session results inform later decisions. The article reports Auto Code benchmark scores in line with Claude Opus 5.5 at 74% lower cost per task, and says the new router was up to 4x better on benchmarks and 3x faster at p90 in a live routing test. Routers can trade some peak performance for consistent cost and quality; the article notes no router beats the best frontier model at every point on the cost-quality curve.

### Source excerpt

The next generation of Auto Router, rebuilt on a new technical foundation, plus Auto Code, a new router specialized for coding work. Both weigh eleven factors per request against live pricing, provider health, benchmarks, and cache state to pick a model and reasoning effort from the model pool your governance policies allow.

## Build an AI Code Reviewer with OpenRouter

DevFeed: [Build an AI Code Reviewer with OpenRouter](<https://devfeed.tech/articles/build-an-ai-code-reviewer-with-openrouter-110210.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/ai-code-review/>)

Author: OpenRouter

Published: 2026-11-01T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [ai code review](<https://devfeed.tech/topics/ai-code-review.md>), [Code review](<https://devfeed.tech/topics/code-review.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>), [Test-driven development](<https://devfeed.tech/topics/tdd.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-code](<https://devfeed.tech/tags/ai-code.md>), [ai-code-review](<https://devfeed.tech/tags/ai-code-review.md>), [api](<https://devfeed.tech/tags/api.md>), [apis](<https://devfeed.tech/tags/apis.md>), [bug](<https://devfeed.tech/tags/bug.md>), [build](<https://devfeed.tech/tags/build.md>), [code](<https://devfeed.tech/tags/code.md>), [costs](<https://devfeed.tech/tags/costs.md>), [files](<https://devfeed.tech/tags/files.md>), [git](<https://devfeed.tech/tags/git.md>), [guide](<https://devfeed.tech/tags/guide.md>), [linux](<https://devfeed.tech/tags/linux.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [patches](<https://devfeed.tech/tags/patches.md>), [python](<https://devfeed.tech/tags/python.md>), [responses](<https://devfeed.tech/tags/responses.md>), [review](<https://devfeed.tech/tags/review.md>), [sandbox](<https://devfeed.tech/tags/sandbox.md>), [uploads](<https://devfeed.tech/tags/uploads.md>)

### AI overview

The guide builds a local Python code reviewer that uses OpenRouter's Responses API to upload changed files, context files, and a Git diff to a hosted Linux sandbox. The model can run checks to confirm suspected bugs and propose V4A patches; findings are marked confirmed with command output or unconfirmed when they cannot be reproduced. Patches are applied to a temporary copy first, and reach the working tree only with `--apply`. The script requires Python 3.11+, Git, and an OpenRouter API key; the Files, Shell, and Apply Patch tools require the global endpoint. In three-run tests on a five-bug change, Claude Haiku 5.5 averaged $0.0153 per review, while Claude Sonnet 5.5 averaged $0.0447 and fixed all five bugs in every run. The sandbox adds cost over diff-only review, and the guide reports model misses and broken patches, so it recommends reading patches before applying them. Changed code is sent to OpenRouter and the provider; uploads are deleted when the script finishes. The proposed GitHub Actions workflow was not tested in GitHub Actions.

### Source excerpt

Code reviewers that only read a Git diff have to guess whether a change breaks anything, and their fixes come back as prose you apply by hand. This guide builds a Python reviewer on the OpenRouter Responses API that uploads your changed files into a Linux sandbox, runs checks to confirm each bug, and returns validated patches. A review takes one to four minutes and costs $0.015 to $0.045.

