The latest AI news and updates of OpenRouter — Unified API for accessing hundreds of LLMs across providers with intelligent routing and optimization. Covering OpenRouter's latest product updates, analysis, and launches from the past 90 days.
OpenRouterOct 24OpenRouter now hosts Moonshot AI's Kimi K3.1 Pro, a 1M-context multimodal model optimized for long-horizon coding and agentic tasks. The model features 2.8T total parameters and is available at $2/M input and $10/M output tokens. A 2–3× faster Highspeed variant is also available at $4/M input and $20/M output.
OpenRouterOct 23OpenRouter’s platform data identifies GLM 5.3 and Kimi K3 as the leading models for security tasks by spend share, capturing 27.8% and 10.7% of the category respectively. Grok 4.7 ranks as the top closed-weight model at 3.6% share. Security auditing accounts for 0.9% of total platform spend, with DeepSeek V4.1 Flash showing the category's largest growth.
OpenRouterOct 21OpenRouter reports that the median cache rate for agentic requests on its platform reached 85% in September 2026, up from 65% at the start of the year. Cached prompt tokens now account for nearly nine in ten tokens used by agents. Most inference providers on the platform serve agentic cache at 70% or higher.
OpenRouterOct 21OpenRouter reports that agentic token usage on its platform reached 132.6 trillion weekly tokens by October 2026, a 34x increase since February 2026. Agentic workloads now consume about eight times more tokens than human interactions, which totaled 17.0 trillion weekly. Agents first surpassed human token volume on the platform during the week of February 2, 2026.
OpenRouterOct 19OpenRouter added Inception's Mercury 3, a diffusion model generating 1,000+ tokens per second with a 1-million-token context window. Inception reports the model outperforms GPT-6 Luna and Claude Haiku 5.5 on IFBench, GPQA Diamond, and SWE-Bench Atlas. It is available via the OpenRouter API at a 50% launch discount of $0.10 per million input tokens and $0.35 per million output tokens.
OpenRouterOct 16OpenRouter now hosts Black Forest Labs' FLUX 3 Image Klein, an open-weights model optimized for single consumer GPUs. The platform also added an improved FLUX Video Upscale in early access for $0.28 per second and enabled 4K video generation from prompts with FLUX 3 Video at $0.41 per second.
OpenRouterOct 16OpenRouter launched Armor in beta, a safety classifier that uses TypeSafe AI's Jev model to block out-of-policy tool calls before execution. Armor evaluates calls against plain-English rules and conversation history, and it extends Claude Code's auto permissions mode to any model served on the platform. The plugin is available for opt-in use via the OpenRouter API.
OpenRouterOct 16OpenRouter added an Open vs Closed model share chart to its public rankings, revealing that open-weight models processed 73% of total tokens through its API so far this week. The new metric tracks weekly token volume, excluding stealth models, to provide a persistent view of usage trends across the platform.
OpenRouterOct 15OpenRouter launched Ori Harness, a tool that runs existing agent CLIs like Claude Code and Codex on any of its 500+ hosted models. The system uses OAuth for authentication, applies organization-wide guardrails, and consolidates billing across agents. A demonstration showed Claude Opus 5.5, Gemini 4 Argon, and GPT-6 Astra each rebuilding the visual engine for the game Royale.
OpenRouterOct 15OpenRouter now hosts Meta's Muse Spark Open, an open-weight model built for agentic coding workflows. The model features a 1-million-token context window and native support for text, image, video, audio, and PDF inputs. It is available via the OpenRouter API at $1.25 per million input tokens and $4.25 per million output tokens across four hosting providers.
OpenRouterOct 15OpenRouter launched Private Endpoints for Enterprise plans, allowing organizations to route fine-tuned models, dedicated deployments, and reserved capacity through its unified API. These endpoints remain visible only to the organization and support automatic fallback to shared capacity. Teams can now manage private infrastructure alongside public models using the same API slugs and routing preferences.
