AI Model Release Tracker

How this tracker stays honest

Every released model's per-token rate is read off the provider's own API pricing page, stored with that exact source URL, and dated to the day it was checked. A model reaches the upcoming table only when the provider has officially announced it or credible reporting has surfaced it, and that entry carries the reporting source. No release date is published unless a primary source states it, and no rumor appears without a citation. When a number cannot be confirmed against an official page, it is left off rather than guessed.

Released, preview, announced, rumored: what the labels mean

• Released: generally available, callable today, with a verified rate card.
• Preview: available to use but still rolling out or limited, so the rate may move.
• Announced: the provider has officially confirmed the model and, usually, a window, but it is not yet callable.
• Rumored: surfaced by credible reporting rather than the provider, with no confirmed date. Treated as a lead, attributed to its source, never as fact.

The sticker rate is not the bill

A low per-token price does not make a model cheap to run. What a coding task costs is set by how many tokens the agent burns reading context, reasoning, and generating output, and that swings by more than an order of magnitude between models and runs. A 2026 Microsoft Research preprint found the cheaper-listed model finished the same work at a higher cost in roughly a third of matchups, the price reversal phenomenon. The honest unit is cost to finish a representative slice of your own tasks: the cost-per-task calculator models it for every row in this tracker, and the Claude Code cost breakdown walks the math end to end.

Frequently asked questions

What new AI models launched in 2026?

As of November 2026, the models tracked here include Mistral's Large 4 (a 1.05-trillion-parameter model in public preview since October 6 at a launch-sale $0.68/$2.09 per Mtok, half its standard $1.36/$4.18, with open weights promised by the end of October), OpenAI's GPT-6 Astra (released September 3 at $10/$50 per Mtok, with a 2x surcharge on prompts over 272,000 tokens), GPT-6 Sol and GPT-6 Luna (released September 22 at $2/$10 and $0.10/$0.50, halving the GPT-5.6 tiers they replace) GPT-6.1 Sol (released September 29 at DevDay at the same $2/$10 with cached input halved to $0.10) and GPT-6.1 Terra (released October 26 at $1/$5 per Mtok, half the GPT-6.1 Sol rate card, replacing GPT-5.6 Terra); Anthropic's Claude Fable 5.2 (released November 9 at $8/$40 per Mtok, 20% under Fable 5.1, alongside the verification-gated Mythos 5.2), Claude Fable 5.1 (released September 1 alongside the invitation-only Mythos 5.1, holding Fable 5's $10/$50 per Mtok and cutting only the cache read, from $1.00 to $0.25), Claude Opus 5.5 (released September 22 at $4/$20 per Mtok, 20% under the Opus 5 it replaces, with the cache read cut 60% to $0.20), Claude Sonnet 5.5 (released September 28 at Sonnet 5's unchanged $2/$10 per Mtok, with its cache read halved to $0.10 on October 7), Claude Haiku 5.5 (released October 7 at $0.10/$0.50 per Mtok for prompts up to 100,000 tokens and $0.50/$2.50 above, 90% under the Haiku 4.5 it replaces), Claude Opus 5 (released July 24), Sonnet 5 and Haiku 4.5, with Claude Opus 4.1 retired on the Claude API on August 5; OpenAI's GPT-5.5 and the GPT-5.6 family (Sol, Terra and Luna, generally available July 9, with Terra and Luna repriced July 30 and Sol taking its own promotional cut on August 21, held through at least November 21); xAI's Grok 4.7 (released September 21 at an unchanged $2/$6 per Mtok) and Grok 4.6 (released August 12, 35 days after Grok 4.5); Meta's Muse Spark Open (released October 15 as open weights, 816B parameters with 42B active, served by third-party hosts at $1.25/$4.25 per Mtok), Muse Wonder (released October 12 as a Meta Model API preview at $4/$20 per Mtok, its first frontier-tier model) and Muse Wonder Full (released the same day inside Muse Code only, with no API price), Muse Spark 1.3 (released September 2, with max reasoning opened to all developers September 4), 1.2 and 1.1 plus Muse Glimmer (an Apache 2.0 open-weight model); Google's Gemini 4 Argon (released September 30 to cyber defenders in its Fairwind Program and generally available on the Gemini API since October 13, at an introductory $2/$10 per Mtok through November 30 that rises to $4/$20 on December 1), Gemini 3.7 Flash (released August 13 at an introductory rate that doubles on January 1, 2027), 3.6 Flash, 3.1 Pro, 3.5 Flash and 3 Flash; Thinking Machines' Inkling (released July 15, a 975B/41B-active MoE under Apache 2.0); and the leading China-built models DeepSeek V4 Pro and V4 Flash (both moving to peak and off-peak billing on August 16) and V4.1 Flash (whose October 17 build, DeepSeek-V4.1-Flash-1017, replaced it in place at unchanged rates), MiniMax-M3.1-Flash (released October 19 at $0.15/$0.60 per Mtok with open weights), Alibaba's Qwen3.7 Max, Qwen3.8-Max (released August 3, with open weights published August 13 under a custom licence rather than Apache 2.0) and the Qwen3.8-Max-0902 post-training snapshot (released September 2 at the same $2/$6 per Mtok), Moonshot's Kimi K3.1 Pro (released October 23 at $2/$10 per Mtok, a third below the Kimi K3 it replaces as flagship, with open weights promised by November 4), Kimi K3 (released July 16) and K2.7 Code, and Zhipu's GLM-5.2. Anthropic's Claude Fable 5 shipped June 9, 2026, was suspended June 12 under a US export-control directive, and returned in early July 2026 after the Commerce Department lifted the directive. Each row links to the provider's official page and is dated to when it was last verified.