## How to Build Your Own Always-On AI Agent

DevFeed: [How to Build Your Own Always-On AI Agent](<https://devfeed.tech/articles/how-to-build-your-own-always-on-ai-agent-109823.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/build-always-on-ai-agent/>)

Author: OpenRouter

Published: 2026-10-31T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [human in the loop agent](<https://devfeed.tech/topics/human-in-the-loop-agent.md>), [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>), [Tutorial](<https://devfeed.tech/topics/tutorial.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-sdk](<https://devfeed.tech/tags/agent-sdk.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [always-on-ai](<https://devfeed.tech/tags/always-on-ai.md>), [build](<https://devfeed.tech/tags/build.md>), [hacker-news](<https://devfeed.tech/tags/hacker-news.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [schedule](<https://devfeed.tech/tags/schedule.md>), [tools](<https://devfeed.tech/tags/tools.md>), [tutorial](<https://devfeed.tech/tags/tutorial.md>), [typescript](<https://devfeed.tech/tags/typescript.md>)

### AI overview

The tutorial builds a scheduled Hacker News watcher in TypeScript with `@openrouter/agent` 0.11.0. It searches for recent stories matching chosen topics, filters previously seen IDs, writes relevant results to a digest, and can request approval before posting to a team channel. Persistent files store seen stories and resumable conversation state; a scheduler runs fresh checks while unfinished or approval-paused runs resume first. Each run is bounded by six steps or $0.05, with an API key credit limit recommended to cap overall spend. In the reported Claude Haiku 5.5 runs, checks cost $0.0003-$0.0012; the SDK may make a call or two beyond a stop threshold before its closing response. Approval pauses all scheduled checks until a decision is recorded, and a failure after writing a digest can cause duplicate entries. The guide also describes running the script under a process manager or scheduled job with persistent storage.

### Source excerpt

An always-on AI agent wakes up on a schedule, checks something with tools, and reports or acts without you asking. This tutorial builds one in TypeScript with the OpenRouter Agent SDK that watches Hacker News, remembers what it has seen across restarts, stops any run at five cents, and waits for your approval before it posts.

## Open-Weight vs Closed-Weight Models: What's the Difference?

DevFeed: [Open-Weight vs Closed-Weight Models: What's the Difference?](<https://devfeed.tech/articles/open-weight-vs-closed-weight-models-what-s-the-difference-109822.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/insights/open-weight-vs-closed-weight-models/>)

Author: OpenRouter

Published: 2026-10-31T00:00:00Z

Content type: comparison

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [artificial-analysis](<https://devfeed.tech/tags/artificial-analysis.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [competition](<https://devfeed.tech/tags/competition.md>), [cost](<https://devfeed.tech/tags/cost.md>), [data](<https://devfeed.tech/tags/data.md>), [difference](<https://devfeed.tech/tags/difference.md>), [models](<https://devfeed.tech/tags/models.md>), [open](<https://devfeed.tech/tags/open.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [providers](<https://devfeed.tech/tags/providers.md>), [quality](<https://devfeed.tech/tags/quality.md>), [run](<https://devfeed.tech/tags/run.md>), [serve](<https://devfeed.tech/tags/serve.md>)

### AI overview

Open-weight models account for a growing share of OpenRouter usage and had roughly one-quarter the token-weighted list prices of closed-weight models in the cited snapshot, largely because multiple providers compete to serve them. The strongest closed-weight models still scored about 11 points higher on the Artificial Analysis Intelligence Index; the article cautions that scores depend on reasoning effort and may not predict differences for a particular workload. It recommends testing an open-weight model on your tasks and using a closed-weight model where its added quality changes the result. Open weights allow self-hosting and fine-tuning, but licenses vary and most models do not publish training data. Hosted privacy depends on the provider's data policy. The excerpt also describes combining models through one API with price caps and fallbacks for routing around cost limits, outages, rate limits, and other errors. Usage and price figures are snapshots with classification and measurement limitations, including excluded unclassified traffic and provider-specific tokenizers.

### Source excerpt

An open-weight model publishes its trained weights, so anyone can run, fine-tune or serve it. A closed-weight model is only available from its maker. We used a year of OpenRouter usage data, our price catalog and benchmark scores to show what that difference changes for cost, providers and quality.

