Changelog
What shipped, and when
Models added, prices changed, surfaces shipped. Only things you can see from outside — a model that appeared in the catalog, a rate that moved, a page that started existing. Current prices are always on the pricing table, and machine-readable at pricing.json.
Latest entry
Showing 49 entries · July 16 – November 6, 2026
November 2026
5 entries · 1 models · 4 pricing5 entries
Price cut: DeepSeek V4.1 Flash
Input falls from $0.125 to $0.0558, output from $0.55 to $0.2231, and cached input from $0.0125 to $0.0011 per million tokens. FP8 service and request compatibility remain available at the same customer price through fallbacks.
Price change: GLM 5.3 Flash
Input falls from $0.07 to $0.0284 and cached input from $0.03 to $0.0284 per million tokens. Output rises from $0.05 to $0.0923 per million tokens. Routing is unchanged.
Tencent Hy4 GA
Hy4 joins the catalog at $0.459 input / $1.375 output per million tokens, with cached input at $0.0231. Endpoint metadata lists 1,048,576 context tokens, 65,536 output tokens, images, reasoning, automatic tool selection and structured output. Forced named and required tool choices are unsupported; top_p is dropped. Hy4 Preview remains available. The catalog now contains 92 models.
MiMo-V2.5 pricing updated
MiMo-V2.5 is now $0.32 input / $1.60 output per million tokens, with cached input at $0.064: 20% below the current $0.40 / $2.00 list price.
MiMo-V2.6-Pro price updated to $0.426 / $0.861 per 1M tokens
MiMo-V2.6-Pro price is now $0.426 input / $0.861 output per million tokens (cache read $0.0036). Requests now go to the lowest-priced available OpenRouter provider.
2 more entries from November — Nov 5 to Nov 3
GPT-6.1 Terra and Terra Pro
OpenAI's GPT-6.1 Terra joins the catalog at OpenRouter list parity: $1 / $5 per million input / output tokens, cached input $0.10, cache write $1.25; the gateway's long-context tier applies above 200K prompt tokens at $2 / $7.50, cached input $0.20 (OpenAI's real threshold is 272K prompt tokens). Terra Pro is the same model in reasoning.mode=pro — same per-token price, far more reasoning tokens spent per request on complex tasks. Both support tools, JSON Schema / structured outputs, image input and reasoning; OpenRouter does not list temperature or top_p for either endpoint. Advertised context is 1,050,000 tokens, max output 128,000. The catalog now contains 91 models.
Bria FIBO 2.0
Bria FIBO 2.0 is available through the OpenAI Images API at $0.055 per image, matching FIBO 1.5, which remains available. Served through DeepInfra. Bria reports training on fully licensed data for commercially safe output. The catalog now contains 89 models.
Kimi K3.1 Pro and Highspeed
Two Moonshot routes at endpoint list parity: Kimi K3.1 Pro $2 / $10 and Highspeed $4 / $20 per million input / output tokens; both charge $0.20 for cached input. Both passed ZDR, optional reasoning, tools with auto or required selection, JSON Schema, images and streaming usage probes. Forced named functions returned 404. Advertised context is 1,048,576 tokens. The catalog now contains 88 models.
All Claude models now 75% below Anthropic list price
Every Claude model now costs 25% of Anthropic list price. Per million input / output tokens: Opus 5.5 $1.00 / $5.00, Sonnet 5.5 $0.50 / $2.50, Haiku 5.5 $0.025 / $0.125, and Fable 5.1 $2.50 / $12.50. Explicit cache prices also fall to 25% of list. Haiku 5.5 requests above 200,000 prompt tokens cost $0.125 input / $0.625 output per million tokens, with cached input at $0.0125.
DeepSeek V4.1 Flash 1017 price cut to $0.098 / $0.294 per 1M tokens (upstream parity)
deepseek-v4.1-flash-1017 now costs $0.098 per million input tokens, $0.294 per million output tokens and $0.00294 per million cached input tokens. These promotional rates match upstream pricing with no markup and no announced end date.
DeepSeek V4.1 Flash 1017
The October 17 snapshot is available as deepseek-v4.1-flash-1017 at base OpenRouter list parity: $0.15 / $0.60 per million input / output tokens, cached input $0.003. The pinned DeepInfra standard FP8 endpoint passed ZDR, tools, forced functions, strict JSON Schema, images and streaming with usage. Advertised limits are 1,048,576 context tokens and 131,072 output tokens. Reasoning is on by default and can be disabled per request. The existing deepseek-v4.1-flash route and price remain available separately. The catalog now contains 86 models.