OpenRouterOct 15OpenRouter has added AssemblyAI's Universal-3.6 Pro speech-to-text model to its unified API. The model features 30% fewer errors on short conversational utterances, 12% fewer errors across supported languages, and improved formatting for entities like phone numbers and addresses. It adds 14 new languages and is available at $0.30 per hour.
OpenRouterOct 13OpenRouter now accepts image input for Inception's Mercury Decide decision model. Developers include images alongside text in the state of a Decisions request to receive structured outputs like choices, scores, or yes/no probabilities. Pricing remains unchanged at $0.02 per million input tokens with zero data retention, and images are billed as input tokens.
OpenRouterOct 13OpenRouter now hosts Google DeepMind's Gemini 4 Argon, a frontier reasoning model built for complex software engineering and multi-step agentic tasks. It leads on AutomationBench at 51.3% and LVBench at 91.7%, while supporting long-decode responses up to 1 million tokens. The model is available at introductory pricing of $2 per million input tokens and $10 per million output tokens.
OpenRouterOct 10OpenRouter has added Microsoft-Decision-1, a model post-trained from Qwen3.5-9B for structured decision-making tasks like routing and classification. The model is priced at $0.042 per million input tokens with free output, and Microsoft reports it achieves the highest accuracy across 36 blind benchmarks while operating 35 times faster than GPT-6 Sol.
OpenRouterOct 8OpenRouter now offers a paid Zero Data Retention endpoint for Inception AI’s Mercury Decide model. The endpoint is priced at $0.02 per million input tokens, a 50% discount from the list price, with free output and cached input tokens. Mercury Decide supports a 66,000-token context window and provides structured, probabilistic decision outputs for classification and routing tasks.
OpenRouterOct 8OpenRouter now hosts Grok Imagine Video 1.5 Lite, a distilled video generation model from xAI. The model supports text-to-video and image-to-video generation with first-frame steering. Pricing starts at 0.02 dollars per second for 480p output and 0.01 dollars per image input, with 1080p rendering available via 720p upscaling.
OpenRouterOct 8OpenRouter built Rasp, an AI sales agent on its Ori platform, to handle inbound triage, call prep, and CRM updates. Rasp saves the five-person sales team 600 hours monthly and costs about $30 a day, as Ori automatically routes tasks to cheaper models. The agent reduced deal cycles by 34% and increased close rates. Ori Slack agents launch soon.
OpenRouterOct 8OpenRouter reports that OpenAI's GPT-6 Luna Decisions is the fastest decision model on its platform, achieving 180ms latency on global requests. The model, which returns typed probabilistic answers for classification and routing tasks rather than generating text, outperforms Jev and Perplexity Decider in platform benchmarks.
OpenRouterOct 8OpenRouter now hosts Anthropic's Claude Haiku 5.5, the first Haiku model to support reasoning efforts. It features a 1-million-token context window and adaptive thinking enabled by default, with adjustable effort levels. The model is roughly 75% cheaper than its predecessor, delivers speeds exceeding 100 tokens per second, and costs $0.10 per million input and $0.50 per million output tokens.
OpenRouterOct 8OpenRouter now hosts ElevenLabs' voice models, providing access to nine text-to-speech and two speech-to-text models through its unified API. The integration supports over 90 languages, speaker labels, and word-level timestamps for voice agents and narration. All ElevenLabs models are available at a 50% discount on OpenRouter through October 19, 8am PT.
OpenRouterOct 8OpenRouter launched a Decision Model Rankings page tracking spend and token share across task types including Relevance, Correctness, and Instruction Following. The leaderboard uses real platform usage data to rank models, with TypeSafe AI’s Jev currently leading all categories. The dashboard provides granular insights into model performance and market share directly on the OpenRouter platform.
OpenRouterOct 6OpenRouter now hosts Google's Nano Banana 2.1, a Flash-tier image generation and editing model. It supports up to 14 reference images, Google Search grounding, and 2K or 4K panoramas in multiple aspect ratios. Pricing starts at $0.0336 per 1K image, scaling to $0.1134 for 4K outputs.