## Guardrails Now Separate Organization Limits from Workspace Defaults

DevFeed: [Guardrails Now Separate Organization Limits from Workspace Defaults](<https://devfeed.tech/articles/guardrails-now-separate-organization-limits-from-workspace-defaults-113595.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/announcements/guardrail-policies/>)

Author: Sasha Blumenfeld

Published: 2026-10-30T00:00:00Z

Content type: release

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [internal developer platform automation](<https://devfeed.tech/topics/internal-developer-platform-automation.md>), [cloud cost management](<https://devfeed.tech/topics/cloud-cost-management.md>), [Single sign-on (SSO)](<https://devfeed.tech/topics/sso.md>)

Tags: [admin](<https://devfeed.tech/tags/admin.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-budgets](<https://devfeed.tech/tags/ai-budgets.md>), [api](<https://devfeed.tech/tags/api.md>), [api-key](<https://devfeed.tech/tags/api-key.md>), [api-keys](<https://devfeed.tech/tags/api-keys.md>), [budget](<https://devfeed.tech/tags/budget.md>), [coding-agents](<https://devfeed.tech/tags/coding-agents.md>), [config](<https://devfeed.tech/tags/config.md>), [content-moderation](<https://devfeed.tech/tags/content-moderation.md>), [costs](<https://devfeed.tech/tags/costs.md>), [guardrails](<https://devfeed.tech/tags/guardrails.md>), [limits](<https://devfeed.tech/tags/limits.md>), [model](<https://devfeed.tech/tags/model.md>), [model-access](<https://devfeed.tech/tags/model-access.md>), [organization](<https://devfeed.tech/tags/organization.md>), [policy](<https://devfeed.tech/tags/policy.md>), [separate](<https://devfeed.tech/tags/separate.md>)

### AI overview

OpenRouter Guardrails now separates organization-wide limits from workspace defaults. Organization admins can cap per-person budgets and restrict models, providers, ZDR, and data regions; workspace admins set member defaults and assign individual budgets or model access within those limits. Budgets reset daily, weekly, or monthly and apply separately per member or API key, not as a shared pool. Budget, model/provider access, and content moderation are independent policy types, so one person's budget can change without altering shared access or moderation rules. Existing guardrails retain their IDs and enforcement, and the existing API and Terraform resources continue to work. Admins can transition workspace by workspace; dashboard changes do not sync back to API or Terraform configuration.

### Source excerpt

Admins can now set organization-wide limits, give people different budgets and model access within them, and change each policy on its own. Existing guardrails keep working.

## Claude Opus vs Sonnet vs Haiku: Which Should You Use?

DevFeed: [Claude Opus vs Sonnet vs Haiku: Which Should You Use?](<https://devfeed.tech/articles/claude-opus-vs-sonnet-vs-haiku-which-should-you-use-107658.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/insights/claude-opus-vs-sonnet-vs-haiku/>)

Author: OpenRouter

Published: 2026-10-30T00:00:00Z

Content type: comparison

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Claude](<https://devfeed.tech/topics/claude.md>), [Claude Code](<https://devfeed.tech/topics/claude-code.md>), [AI-assisted coding](<https://devfeed.tech/topics/ai-assisted-coding.md>), [agentic-coding](<https://devfeed.tech/topics/agentic-coding.md>), [Python](<https://devfeed.tech/topics/python.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [bug](<https://devfeed.tech/tags/bug.md>), [bug-fix](<https://devfeed.tech/tags/bug-fix.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [claude-sonnet](<https://devfeed.tech/tags/claude-sonnet.md>), [claude-sonnet-5-5](<https://devfeed.tech/tags/claude-sonnet-5-5.md>), [code](<https://devfeed.tech/tags/code.md>), [coding](<https://devfeed.tech/tags/coding.md>), [coding-tasks](<https://devfeed.tech/tags/coding-tasks.md>), [cost](<https://devfeed.tech/tags/cost.md>), [haiku](<https://devfeed.tech/tags/haiku.md>), [opus](<https://devfeed.tech/tags/opus.md>), [opus-5](<https://devfeed.tech/tags/opus-5.md>), [opus-5-5](<https://devfeed.tech/tags/opus-5-5.md>), [tasks](<https://devfeed.tech/tags/tasks.md>), [tests](<https://devfeed.tech/tags/tests.md>), [volume](<https://devfeed.tech/tags/volume.md>), [vs](<https://devfeed.tech/tags/vs.md>)

