OpenRouter

DeepSeek V4.1 Flash 1017 is a sparse MoE with native vision and a 1M-token context, priced at $0.10 input and $0.50 output per million tokens. DeepSeek targets coding, terminal and computer-use agents and claims it beats V4 Pro.

A method for regression testing agents after prompt, model, tool or retrieval changes: lock the case set, write a behavioral contract with tool-call assertions and invariants, pin concrete model slugs, and read baseline results before candidate results.

OpenRouter added a paid zero-data-retention endpoint for Inception's Mercury Decide, alongside the existing free one. Listed pricing is $0.02/M input (half the $0.04 list price), free output and cached input, and 66K context.

OpenRouter released Ori Harness, which configures Codex, Claude Code, OpenCode and Hermes to run through OpenRouter with per-harness tuning. It routes Opus across eight endpoints for failover and sets a Claude Code option that OpenRouter says nearly halves the system prompt. The linked blog was not

A tutorial on testing agent tool calling in layers: exact-match checks where the expected tool is known, an LLM judge where several choices are valid, schema plus value checks on arguments, and trajectory comparison. Ends with running the same eval across models.

DeepSeek charges different rates by hour of day. OpenRouter now shows the peak and off-peak windows in your local timezone on the model page and factors the schedule into which endpoint it sources, with the V4 Flash 0731 endpoint currently showing a 64 percent discount.

Two reliability layers, often confused: provider failover recovers inside a single model automatically, while model fallbacks are opt-in and recover across models. Covers what triggers each, why the last entry in your list should be a dependable floor model, and where failover stops helping.

Breaks down why Kimi K3 is open-weight rather than open-source: Moonshot's custom license lacks OSI approval and adds a Model-as-a-Service revenue gate and UI attribution requirement above certain scale, despite granting broad use and resale rights.

OpenRouter compares Veo 3.1, Seedance, Kling and Grok Imagine Video for image-to-video work. It covers clip length, resolution, frame roles, audio and reference inputs, and shows how to submit a job through the unified video API.

OpenRouter breaks down Qwen 3.8's four variants: which weights are downloadable, their differing licenses (Apache 2.0, Qwen3.8-Max License, Qwen Community License), native versus extended context windows, and per-token pricing by provider.

Dedicated speech and transcription endpoints arrive behind an existing gateway key, with Whisper and GPT-4o Transcribe for speech to text and OpenAI, Google, and Mistral voices for synthesis, plus guidance on when a voice-native audio model is the wrong tool.

A server-side tool lets an orchestrating model delegate self-contained work to a smaller worker model mid-generation. The worker sees only the task description, never the parent conversation, and you pin which model handles it, with the price spread that makes the split worth it.

OpenRouter uses 30 days of real Claude Code sessions to give median per-session API costs by model and length. It attributes most cost to prompt-cache context and suggests Sonnet defaults, Haiku sub-agents and key limits.

OpenRouter benchmarked Exa, Parallel, Perplexity and native search across four models and four depths. Raising the budget from 1 to 25 turns roughly doubled BrowseComp scores. Model choice shifted results about 15 points, engine about 10, and incorrect runs averaged 19.7 searches versus 10.3.

OpenRouter released an openrouter-decisions skill for coding agents. In its tests across three models and four tasks, decision-model use rose from 4 of 27 runs to 27 of 27, and a grader preferred the skill output every time.

OpenRouter tests layered prompt injection defenses against four models: data labeling, key guardrails, and tool-call checking. Reports attack counts per defense, latency overhead, and a case where legitimate work got blocked.

Where should the routing layer run? A comparison of managed versus self-hosted LLM gateways on operational burden, data residency, RBAC and cost, with the spend arithmetic that identifies the point where running your own proxy becomes the cheaper option.

Live leaderboards grading web-search configurations for agents across BrowseComp, DeepSearchQA, WideSearch and HLE. The findings that survive the leaderboard churn: search budget dominates, the querying model outweighs the engine, and cost is governed by how often runs fail.

Grok 4.6 is routable on OpenRouter, with gains over Grok 4.5 claimed on GPDVal-AA V2 and pricing unchanged at $2 per million input tokens and $6 per million output.

An opt-in header that returns identical requests from an edge cache in 80 to 300ms with zero tokens billed, with configurable TTL up to 24 hours, explicit cache busting, hit and miss response headers, and streaming replay. Distinct from provider prompt caching.

OpenRouter presets move model choice, system prompt, provider preferences and sampling parameters to the server. Any app that accepts a model string can reference one with the preset field or model@preset syntax, and version history reverts a change without touching client code.

OpenRouter's Ori Eval, a model-comparison CLI, now reports failure causes and the serving provider, counts crashed runs as failures, estimates run cost with ori eval --pilot, warns on same-family judges, and supports headless login for CI and agent runs.

OpenRouter launches a Security Center that lists every API key across workspaces, flags risky ones with reasons, and applies bulk disable, archive or spend caps to up to 500 keys. The team found 1,000+ active keys for 85 employees in its own audit.

Gateway-level measurement of what a three-week frontier-model discount did to demand: token volumes multiplied several-fold against a flat undiscounted control, most of the share came from competing labs, and roughly a third of new users stayed once list pricing returned.