OpenRouterOct 6OpenRouter added Mistral AI's Mistral Large 4 in public preview. The multimodal model features 1T parameters (49B active), a 1-million-token context window, and up to 256K output tokens. It is available at a 50% discount for the first two weeks, priced at $0.68 per million input tokens and $2.09 per million output tokens.
OpenRouterOct 3OpenRouter launched a Model Router Benchmarks page comparing seven routers across six benchmarks. The Router Index scores performance by weighting quality at 60%, speed at 20%, and cost at 20%. The data shows that routers can introduce latency or increase costs through cache rebuilding, revealing when a single model outperforms a routing layer.
OpenRouterOct 2OpenRouter now hosts ByteDance Seed's Seedream 5.0 Flash, a low-latency model priced at a flat $0.018 per image. The model supports 1K or 2K output in aspect ratios from 21:9 to 9:21 and accepts up to 14 reference images for precise, single-prompt edits. Reference images are included at no additional cost.
OpenRouterOct 1OpenRouter added Liquid AI's D1, a structured decision model that processes app state and typed questions to return probabilistic answers. The model supports yes/no, choice, and score outputs with a 65,536-token context window. It is priced at $0.04 per million input tokens with no output-token charges and features zero data retention.
OpenRouterOct 1OpenRouter launched a Security Center to help teams identify and remediate risky API keys. The dashboard surfaces keys lacking spend limits or expiration dates, or those unused for 90+ days, and allows admins to disable, archive, or cap up to 500 keys at once. Enterprise plans also include IP allowlisting to restrict API access to approved network ranges.
OpenRouterSep 30OpenRouter added Voyage AI's rerank-3 and rerank-3-lite models to its unified API. These drop-in upgrades for long document and code retrieval support a 32K-token context length and instruction-following. Pricing is set at $0.05 per million tokens for rerank-3 and $0.02 per million tokens for rerank-3-lite.
OpenRouterSep 30OpenRouter launched a new /tools API that provides programmatic access to its Server Tools Marketplace. The endpoint returns tool configurations and real-time pricing, allowing agents to discover and load tools dynamically during execution. This update integrates tool discovery directly into agent workflows, removing the need for manual marketplace browsing.
OpenRouterSep 30OpenRouter now hosts OpenAI's GPT-6 Astra through its Ultrafast endpoint, delivering 147 tokens per second throughput. This premium tier is priced at $60 per million input tokens and $300 per million output tokens. The endpoint is available for opt-in use, providing a high-speed alternative to the standard Astra model routing.
OpenRouterSep 29OpenRouter released the openrouter-decisions skill to help coding agents correctly implement decision models via its Decisions API. The skill guides agents to use probabilistic outputs instead of text, ensuring consistent classification and routing. In OpenRouter's tests across three models, the skill increased decision-model usage from 4 of 27 runs to all 27, with blind preference in every comparison.
OpenRouterSep 28One week after launch, Claude Opus 5.5 has become the most used Anthropic model on OpenRouter by both share of spend and share of tokens. Platform data shows a rapid migration of request volume from the previous Opus 5 model to the new version.
OpenRouterSep 28OpenRouter has added Upstage's Solar Mini 4, a 35B mixture-of-experts model with a 524K context window, and Solar Decide, a structured decision model. Solar Mini 4 is priced at $0.05 per million input tokens and $0.20 per million output tokens, while Solar Decide costs $0.10 per million input tokens with no output-token charges. Both are 50% off.
OpenRouterSep 27OpenRouter reports that TypeSafe AI's Jev model now accounts for 27% of weekly classification request volume on its platform. This figure nearly doubles the share of DeepSeek V4 Flash, which previously held the top spot. The data tracks weekly request volume across the platform from June 8 through September 24, 2026.
OpenRouterSep 24OpenRouter launched a Server Tools Marketplace, providing server-side tools for agent grounding and execution. The suite includes web search with pinnable engines like Exa and Perplexity, a hosted Shell sandbox at 0.0001 dollars per second, and tool search with defer_loading. Most tools run server-side during the request, eliminating the need for client-side tool loops.