### AI overview

In the article's tests, Opus 5.5, Sonnet 5.5, and Haiku 5.5 all passed 11 hidden-test coding tasks and three Claude Code bug-fix runs; their main difference on those tasks was cost. Sonnet matched or exceeded Opus on the reported GPQA Diamond and τ-bench airline scores at lower cost, leading the authors to recommend it as a default. Haiku passed the coding tests at much lower cost and is suggested for routine, high-volume, or sub-agent work; prompts under 100,000 tokens retain its lower listed price. Opus is reserved for long, ambiguous tasks where mistakes are costly, an advantage these short tests cannot assess. The authors recommend evaluating models on one's own workload. They also report that Claude Code's cost field overstated Haiku spend through OpenRouter, advising users to check OpenRouter Activity instead.

### Source excerpt

Use Claude Sonnet 5.5 as your default, Haiku 5.5 for high-volume and sub-agent work, and Opus 5.5 for long tasks where a mistake is costly. We compared the three on price, benchmarks, 11 coding tasks with hidden tests and a Claude Code bug fix. All three passed everything, so the difference was cost.

## How Much Does Claude Code Cost? Real Session Costs by Model

DevFeed: [How Much Does Claude Code Cost? Real Session Costs by Model](<https://devfeed.tech/articles/how-much-does-claude-code-cost-real-session-costs-by-model-107657.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/insights/claude-code-cost/>)

Author: OpenRouter

Published: 2026-10-30T00:00:00Z

Content type: article

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Claude Code](<https://devfeed.tech/topics/claude-code.md>), [Claude](<https://devfeed.tech/topics/claude.md>), [Model Routing](<https://devfeed.tech/topics/model-routing.md>), [cost-optimization](<https://devfeed.tech/topics/cost-optimization.md>), [AI-assisted coding](<https://devfeed.tech/topics/ai-assisted-coding.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [alternatives](<https://devfeed.tech/tags/alternatives.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [anthropic-claude-haiku-5-5](<https://devfeed.tech/tags/anthropic-claude-haiku-5-5.md>), [api](<https://devfeed.tech/tags/api.md>), [api-key](<https://devfeed.tech/tags/api-key.md>), [app](<https://devfeed.tech/tags/app.md>), [bill](<https://devfeed.tech/tags/bill.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [cost](<https://devfeed.tech/tags/cost.md>), [costs](<https://devfeed.tech/tags/costs.md>), [free](<https://devfeed.tech/tags/free.md>), [install](<https://devfeed.tech/tags/install.md>), [median](<https://devfeed.tech/tags/median.md>), [model](<https://devfeed.tech/tags/model.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [request](<https://devfeed.tech/tags/request.md>), [sessions](<https://devfeed.tech/tags/sessions.md>)

### AI overview

Claude Code is free to install; API-key users pay for the model it calls. OpenRouter's medians from real sessions in the 30 days to October 18, 2026, put a 2-9-turn session at $0.31 on Opus 5.5 and $0.16 on Sonnet 5.5; longer sessions cost more, and Sonnet is about half Opus at each length. The first request is often costliest because it writes the system prompt, tool definitions, and project context to the prompt cache; later turns read that cached context at a lower rate. In one empty-folder test, 17,047 cache-write tokens cost $0.085, while real projects can add context. The post recommends Sonnet as the default, Haiku for background jobs and sub-agents, shorter sessions or `/compact`, and API-key credit limits. It cautions that Claude Code's `total_cost_usd` may misprice OpenRouter models; the Activity page reflects actual charges. Other providers' models can be used through OpenRouter, but should be tested on your own tasks because Claude Code is built for Anthropic models.

### Source excerpt

Claude Code is free to install, and what you pay for is the model it calls. We publish the median cost of real Claude Code sessions on OpenRouter by model and session length. This post shows those costs, why the first request costs the most, what a month of use comes to, and how to lower the bill.

