Model releases
Notable AI model releases with their access terms, context windows, and list prices. Specifications come from the developers unless marked otherwise.
| Model | Released | Access | Context (tokens) | Input $/M | Output $/M | Notes | Sources |
|---|---|---|---|---|---|---|---|
| Nemotron 3.5 Super VLNVIDIA | Open weights | — | — | — | Open 120-billion-parameter multimodal model with 12 billion parameters active per token, released with smaller Bolt models for local hardware; NVIDIA did not publish benchmark scores with the announcement, and we could not confirm the license text.Figures: company | 1 2 | |
| Muse WonderMeta | APIProprietary | 1,000,000 | $4 | $20 | Meta Superintelligence Labs' first frontier model; a stronger Muse Wonder Full is limited to Muse Code subscribers and not sold through the API, and the benchmark comparisons in Meta's model card are company-reported.Figures: company | 1 2 3 | |
| Claude Haiku 5.5Anthropic | APIProprietary | — | $0.1 | $0.5 | Anthropic's small, low-cost model for high-volume work such as summarization and subagents; prices shown are the starting rates Anthropic lists, and Anthropic says it costs about 75% less to run than Haiku 4.5 on average.Figures: company | 1 2 | |
| Mistral Large 4Mistral AI | API | 1,000,000 | $1.36 | $4.18 | Released as a public preview with weights promised later; Mistral lists these as regular prices with a half-price launch sale, gives about 1 trillion total parameters (its docs and an NVIDIA announcement differ on active parameters), and its benchmark figures are company-reported.Figures: company | 1 2 3 | |
| Gemini 4 ArgonGoogle DeepMind | APIProprietary | — | $4 | $20 | Announced on 30 September for vetted cyber defenders only and made generally available to paid tiers on 13 October; the prices shown are Google's standard rates from 1 December 2026, after an introductory $2 input and $10 output, and Google's benchmark scores (such as 68% on CWE-bench v1) are company-reported.Figures: company | 1 2 3 4 | |
| GPT-6.1 SolOpenAI | APIProprietary | — | $2 | $10 | Released at DevDay as an upgrade to GPT-6 Sol; OpenAI says it nearly matches GPT-6 Astra on agentic coding and computer use (company-reported) and rates it Critical for cybersecurity under its Preparedness Framework.Figures: company | 1 2 3 | |
| Claude Sonnet 5.5Anthropic | APIProprietary | 1,000,000 | $2 | $10 | Positioned as a faster, cheaper companion to Opus 5.5; Anthropic says it outperforms Opus 5.5 on agentic coding (company-reported) and halved its cache-read price on 7 October.Figures: company | 1 2 | |
| Claude Opus 5.5Anthropic | APIProprietary | 1,000,000 | $4 | $20 | First model in Anthropic's Claude 5.5 family, priced 20% below Opus 5; Anthropic says it performs close to Claude Fable 5.1 on most work, a company-reported comparison.Figures: company | 1 2 3 4 | |
| GPT-6 SolOpenAI | APIProprietary | — | $2 | $10 | Mid-tier GPT-6 model released in the API alongside GPT-6 Luna about three weeks after GPT-6 Astra; succeeded by GPT-6.1 Sol on 29 September.Figures: company | 1 2 | |
| GPT-6 LunaOpenAI | APIProprietary | — | $0.1 | $0.5 | Lowest-cost GPT-6 model, released in the API with GPT-6 Sol; OpenAI also lists Batch and Fast tiers at different rates.Figures: company | 1 2 | |
| Grok 4.7xAI | APIProprietary | 500,000 | $2 | $6 | xAI's coding and knowledge-work model, with text and image input and text output; prompts over 200,000 tokens are billed at a higher rate, and xAI's release notes carry an earlier 17 September entry than the 21 September announcement.Figures: company | 1 2 3 | |
| GPT-6 AstraOpenAI | APIProprietary | — | $10 | $50 | OpenAI's flagship, rolled out first to a limited set of organizations and the first OpenAI model it classifies as Critical for cybersecurity; its benchmark scores, such as 98% on FrontierMath Tier 4, are company-reported.Figures: company | 1 2 3 | |
| Gemini 3.8 FlashGoogle DeepMind | APIProprietary | — | $1.5 | $7.5 | Successor to Gemini 3.7 Flash, released with a cybersecurity variant; prices shown are Google's standard rates from 1 January 2027, after an introductory $0.75 input and $3.75 output through 31 December 2026.Figures: company | 1 2 | |