Muse Spark Open
Meta’s open-weight Muse Spark Open is available as muse-spark-open at OpenRouter list parity: $1.25 / $4.25 per million input / output tokens, cached input $0.15. Modal and Together both passed zero data retention, tools, forced functions, strict JSON Schema, image input and streaming probes. The route lists a common 1,000,000-token context and 262,144-token output limit. Reasoning is mandatory; low effort is supported. Weights use the custom Muse Model License. The catalog now contains 85 models.
Step 5 Preview
StepFun's Step 5 Preview is available as step-5-preview at OpenRouter list parity: $1 / $2.70 per million input / output tokens, cached input $0.10. The pinned DeepInfra FP8 endpoint lists 1,048,576 context tokens and 65,536 output tokens. Live probes verified mandatory reasoning, tools, forced named functions, strict JSON Schema, image input and streaming with zero data retention. Reasoning cannot be disabled and consumes the output budget. The catalog now contains 84 models.
Qwen3.8 27B Decentered and cloud-only Qwen3.8 27B
qwen3.8-27b is now served only by cloud providers. Choose qwen3.8-27b-decentered for independent Share Compute community Macs at $0.238 input / $1.428 output per 1M tokens, 30% less. When no community machine can serve a request, cloud fallback is billed at qwen3.8-27b prices ($0.34 / $2.04, with its cache rates). Community operators see prompts and completions; see Privacy Policy Section 5. Contributor earnings follow the lower customer price.
Trio-Spark v1.3
MachineFi's System One model is available as trio-spark-v1.3 on POST /v1/systemone for typed choice, score and noul decisions with calibrated probabilities. Text input only on QSP, with a 1,024-token input limit; requests over it are rejected. $0.042 per 1M input tokens, output free, at list price. MachineFi reports 90.9% on a ten-task panel of 2,977 decisions and expected calibration error 0.019. The catalog now contains 82 models.
Nano Banana 2.1 and GPT Image 2.5; Nano Banana Pro moves to the Images API
Three image models on /v1/images/generations, priced per 1024x1024 image: gemini-nano-banana-2.1 (Google's newest Nano Banana) at $0.036, and gpt-image-2.5-flare (speed) and gpt-image-2.5-sunburst (precision) at $0.018, generated at medium quality with prompts up to 1,000 characters. Breaking change: gemini-3-pro-image (Nano Banana Pro) is now an image model at $0.114 per 1024x1024 image on /v1/images/generations, with n=1; calls to it on /v1/chat/completions return 400. Image model pages now have a playground for signed-in accounts. The catalog now contains 81 models.
Gemini 4 Argon
Google's new flagship reasoning model is available as gemini-4-argon the day Google opened the API. 1M context, text and image input, up to 65,536 output tokens, always-on reasoning, tools and strict JSON Schema output. Introductory price: $2 / $10 per 1M input / output tokens, cache read $0.10, at standard Google endpoint parity. Google announced $4 / $20 after the introductory window; the current price is not permanent. Reasoning text is returned when present and reasoning consumes the total output budget. Assistant prefill is unsupported. The catalog now contains 78 models.
Ten models at upstream parity; Kimi K3 rule updated
These ten models now track the upstream endpoint price selected by their route, with no markup or discount. Prices per 1M input/output tokens, with cache read: deepseek-v4-flash $0.086/$0.173 -> $0.06/$0.18 (cache $0.0086 -> $0.015); glm-5.2 $1.12/$3.52 -> $0.70/$2.20 (cache $0.208 -> $0.105); glm-5.3 $1.12/$3.52 -> $0.60/$2.20 (cache $0.208 -> $0.20); glm-5.3-flash $0.06/$0.2 -> $0.07/$0.05 (cache $0.012 -> $0.03); gpt-oss-120b $0.041/$0.187 -> $0.03/$0.15 (cache $0.041 -> $0.03); kimi-k2.6 $0.5472/$2.728 -> $0.45/$1.95 (cache $0.292 -> $0.045); kimi-k2.7-code $0.584/$2.8 -> $0.6712/$3.35 (cache $0.1278 -> $0.18); muse-glimmer-30b $0.28/$1.2 -> $0.30/$1.20 (cache $0.032 -> $0.04); nemotron-3-ultra $0.5/$2.2 -> $0.50/$2.20 (cache $0.1 -> $0.10) (unchanged); qwen3.6-35b $0.112/$0.8 -> $0.05/$0.70 (cache $0.04 -> $0.025). Kimi K3 keeps $2.55/$12.75 and $0.255 cache read; only its rule changes to openrouter_market x 1.0 because Makora flaps. Pinned provider order and quantization filters determine the reference; OpenRouter headline prices can differ. More expensive fallbacks keep the same customer price.