## How to Prevent Prompt Injection: 13 Attacks, Tested

DevFeed: [How to Prevent Prompt Injection: 13 Attacks, Tested](<https://devfeed.tech/articles/how-to-prevent-prompt-injection-13-attacks-tested-105749.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/prevent-prompt-injection/>)

Author: OpenRouter

Published: 2026-10-29T00:00:00Z

Content type: article

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Cybersecurity](<https://devfeed.tech/topics/cybersecurity.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>), [Prompt Engineering](<https://devfeed.tech/topics/prompt-engineering.md>), [human review](<https://devfeed.tech/topics/human-review.md>)

Tags: [attacks](<https://devfeed.tech/tags/attacks.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [prevent](<https://devfeed.tech/tags/prevent.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>)

### AI overview

The tests suggest that prompt injection risk centers on tool-using agents, and that labeling content as untrusted data reduces attacks but does not reliably stop them. Across four models, labels alone still allowed 13 attack emails in 720 attempts; OpenRouter Armor, which checks each tool call against rules before it reaches the app, blocked all 73 attack emails the models attempted. It added about 0.3 seconds at the median for tool-using requests. Armor allowed emails when users named the recipient, but blocked 55 of 80 cases where users approved an instruction from a document without naming its target. The authors recommend combining labels with model allowlists and spending caps, rules for each external action, and keeping secrets out of prompts. They also caution that blocked-call details can contain secrets, and advise auditing verdicts before enforcing rules.

### Source excerpt

If your app has a model read documents, emails or web pages, anyone who writes in them can try to give it instructions. We ran 25 attacks against four current models with each defense on and off. Labels cut attack emails, but only checking every tool call with Armor stopped all of them.

## Hit Your Claude Code Usage Limit? How to Keep Coding

DevFeed: [Hit Your Claude Code Usage Limit? How to Keep Coding](<https://devfeed.tech/articles/hit-your-claude-code-usage-limit-how-to-keep-coding-105747.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/claude-code-usage-limit/>)

Author: OpenRouter

Published: 2026-10-28T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Claude Code](<https://devfeed.tech/topics/claude-code.md>), [API Business Models](<https://devfeed.tech/topics/api-business-models.md>), [coding assistant](<https://devfeed.tech/topics/coding-assistant.md>), [Test-driven development](<https://devfeed.tech/topics/tdd.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [api-key](<https://devfeed.tech/tags/api-key.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [cost](<https://devfeed.tech/tags/cost.md>), [guide](<https://devfeed.tech/tags/guide.md>), [models](<https://devfeed.tech/tags/models.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [plan](<https://devfeed.tech/tags/plan.md>), [token](<https://devfeed.tech/tags/token.md>), [weekly](<https://devfeed.tech/tags/weekly.md>), [window](<https://devfeed.tech/tags/window.md>)

### AI overview

When Claude Code hits a plan's five-hour or weekly limit, the excerpt explains how to continue through OpenRouter, where usage is billed per token and governed by credit and spending limits. The quickest route is Ori Harness, which uses browser sign-in; the manual setup uses three environment variables in a separate `claude-or` shell function so plain `claude` remains on the plan. The excerpt recommends checking `limit_remaining` or the Activity page for actual charges: Claude Code's displayed cost can be inaccurate for OpenRouter runs. Its reported tests found median short-session costs of $0.16 on Sonnet 5.5 and $0.31 on Opus 5.5, while a Haiku 5.5 bug-fix run cost $0.008. It also advises Anthropic as the first-choice provider because Claude Code is only guaranteed to work with Anthropic; other models should be tried on a small task first.

### Source excerpt

When Claude Code hits your plan's five-hour or weekly limit, you can keep working by pointing it at an OpenRouter API key. You pay per token for the same Claude models, with no usage window. This guide sets it up as a separate command, checks the cost, and switches back when your plan resets.