| Claude Fable 5.1Anthropic | APIProprietary | 1,000,000 | $10 | $50 | Anthropic's most capable generally available model, released with Claude Mythos 5.1, the same model with fewer safeguards for trusted-access programs only; Anthropic reports 52.6% on Terminal-Bench-Science (company-reported).Figures: company | 1 2 3 | |
| DeepSeek-V4-ProDeepSeek | Open weightsMIT | 1,000,000 | — | — | General-availability release of the V4 model previewed on 24 April, a 1.6-trillion-parameter mixture-of-experts model with 49 billion active; prices are omitted because DeepSeek's pricing page uses peak and off-peak rates and we could not match its columns to models.Figures: company | 1 2 3 | |
| Qwen3.8-MaxAlibaba (Qwen) | Open weightsCustom Qwen license with commercial restrictions | 1,000,000 | — | — | A 2.4-trillion-parameter mixture-of-experts model with 95 billion active, first offered through Alibaba Cloud's API and released as open weights about two weeks later; companies above $50 million in revenue need a separate commercial license, and its Text Arena ranking is as reported by Alibaba.Figures: company | 1 2 3 | |
| GLM-5.3Zhipu AI (Z.ai) | Open weightsGLM-5.3 License | 1,000,000 | — | — | Release month only (mid-August 2026; we could not confirm the day from Z.ai); weights followed about two weeks after API launch under a license that requires a security review for hosts with more than $10 billion in revenue, and its CyberGym result is company-reported.Figures: company | 1 2 3 | |
| Claude Opus 5Anthropic | APIProprietary | 1,000,000 | $5 | $25 | Priced the same as Opus 4.8; Anthropic said it was state of the art on coding and knowledge-work evaluations (company-reported), and it is now listed as a legacy model after Opus 5.5.Figures: company | 1 2 3 4 | |
| Gemini 3.6 FlashGoogle DeepMind | APIProprietary | — | $1.5 | $7.5 | Released with Gemini 3.5 Flash-Lite and 3.5 Flash Cyber while the expected Gemini 3.5 Pro did not appear; Google now describes it as a previous-generation model.Figures: company | 1 2 | |
| Kimi K3Moonshot AI | Open weightsKimi K3 License | 1,000,000 | $3 | $15 | A 2.8-trillion-parameter model with native vision, which Moonshot calls the largest open-weight model so far; Artificial Analysis independently scored it 57 on its Intelligence Index, and its license requires a paid arrangement for services earning over $20 million a year.Figures: company | 1 2 3 | |
| GPT-5.6 SolOpenAI | APIProprietary | 1,050,000 | — | — | Most capable of the three GPT-5.6 variants (Luna, Terra, Sol), first shipped on 26 June to trusted partners after a US government safety review and released publicly on 9 July; prices are omitted because OpenAI cut them from $5 input and $30 output to $4 and $20 on 21 August under a promotion.Figures: company | 1 2 3 | |
| MAI-Thinking-1Microsoft | APIProprietary | 256,000 | — | — | Microsoft's first in-house reasoning model, shown in private preview at Build on 2 June and in public preview on Microsoft Foundry from 12 August; Microsoft has not published per-token prices.Figures: company | 1 2 3 | |
| Gemini 3.5 FlashGoogle DeepMind | APIProprietary | — | — | — | First Gemini 3.5 model, launched at Google I/O and made the default in the Gemini app; Google's benchmark claims (such as 76.2% on Terminal-Bench 2.1) are company-reported, prices are omitted because sources conflict, and the API model is now deprecated in favor of 3.6 Flash.Figures: company | 1 2 3 | |
| Muse SparkMeta | APIProprietary | — | — | — | First model from Meta Superintelligence Labs, succeeding the Llama family and, unlike Llama, released as a closed model; Meta later offered it on its Meta Model API.Figures: company | 1 2 3 | |
| Gemini 3.1 ProGoogle DeepMind | APIProprietary | 1,000,000 | $2 | $12 | Released in preview, replacing Gemini 3 Pro Preview; prices shown apply to prompts up to 200,000 tokens, with higher rates ($4 input, $18 output) above that.Figures: company | 1 2 3 |
Every row links to its sources (numbered links). Figures supplied by a company are marked as company figures. See How we report. Spotted an error? Reply to any newsletter issue.