OpenRouterSep 24OpenRouter now hosts Recraft V4.1 Flash, a text-to-image model priced at $0.007 per image with 1.3-second generation speeds. The model supports five aspect ratios and generates up to six images per request, targeting high-volume creative iteration and prototyping. It maintains Recraft's aesthetic quality for photography, portraits, and logos without requiring image inputs or style references.
OpenRouterSep 23OpenRouter added Space Bunny Alpha, a free stealth model from an anonymous provider. It features fast inference, adjustable reasoning, and a 1-million-token context window with support for text, image, and video inputs. The provider does not use prompts or completions for training. The model is available via the OpenRouter API using the stealth/space-bunny-alpha slug.
OpenRouterSep 22OpenRouter has added OpenAI's GPT-6 Sol and Luna models to its unified API, both priced at half their GPT-5.6 predecessors. Sol, positioned below flagship Astra, offers improved factual reliability and coding performance, while Luna serves as the fast, low-cost tier. Both models feature a 1.1-million-token context window and support OpenAI's improved prompt caching.
OpenRouterSep 22OpenRouter now hosts Anthropic's Claude Opus 5.5, the first model in the Claude 5.5 family. It features a 1-million-token context window and is priced at $4 per million input tokens and $20 per million output tokens, a 20% reduction from Opus 5. The model leads previous versions in agentic coding, knowledge work, and computer use performance.
OpenRouterSep 22OpenRouter launched a Batch API that cuts per-token pricing in half on most models for requests returned within a 24-hour window. Across 230,000 beta batches, the service achieved a 7-minute median turnaround, with 90% of requests finishing within an hour. The API supports over 70 models and processes chat completions, responses, messages, and embeddings.
OpenRouterSep 20OpenRouter tested TypeSafe AI's Jev decision model against popular LLMs using its Ori Eval tool. Jev processed requests over 5x faster than the next fastest model while matching the accuracy of top classification LLMs. It also ranked as the second cheapest model in the benchmark, providing a high-performance option for structured, latency-sensitive decision tasks.
OpenRouterSep 18OpenRouter now hosts TypeSafe AI's Jev, a System One model that returns typed, probabilistic decisions instead of generated text. It processes app state and typed questions to provide structured answers with confidence scores. The model supports a 32,000-token context window and is available via the OpenRouter API for $0.042 per million input tokens.
OpenRouterSep 18OpenRouter revealed its stealth model Union Alpha as Pareto, a multimodal composite model from Unbiased. Community demand and detective work ended the stealth period early, and the model remains available for use. Pareto features a 262,144-token context window and supports tool calling, priced at $2.50 per million input tokens and $7.50 per million output tokens.
OpenRouterSep 16OpenRouter reports that monthly spending on models from three open-weight labs—Moonshot AI, Z.ai, and DeepSeek—has grown by 10x or more so far in 2026. While closed-source models maintain higher total spending, they have experienced comparatively slower growth rates. Data shows Moonshot AI led with a 2,425% increase, followed by Z.ai at 1,925% and DeepSeek at 1,000%.
OpenRouterSep 16OpenRouter added Union Alpha, a multimodal stealth model designed for research, coding, and agentic workflows. The model features a 262,144-token context window and supports tool calling. It is free to use, and the third-party provider does not use prompts or completions for training.
OpenRouterSep 15OpenRouter added US and Europe flags to its platform to identify providers that support geofencing requests. These indicators help verify that AI prompts and completions remain within designated geographic regions throughout the request lifecycle. This feature is available for Business and Enterprise plans to ensure data residency compliance.
OpenRouterSep 15OpenRouter reports that users spent more on OpenAI models than on Anthropic models last week. This shift marks the first time OpenAI has led in platform wallet share since the week of February 26, 2024. The data breaks a trend that had persisted for more than two and a half years.
