The latest AI models

GPT-6 Astra

3 Sep 2026OpenAI describes Astra as its most intelligent and best-aligned model to date. It launched to Plus, Pro, Business and Enterprise users in ChatGPT and to the API at $10 / $50, and is also offered through Microsoft Azure and AWS Bedrock. In OpenAI's own results it scores 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 72.6% on OSWorld 2.0.

Astra is also the first OpenAI model the company says meets the "Critical" cybersecurity threshold in its Preparedness Framework. Advanced offensive-cyber requests are refused by default, and OpenAI says a separate programme (Daybreak) will give vetted defenders broader access. Read my explainer on Astra.

GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol

22 Sep 2026The smaller siblings arrived three weeks later: Sol at $2 / $10 and Luna at $0.10 / $0.50. OpenAI says this is 50% cheaper than the GPT-5.6 promotional pricing. 29 Sep 2026At DevDay, OpenAI announced GPT-6.1 Sol, which it says delivers "near-Astra intelligence" at one-fifth of Astra's prices.

GPT-6 models on Google Cloud

21 Oct 2026Google Cloud now offers GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna in its Model Garden, in private preview for selected enterprise customers. Usage is billed on the customer's Google Cloud invoice and counts towards existing spending commitments, and OpenAI's Codex tools can connect to it for inference. Google says the models will come to its business Gemini products and Antigravity later, and that it is now the only hyperscaler offering models from all three leading closed-model labs. That is Google's description; Astra was already available through Microsoft Azure and AWS Bedrock.

GPT-6.1 Terra

26 Oct 2026OpenAI's new mid-priced model costs $1 / $5, with cached input at $0.05, which OpenAI says is 50% below GPT-5.6 Terra and GPT-6.1 Sol. OpenAI says it matches GPT-6 Astra on Agents' Last Exam at medium effort for a tenth of the price and more than three times the speed, and reports 50.3% on Terminal-Bench 4.0 at maximum effort. It is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra. OpenAI also says Terra meets the "Critical" cybersecurity threshold in its Preparedness Framework and keeps Astra's protections.

GPT-6 Cyber

5 Nov 2026OpenAI launched GPT-6 Cyber, a model specialised for security research, and rates it "Critical" for cybersecurity under its Preparedness Framework. OpenAI reports 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, and says it stayed within its assigned scope in 99.93% of a new honeypot test without its deployment prompt. It is available through the API and Codex to a select group of approved Daybreak Red customers first. Read my post on GPT-6 Cyber and the Linux results.

Claude Fable 5.2 and Mythos 5.2

9 Nov 2026Anthropic upgraded its top tier and cut the price from $10 / $50 (the 5.1 models, released on 1 September) to $8 / $40, with an optional fast mode at $16 / $80. Fable and Mythos are the same underlying model: Fable has cybersecurity and biology safeguards and is available to everyone, including on AWS, Google Cloud and Azure, while Mythos is only for organisations in Anthropic's Cyber and Life Sciences Verification Programs. Anthropic reports 49.1% on FrontierCode 2.0 at its highest effort setting, against 45.5% for Opus 5.5, and says a team of about 1,000 Fable 5.2 agents formalised the proof of Fermat's Last Theorem in 30 hours. It still recommends Opus 5.5 for most everyday work. Read my post on Fable and Mythos 5.2.

Claude Opus 5.5

22 Sep 2026Priced at $4 / $20, with cache reads at $0.20. Anthropic says it costs about 40% less than Opus 5 on typical workloads and generates output more than 30% faster. METR and Frontier Design tested it before release. Anthropic reports 66.4% on Terminal-Bench 4.0.

It is the first release since Anthropic's chief executive called for "pacing the frontier". It ships with Fable-class cyber and biology safeguards, a Life Sciences Verification Program for legitimate researchers, and output watermarking aimed at EU AI Act compliance.

Claude Sonnet 5.5 and Haiku 5.5

28 Sep 2026Anthropic says Sonnet 5.5 is more than 30% faster than Sonnet 5 and up to 30% cheaper. 7 Oct 2026Haiku 5.5 costs $0.10 / $0.50 for prompts up to 100k tokens, which Anthropic puts at about 75% cheaper to run than Haiku 4.5. It is the first Haiku with an adjustable "effort" setting. The same announcement halved Sonnet 5.5 cache reads to $0.10 and introduced monthly API credits for Max and Team subscribers. Read my piece on the Claude 5.5 family.

Gemini 4 Argon

30 Sep 2026Google announced Argon and gave it first to trusted cyber defenders through its Fairwind Program. The output limit rises to 1M tokens, up from 64K. Google reports 77.9% on DeepSWE v1.1, 51.3% on AutomationBench (first place in Google's comparison) and 91.7% on LVBench. Google also says it monitors Argon's chain of thought for signs of misalignment.

13 Oct 2026Argon became generally available at an introductory price of $2 / $10, rising later to $4 / $20. Google says it ranks first on the Vals Index and describes Argon refactoring roughly 6 million lines of legacy Chrome code.