What AI models are coming next?

Two entries. Google's Gemini 4 Carbon, announced October 28, 2026 as its strongest model, is available only to cyber defenders in Google's Fairwind Program, with no public date; Google lists Gemini 4 Argon's introductory $2 and $10 per million tokens, then $4 and $20, and Artificial Analysis ranked it first on its v5 index at launch and now scores it 60, second to Claude Fable 5.2 at 65. The other is Reflection AI's Beam, announced October 5, 2026, a 501-billion-parameter open-weight model (23 billion active) that is in early access by waitlist only, with no public API and no price. Reflection says its Apache 2.0 weights ship later in October. Google's last announced-but-unshipped entry before it, Gemini 4, shipped as Gemini 4 Argon on September 30, 2026, as a limited preview for cyber defenders in Google's Fairwind Program at an introductory $2 and $10 per million tokens, and opened to every paid Gemini API developer on October 13. The list was empty for a week because the labs cleared their pipelines in September: Claude Fable 5.1 on the 1st, Gemini 3.8 Flash, Muse Spark 1.3 and Qwen3.8-Max-0902 on the 2nd, GPT-6 Astra on the 3rd, Grok 4.7 on the 21st, Claude Opus 5.5 on the 22nd, Claude Sonnet 5.5 on the 28th, GPT-6.1 Sol on the 29th and Gemini 4 Argon on the 30th. Google's Gemini models page lists nothing unreleased (checked October 1, 2026), and no announced-but-unshipped model was found at OpenAI, Anthropic, Meta, xAI, DeepSeek, Alibaba, Moonshot or Zhipu on September 23, 2026; xAI's Grok 5 is widely reported but xAI's own site could not be read to confirm it, so it is not listed. A model appears here only once its lab announces it or credible reporting surfaces it, attributed to its source.

How much do new AI models cost to use?

Per-token API rates span two orders of magnitude, from GPT-6 Luna and Claude Haiku 5.5 (on prompts up to 100,000 tokens) at $0.10 per million input tokens to Claude Fable 5, Fable 5.1 and GPT-6 Astra at $10, with Claude Fable 5.2 at $8. DeepSeek held the bottom of that range until 16:00 UTC on August 16, 2026, when it moved to peak and off-peak billing: V4 Flash input went from $0.14 flat to $0.22 off-peak and $0.44 at peak, so the rates shown here for both DeepSeek rows are the peak figures. But the sticker rate is not the bill: what a task actually costs is set by how many tokens an agent burns, and the cheapest per-token model is often not the cheapest to finish the work. GPT-6 Astra is the clearest case on this table. It costs 2.5 times GPT-5.6 Sol per token and matches Fable 5.1 exactly on the sticker, yet on the independent Terminal-Bench 4.0 board it finishes a task for $17.02 against the $32.69 Fable 5.1 costs, because it burns 1.53 billion tokens where Fable 5.1 burns 2.75 billion. Fable 5.1 makes the same point from the other direction: it matches Fable 5 on every headline rate and still costs about 21% less on a long agentic run, because the only number it changed was the cache read. Claude Opus 5.5, released September 22, is the rarer case where the sticker itself fell: $4/$20 against the $5/$25 Opus 5 charged, with the cache read down 60% to $0.20. Model any row in the cost-per-task calculator to see the real figure.