Speech to text and text to speech
Five speech models added to the catalog. OpenAI-compatible transcription and streaming speech endpoints, with whisper-1 and tts-1 compatibility aliases.
Qwen Image 3.0 replaces Qwen Image Max
Qwen Image Max was retired by Alibaba; qwen-image-3.0 replaces it at $0.04 per image. Requests that still use the old id keep working at the new price. Text-to-image supports 512x512 to 2560x2560, with optional prompt_extend (off by default).
Muse Wonder
Meta Muse Wonder is live as muse-wonder for agentic personal and professional work, with 1M context, up to 943,718 output tokens and text and image input. $4 / $20 per 1M input / output tokens, cache read $0.20, at OpenRouter list parity. Reasoning is mandatory and billed as output; its text is not returned, and small max_tokens budgets can leave no visible answer. Tools accept only tool_choice auto. Strict JSON Schema is supported. Meta reports, at max effort, SWE-Marathon 53.1%, DeepSWE 78.8% and AutomationBench-AA 39.1%.
GPT Image 2
OpenAI's GPT Image 2 is live as gpt-image-2 for image generation only via /v1/images/generations, at a flat $0.01 per image. The upstream currently ignores size and quality and returns roughly 1.5-megapixel images.
Claude Haiku 5.5
Anthropic's new small, fast model is live as claude-haiku-5-5 on its launch day, at half of Anthropic's list price like every Claude here: $0.05 / $0.25 per 1M input / output tokens (list $0.10 / $0.50), with 1M context, text and image input and tool calling. Prompts over 200K tokens are billed at $0.25 / $1.25.
Mistral Large 4
Mistral AI's new frontier multimodal model is live as mistral-large-4 on its launch day: text and image input, 524K context, tools, JSON Schema output and reasoning you can switch off per request. $0.68 / $2.09 per 1M input / output tokens, the same as its list price on OpenRouter.
Nemotron 3 Ultra
NVIDIA's largest open model is live as nemotron-3-ultra: a 550B-parameter Mixture-of-Experts with 55B active per token, built for deep reasoning and agent orchestration. $0.50 / $2.20 per 1M input / output tokens, the same as the lowest stable paid market rate, with 262K context, tools and JSON Schema output.
Claude prices cut in half
Every Claude model now costs half what it did, per 1M input / output tokens: Claude Opus 5.5 $1.60 / $8 (Anthropic list $4 / $20), Claude Sonnet 5.5 $1 / $5 (list $2 / $10), Claude Fable 5.1 and Fable 5 $4 / $20 (list $10 / $50), Claude Opus 5 and Opus 4.8 $2 / $10 (list $5 / $25), Claude Sonnet 5 $1 / $5 (list $3 / $15), and Claude Haiku 4.5 $0.40 / $2 (list $1 / $5). Cache-read and cache-write rates are halved too. Same models, same endpoint and same key; nothing to change in your code.
Forced tool_choice accepted on Claude 5.5 models
Claude Opus 5.5, Claude Sonnet 5.5 and Claude Fable 5.1 no longer reject a forced `tool_choice` (a named function or `"required"`). Anthropic does not accept one on these models, so QuickSilver Pro sends it as `"auto"` plus an instruction to call the tool, which steers the model toward the call but does not guarantee it. Requests that end with a prefilled assistant message are also accepted on these models and on Claude Opus 5, Sonnet 5, Fable 5 and Opus 4.8: the prefix is passed on as an instruction to continue from it.
Claude capability listings corrected
The Claude models no longer list JSON Schema structured output. `response_format: {type: "json_schema"}` is accepted but not enforced on any Claude model — the reply comes back as ordinary prose — so the capability has been removed from the catalog rather than left standing as a promise the API does not keep. A tool declared with `strict: true` does return schema-shaped arguments, and that is the supported way to get typed output from a Claude model. Separately, claude-opus-5-5 chooses its own tools: tool_choice takes "auto" or "none", and forcing one specific function is not supported on that model — a forced choice now returns an error that says exactly that.