OpenRouterSep 11OpenRouter reports DeepSeek V4.1 Flash processed one trillion tokens in its first 24 hours, on pace for 2.8 trillion in 48 hours. 90% of this volume consisted of cache reads priced at approximately $0.006 per million tokens. This rate is five times cheaper than comparable models like GLM-5.3 Flash.
OpenRouterSep 10OpenRouter launched a beta server tool, openrouter:shell, that lets any model run commands in a hosted Linux container. The accompanying Files API enables moving files into and out of the sandbox. Usage costs $0.0001 per active second, with a 30-second minimum for cold container starts, and supports both OpenAI and Anthropic tool specifications.
OpenRouterSep 9OpenRouter has made US In-Region Routing generally available on Business and Enterprise plans. Requests sent to the US endpoint are decrypted and processed entirely within the region, ensuring end-to-end data residency. This update allows teams to access Chinese open-weight models like DeepSeek V4 Pro and Kimi K3 through US and EU providers without data leaving the specified jurisdiction.
OpenRouterSep 8OpenRouter added Nex AGI's agentic N2.5-Pro and N2.5-Mini models to its unified API. Both models feature tool calling and adjustable reasoning effort for agentic coding tasks. Nex-N2.5-Pro supports text and image inputs for visual feedback loops, while Nex-N2.5-Mini provides a lighter, text-only alternative. Both models are available for free for a limited time.
OpenRouterSep 8OpenRouter launched Inception AI's Mercury 2.5, a diffusion-based reasoning model that generates tokens in parallel. It features a 260,000-token context window, tool calling, and structured output support. The model is available at $0.04 per million input tokens and $0.15 per million output tokens, ranking as the fastest model on the platform that runs on standard hardware.
OpenRouterSep 5OpenRouter added Microsoft's MAI-Image-2.6 and MAI-Image-2.6-Flash models to its unified API. Both support multi-image editing, web grounding, dynamic aspect ratios, and up to 1.5K output. The flagship model costs $5 per million tokens, while the Flash variant provides 2.8x faster generation than GPT-Image-2-Medium at $1.75 per million tokens.
OpenRouterSep 3OpenRouter has added Microsoft AI's MAI-Transcribe 2 to its unified API. The model supports 60 languages with automatic language identification, code switching, speaker diarization, and word-level timestamps. It is priced at $0.10 per hour and ranks #1 on the FLEURS multilingual benchmark.
OpenRouterSep 3OpenRouter launched a video benchmark suite that compares generation quality, cost, and speed across multiple AI video models. The platform provides side-by-side testing on curated prompts, including physics and prompt-adherence challenges. Users can generate custom prompts to evaluate model performance, with specific data showing Seedance 2.5 excels at physics while MiniMax H3 Max offers lower cost and faster generation.
OpenRouterSep 2OpenRouter has added the MiniMax H3 Max video-generation model to its unified API. The model generates 5–15 second clips from text or keyframe images in ~30 seconds for $0.40, significantly faster and cheaper than the previous H3 version. It supports 480p and 768p resolutions across six aspect ratios, with pricing starting at $0.05 per second.
OpenRouterSep 1OpenRouter has added Anthropic’s Claude Fable 5.1 to its unified API, offering a direct upgrade for Fable 5 workloads. The model improves performance in agentic coding, visual generation, and analysis tasks. It maintains existing input and output pricing while reducing prompt cache read costs by 75% to 0.25 dollars per million tokens.
OpenRouterSep 1OpenRouter integrated with Render Workflows to process LLM prompt batches as individual, retriable task runs. This architecture fans out prompts to avoid holding HTTP requests open, allowing Render to provision compute on-demand for each call. The integration supports Node and Python, using OpenRouter’s Auto Router to select models for each prompt while tracking model usage per task.
OpenRouterSep 1OpenRouter exclusively launched Inception AI's Mercury 2.5 Preview, a reasoning model that uses parallel token generation to reach 1,107 tokens per second. It features a 260,000-token context window, tunable reasoning, and parallel tool calling. The model is priced at $0.04 per million input tokens and $0.15 per million output tokens, targeting latency-sensitive production workloads like voice and coding agents.
Frequently asked questions