Gemini 4 Carbon

28 Oct 2026Two weeks after Argon's general release, Google announced Carbon, which it calls its strongest model. Google says it ranks first on the Artificial Analysis Intelligence Index v5.0 with 59, and reports 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploiting them) and 71.0% on CWE-bench (patching them), measured without guardrails. Because of those security skills it is going first to a subset of cyber defenders and then to Fairwind partners, with developers, businesses and consumers to follow "as soon as possible". It will start at Argon's introductory $2 / $10, rising later to $4 / $20. Read my post on Carbon.

10 Nov 2026Carbon is now available to all paying developers and businesses through the Gemini API and Gemini for business, and is rolling out to Google AI Ultra subscribers in the Gemini app, Antigravity and AI Mode in Search. SecurityWeek reports that the introductory price runs until 30 November and then doubles. The specialised "Carbon for cyber" stays limited to Fairwind Program members. The same day, Google DeepMind said agents built on Carbon had formally verified the code behind the Signal protocol (my post).

Gemini 4 Hydrogen

12 Nov 2026Hydrogen replaces Gemini 3.8 Flash as Google's fast, cheaper model for everyday coding and agent work. It costs $1.50 / $7.50, flat across its whole 1M-token context, and adds a fourth "xhigh" thinking level. Google says it scores 51 on the Artificial Analysis Intelligence Index at its default setting, just above Argon's 50, and 57% on Terminal-Bench 4.1 against 18.2% for 3.8 Flash. It is available on the free and paid tiers of the Gemini API, in the Gemini app, and in Google Search for all users, which Google calls the biggest model upgrade to Search yet. In Antigravity, a new "Orchestrator" option pairs it with Carbon or Argon.

Gemini Decision 1

11 Nov 2026A different kind of model: instead of writing text, it scores fixed options, such as yes/no checks, multiple choice and rubric grades, for up to 1,000 questions about one input in a single pass. It accepts text, images, audio, video and PDFs. Google prices it at $0.10 per million input tokens with no charge for output, and says answers take 150 to 250 milliseconds end to end. It is aimed at jobs like routing requests to the right model and policy checks.

Gemini Omni 1.2 Flash

5 Nov 2026Google's video model is now generally available in the Gemini API and Gemini for business. It makes and edits clips of up to 15 seconds, up from 10, in one-second steps from 3 to 15 seconds, and lets you refine a clip through follow-up prompts instead of starting again. Google cut the price by 30% to $0.07 per second at 720p. You can draft at 360p or 720p and upscale the best results to 1080p or 4K, and outputs carry a SynthID watermark. In Google Vids, anyone with a Google account can make up to three 15-second videos a month at no cost.

Meta Muse Ember

9 Nov 2026Meta's smallest and most efficient model is built for the work agents repeat all day: tool calls, coding subtasks, routing and extracting information from long documents. It costs $0.10 / $0.50 for prompts up to 200K tokens, the same as GPT-6 Luna and Claude Haiku 5.5, with a 1M-token context window. Meta says it matches Haiku 5.5's best Artificial Analysis Intelligence Index score (36) at about a third of the cost per task, though Haiku leads on Meta's FrontierCode 2.0 comparison (34.9% against 27.2%). It is available through the Meta Model API and Muse Code, and is the first model Meta says it trained using Muse Wonder Full.

Meta Muse Wonder and Muse Spark Open

12 Oct 2026Meta launched Muse Wonder and Muse Wonder Full, with features it calls Active Memory and Active Focus, through the Meta Model API and its Muse Code tool. 15 Oct 2026Three days later it released Muse Spark Open with downloadable weights: 816B total parameters, 42B active per token, and a 1M-token context. Meta says it runs on two Apple Mac Studios with M5 Ultra chips. It is released under Meta's own Muse Model License rather than a standard open-source licence.

Mistral Large 4

6 Oct 2026Mistral released a public preview of Large 4 (internal nickname "Le Chonk"). It is a multimodal model with about 1 trillion parameters, 52B of them active, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres. The work is funded by a €3 billion Series D. Read my piece on open-weight models.

29 Oct 2026As promised, the weights are now out, under Mistral's Ecosystem Open-Weight licence, a modified MIT licence that requires companies offering the model as a hosted service to get a commercial licence. The API price is $0.70 / $2.10. Mistral reports 48% on the Artificial Analysis Coding Agent Index, 41% on Terminal-Bench 4.0 and 51% on the Artificial Analysis Cyber Index, and says it refuses 96.6% of malicious cyber requests. A specialised security version is limited to vetted partners. It is also on Microsoft Foundry.

Thinking Machines Inkling 2 Small

5 Nov 2026Thinking Machines Lab's second-generation model is a mixture-of-experts model with 594 billion parameters in total, 17 billion of them active at a time, that accepts text and images. The company says it scores 31 on the Artificial Analysis Intelligence Index and leads on several vision, finance and enterprise tests among the open-weight models it compared, while it trails others on some coding tests. These are the company's own benchmark claims. This weekAs promised, the full weights are now on Hugging Face under the Apache 2.0 licence.

IFM PAN 2 world model

10 Nov 2026The Institute of Foundation Models, set up by Mohamed bin Zayed University of Artificial Intelligence, released PAN 2, an open "world model": give it a starting scene and an action in words, and it generates a video of what happens next. It is released with model weights and full training and inference code under Apache 2.0. IFM says it placed third overall on the Physics-IQ Verified leaderboard and first among open-weight models, and says long-term consistency is still an open challenge.