Claude Opus 5.5 is live
Anthropic Claude Opus 5.5 joins the catalog — the first model of the Claude 5.5 family and the most capable Claude to date. It leads Claude Opus 5 and Claude Fable 5.1 on agentic coding (Terminal-Bench 4.0: 66.4%, versus 55.8% and 52.3%), knowledge work (GDPval-AA 1846) and computer use (OSWorld 2.0: 81.8%), and ranks #1 on the Artificial Analysis Intelligence Index v4.3 with a score of 58. It takes text and image input across a 1M-token context with up to 128K output tokens. On QuickSilver Pro it's $3.20 input / $16 output per 1M tokens, 20% below Anthropic's $4 / $20 list and 20% less per token than Opus 5. Point base_url at https://api.quicksilverpro.io/v1 and use model="claude-opus-5-5".
Grok 4.7 now at xAI's list price
Grok 4.7 moves to xAI's published list price, the same as Grok 4.6: $2 input / $6 output per 1M tokens and $0.50 cached input. A request whose prompt exceeds 200K tokens bills at $4 / $12 ($1 cached). Launch-day pricing was $1.60 / $4.80.
MiMo-V2.6-Pro and MiMo-V2.6-Flash are live
Xiaomi's MiMo-V2.6 generation joins the catalog. MiMo-V2.6-Pro is Xiaomi's flagship open-weights model (1.02T-parameter MoE, 42B active) and the top open-weights model on the Artificial Analysis Intelligence Index v4.3 with a score of 46, on AA's intelligence-versus-cost Pareto frontier; its current price is $0.426 input / $0.861 output per 1M tokens (cache read $0.0036; updated 2026-11-03). MiMo-V2.6-Flash is the fast, low-cost sibling (309B MoE, 15B active) at $0.112 / $0.224, the same price as MiMo-V2.5. MiMo-V2.6-Flash is 20% below Xiaomi's list price. Both take text and image input across a 1M-token context, and support tools and structured output. Reasoning is off by default; pass reasoning: { enabled: true } to opt in. Use model="mimo-v2.6-pro" or model="mimo-v2.6-flash".
Grok 4.7 is live
xAI Grok 4.7 joins the catalog — xAI's new flagship for coding, agentic tasks and knowledge work, succeeding Grok 4.6, built for long-running software engineering and for checking its own work. It accepts text and image input across a 500K-token context and always reasons (reasoning.enabled=false is rejected; set reasoning_effort to control depth). On QuickSilver Pro it's $1.60 input / $4.80 output per 1M tokens, xAI list parity and 20% below Grok 4.6; a request whose prompt exceeds 200K tokens bills at $3.20 / $9.60, xAI's long-context rate. Point base_url at https://api.quicksilverpro.io/v1 and use model="grok-4.7".
System One: Jev 1.13 typed decisions on /v1/systemone
A new kind of model joins the catalog. TypeSafe's Jev 1.13 is a System One model: it does not generate text — you POST a state and up to 64 typed questions (choice, score or noul) to https://api.quicksilverpro.io/v1/systemone and it answers every one in a single forward pass with calibrated probabilities, for routing, classification, triage, scoring and guardrail decisions. It bills input tokens only, at $0.042 per 1M, with output tokens free — TypeSafe's own list rate — and every response's usage block carries the call's USD cost. 32K-token context, same key and balance as the rest of the catalog; it is not reachable through Chat Completions. Docs: /docs/systemone.
Qwen3.8 Omni Flash
Qwen's first agentic omni model joins the catalog at $0.12 input / $0.376 output per 1M tokens, with cached input at $0.0128. It natively understands image, audio and video input and can drive tools across long workflows, with the family's full 1M-token context window and 131K max output. A single unified input rate covers every modality, and QuickSilver Pro sells it ~20% below Alibaba's own list price of $0.15 / $0.47 — omni capabilities at flash pricing. Point base_url at https://api.quicksilverpro.io/v1 and use model="qwen3.8-omni-flash".
GPT-6 Astra
OpenAI GPT-6 Astra is live — OpenAI's new flagship for long-horizon agentic coding, deep research and document work, with a 1M-token context, 128K max output and image input. On QuickSilver Pro it's $10 input / $50 output per 1M tokens, OpenAI list parity, on the same OpenAI-compatible endpoint as the rest of the catalog; a request whose prompt exceeds 200K tokens bills at $20 / $75, OpenAI's long-context rate. Point base_url at https://api.quicksilverpro.io/v1 and use model="gpt-6-astra".