## Claude 529 Overloaded: What It Means and How to Fix It

DevFeed: [Claude 529 Overloaded: What It Means and How to Fix It](<https://devfeed.tech/articles/claude-529-overloaded-what-it-means-and-how-to-fix-it-105746.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/tutorials/claude-529-overloaded/>)

Author: OpenRouter

Published: 2026-10-28T00:00:00Z

Content type: tutorial

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Claude](<https://devfeed.tech/topics/claude.md>), [Error Handling](<https://devfeed.tech/topics/error-handling.md>), [Claude Code](<https://devfeed.tech/topics/claude-code.md>), [anthropic](<https://devfeed.tech/topics/anthropic.md>), [rate-limiting](<https://devfeed.tech/topics/rate-limiting.md>)

Tags: [anthropic](<https://devfeed.tech/tags/anthropic.md>), [api](<https://devfeed.tech/tags/api.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [error](<https://devfeed.tech/tags/error.md>), [fallback](<https://devfeed.tech/tags/fallback.md>), [guide](<https://devfeed.tech/tags/guide.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [model](<https://devfeed.tech/tags/model.md>), [openrouter](<https://devfeed.tech/tags/openrouter.md>), [retry](<https://devfeed.tech/tags/retry.md>)

### AI overview

A Claude 529 `overloaded_error` indicates Anthropic temporarily lacks capacity for the requested model across users; the article distinguishes this from a 429 caused by an account's own rate limit or spend cap. It recommends exponential backoff with randomness for brief spikes, noting Anthropic's SDKs and Claude Code already retry transient errors. For longer overloads, OpenRouter can route the same Claude model across providers, retry failed requests, and report which provider served each response. Custom provider ordering changes default load balancing, while `allow_fallbacks: false` and restrictive `only` settings can prevent failover. A `models` list can specify another model if all providers for the first fail, though fallback triggers on any error and may change capability. In Claude Code, users can switch models or route sessions through OpenRouter; increasing retry counts may prolong waits. The excerpt reports a successful nine-session trial, but does not establish broader reliability.

### Source excerpt

A 529 overloaded_error from Claude means Anthropic's API is out of capacity for the model you asked for, not that your key or plan is wrong. This guide shows how to retry with backoff, send the same Claude model to another provider through OpenRouter, add a fallback model, and stop the error in Claude Code.

## How we rebuilt provider routing for better, faster inference

DevFeed: [How we rebuilt provider routing for better, faster inference](<https://devfeed.tech/articles/how-we-rebuilt-provider-routing-for-better-faster-inference-105745.md>)

Original publisher: [Read original article](<https://openrouter.ai/blog/announcements/how-we-rebuilt-provider-routing/>)

Author: Scott Morse

Published: 2026-10-28T00:00:00Z

Content type: release

Language: en

Sources: [OpenRouter](<https://devfeed.tech/sources/openrouter.md>)

Topics: [Model Routing](<https://devfeed.tech/topics/model-routing.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [LLM observability](<https://devfeed.tech/topics/llm-observability.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [Cache](<https://devfeed.tech/topics/cache.md>)

Tags: [algorithm](<https://devfeed.tech/tags/algorithm.md>), [cost](<https://devfeed.tech/tags/cost.md>), [performance](<https://devfeed.tech/tags/performance.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [routing](<https://devfeed.tech/tags/routing.md>)

### AI overview

OpenRouter's new default router scores each provider endpoint using effective price, throughput, time to first token, rate limits, reliability errors, and response quality, then shifts traffic as endpoint performance changes. It shares traffic among competitive providers, maintains a small test flow to each, backs off failing or rate-limited endpoints, and caps load based on capacity. Conversations stay pinned to the selected provider while its cache is warm; failed requests can retry elsewhere. `provider.order` bypasses the algorithm, explicit `provider.sort` uses the chosen sort, and `provider.only` or `provider.ignore` filter candidates before algorithmic selection. In a one-week switchback experiment covering over 24 million requests, the article reports 16.9% better TTFT, 16.3% better throughput, 67% fewer 429 capacity errors, and 6.0% higher cost per request. Price-only routing remains available through `provider.sort: "price"`, `:floor`, or the account default.

### Source excerpt

Our default routing used to send most traffic to the cheapest provider for a model. The new routing algorithm constantly monitors the performance of every provider endpoint to optimize the balance of cost, speed, and reliability for each new request.

[Next page](<https://devfeed.tech/sources/openrouter.md?cursor=WyIyMDI2LTEwLTI4VDAwOjAwOjAwKzAwOjAwIiwgIjE3ZjQ5ZmQ3LTk4NTgtNDRkYy1hN2U2LTQyYzRmOWFhNDAzOCJd>)