Where does this AI model data come from?

Every per-token rate is read off the provider's own official API pricing page, stored with that source URL, and stamped with the date it was checked. Release and announcement dates are recorded only when a primary source states them. Nothing here is taken from memory or an unsourced tracker.

Sources

Per-token rates for the released models are grounded to each provider's official API pricing page, verified November 2026:

Spotted a rate that has changed, or a model that is missing? Submit it or send a correction. Featured and enhanced listings are available.

• Anthropic. (2026). Claude pricing (Fable 5.2, Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5, Haiku 5.5, Haiku 4.5). claude.com/pricing
• OpenAI. (2026). API pricing (GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol, Luna; GPT-5.6). openai.com/api/pricing
• Google. (2026). Gemini Developer API pricing (Gemini 4 Argon, Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash, 3.5 Flash, 3.1 Pro, 3 Flash). ai.google.dev/gemini-api/docs/pricing
• Google. (2026, September 30). Gemini 4 Argon: our next era of frontier intelligence. Company announcement, the primary source for Argon's introductory and standard rates and its Fairwind-first rollout. blog.google
• Google. (2026, October 13). Introducing Gemini 4 Argon. Company announcement of Argon's global availability on the Gemini API, enterprise and the Gemini app. blog.google
• DeepSeek. (2026). API pricing (DeepSeek V4 Pro, V4 Flash and V4.1 Flash, now the 1017 build). api-docs.deepseek.com/quick_start/pricing
• MiniMax. (2026). Pricing overview (MiniMax-M3.1-Flash). platform.minimax.io/docs/pricing/overview
• Alibaba Cloud. (2026). Model Studio model pricing (Qwen3.7 Max). alibabacloud.com/help/en/model-studio/models
• QwenCloud. (2026). Qwen3.8-Max-0902 model page and pricing (same $2/$6 rate as Qwen3.8-Max). qwencloud.com/models/qwen3.8-max-0902
• Moonshot AI. (2026). Kimi API pricing (Kimi K3.1 Pro, Kimi K3 and K2.7 Code). platform.kimi.ai/docs/pricing
• Meta. (2026). Meta Model API pricing and rate limits (Muse Wonder, Muse Spark 1.3 and its contributor tier). dev.meta.ai/docs/pricing-rate-limits
• Meta. (2026, October 15). Muse Spark Open. Meta research blog, the primary source for the open-weight release; host rate from OpenRouter's Muse Spark Open listing. research.meta.ai/blog/muse-spark-open
• Meta. (2026). Models (Muse Wonder Full availability: Muse Code Power and Maximum plans only). dev.meta.ai/docs/models
• Mistral AI. (2026). Mistral Large 4 model page (public preview, launch-sale and standard rates). docs.mistral.ai/models/mistral-large-4-0
• Google. (2026, October 28). Gemini 4 Carbon: our next frontier model. Company announcement, the primary source for Carbon's Fairwind-only access and its introductory and standard rates. blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon
• Reflection AI. (2026, October 5). Introducing Beam: Reflection's 501B open-weight model. Company announcement, the primary source for this upcoming entry. reflection.ai/blog/introducing-beam
• Zhipu AI (via OpenRouter). (2026). GLM-5.2 API pricing (Z.ai's own rate card still rolling out). openrouter.ai/z-ai/glm-5.2

For the cadence figures:

• Capital & Compute. (2026). AI model releases by month. The site's own dated release record, from which every cadence figure on this page is derived. capitalandcompute.net/ai-model-releases