Claude Fable 5.1
Anthropic Claude Fable 5.1 is live — the Mythos-class flagship, successor to Fable 5 for long-horizon agentic work and the most capable reasoning, with a 1M-token context and image input. On QuickSilver Pro it's $8 input / $40 output per 1M tokens, 20% below Anthropic's $10 / $50 list. Point base_url at https://api.quicksilverpro.io/v1 and use model="claude-fable-5-1".
Hy4 (preview)
Tencent Hy4 is live as a preview — the successor to Hy3 for coding and agentic workflows with reasoning over a 1M-token context. On QuickSilver Pro it's $0.6672 input / $2.0008 output per 1M tokens (cached input $0.0336), 20% below Tencent's $0.834 / $2.501 OpenRouter rate. Preview endpoint: availability and behavior may change. Point base_url at https://api.quicksilverpro.io/v1 and use model="hy4-preview".
Muse Spark 1.3
Meta Muse Spark 1.3 is live — Meta's flagship coding and agentic reasoning model, with stronger agentic and coding performance and better usability than Muse Spark 1.2 over a 1M-token context with multimodal input. On QuickSilver Pro it's $1.00 input / $3.40 output per 1M tokens (cached input $0.12), ~20% below Meta's $1.25 / $4.25 list price. Reasoning is always on; use reasoning_effort to trade depth against cost. Muse Spark 1.2 remains available at model="muse-spark-1.2". Point base_url at https://api.quicksilverpro.io/v1 and use model="muse-spark-1.3".
Kimi K3 re-priced to $2.55 / $12.75
Kimi K3 moves from $2.40 / $12.00 to $2.55 / $12.75 per 1M input/output tokens (cached input $0.255). It remains 15% below Moonshot's $3.00 / $15.00 list price; the adjustment aligns the sell price with the current cheapest production route.
Gemini 3.8 Flash
Google's most capable Flash model joins the catalog at $0.6375 input / $3.1875 output per 1M tokens, with cached input at $0.06375 — 15% below Google's promotional API price (through 2026-12-31). With high reasoning it scores 59 on the Artificial Analysis Intelligence Index, up 3 points from Gemini 3.7 Flash and on the intelligence-vs-cost Pareto frontier, with a 1M-token context and text + image input.
Qwen3.8 Flash Next
Alibaba's open-weight preview of the Qwen4 architecture joins the catalog at $0.12 input / $0.376 output per 1M tokens, with cached input at $0.0128. A 6B-active mixture-of-experts model with the full 1M-token context window and 131K max output, tuned for high-volume long-context agents — ~20% below Alibaba's own list price of $0.15 / $0.47.
Nemotron 3.5 Lightning
NVIDIA's open 30B-A3B efficiency model joins the catalog for fast, tool-heavy agents at $0.08 input / $0.20 output per 1M tokens, with cached input at $0.04 and a stable 262K context window.
Browser chat, no API key required
Every account now gets a free trial balance to spend at /chat — no card and no key. Four models are available on it, and each reply shows the exact cost of that turn at the same per-token rate the API charges.
Qwen3.8 Max
Alibaba's Qwen3.8 Max joins the catalog at $2.00 input / $6.00 output per 1M tokens, with implicit caching at $0.25, passed through at Alibaba's own published price.
Qwen3.7 Max repriced to Alibaba's own list
Now $1.25 input / $3.75 output per 1M tokens, down from $1.475 / $4.425. Alibaba is the only provider of this model, so we sell it at their published price rather than marking it up.
DeepSeek V4 Flash 0731, refreshed Gemini, audited open catalog
V4 Flash 0731 brings a 1M context window and native Responses API support. The Gemini line was refreshed and the open-weight catalog re-audited against each publisher's current list; retired models were removed rather than left to rot.
Claude Opus 5, and Google sign-in
Claude Opus 5 is available through the same endpoint and the same key as the open-weight catalog. Signing in with Google now works alongside email and password.
Kimi K3
Moonshot's K3 lands on release day at $2.40 input / $12.00 output per 1M tokens, 20% below Moonshot's own published rate.
MiniMax M3, MiMo-V2.5 and Hy3
Three additions to the open-weight side of the catalog, each with its reference list price published alongside ours.