The Day in AI
In short
- Microsoft released MAI Code 2 at Ignite and says it reaches 98% of GPT-6.1 Sol's coding performance at 40% of the price.
- Microsoft will license the weights of its new MAI models so businesses can build customized versions, according to The Information.
- OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, with Astra priced at $8 per million input tokens and $40 per million output tokens.
- Google made Tasks in Gemini, the agent first tested as Gemini Spark, free in six countries, and it can now call businesses and pay through Google Pay.
- Jasjeet Sekhon, chief strategy officer at Google DeepMind, is leaving to join Anthropic.
The day in brief
Microsoft released MAI Code 2 at its Ignite conference on 17 November and says the model gives 98% of GPT-6.1 Sol's coding performance at 40% of the price.
Microsoft's figure is an average over four coding benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), with every model at maximum reasoning effort in GitHub Copilot CLI. On the same average, Microsoft puts Code 2 at 96% of Claude Opus 5.5 for 5% of the cost. A smaller MAI Code 2 Flash handles high-volume work such as targeted edits and routine fixes. Microsoft says neither model is yet tuned for very long agentic work or hard science problems.
Code 2 is one of a batch of MAI models Microsoft showed at Ignite, covering coding, cybersecurity, images and voice. Microsoft says its voice model MAI Realtime 1 starts at $0.01 per audio minute, and Artificial Analysis ranks MAI Image 2.7 first on image editing. The Information reports that Microsoft will license the models' weights so business customers can customize them.
OpenAI released GPT-6.2 Astra, its new frontier model, and GPT-6.2 Sol, an efficient model, on the same day. Astra costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens, which OpenAI says is 20% below GPT-6 Astra. OpenAI reports that Astra scores 9 points above GPT-6 Astra on Terminal-Bench-Science 0.1 at under one-third the cost. It's also trained to produce slides, spreadsheets and documents that follow professional conventions.
Microsoft's comparison uses GPT-6.1 Sol, the version before today's. In Cognition's FrontierCode 2.0 tests, GPT-6.2 Sol comes within 0.5 points of Astra at about one-fifth the cost. OpenAI says Sol keeps GPT-6.1 Sol's prices, which 9to5Mac puts at $2 per million input tokens and $10 per million output tokens.
Also in the news
- Jasjeet Sekhon, DeepMind's chief strategy officer since March 2026 and formerly chief scientist at Bridgewater Associates, is joining Anthropic to work on epistemic rigor in Claude.
- Google made Tasks in Gemini free in six countries, and TechCrunch and The Verge report that it runs on the new Gemini 4 Hydrogen model.
- OpenAI added two-hour and eight-hour background windows to the Responses API in public beta, with token rates 30% and 40% below Standard.
- Anthropic launched Linked billing, which lets customers pay for Claude usage inside a developer's app through their own Claude organization, with spend caps in the Claude Console.
- OpenAI released a free image classifier for vetted organizations that gives separate scores for potential CSAM and potential AI-generated CSAM.
- Alibaba's QwenWork office agent now lets users pick models from Kimi, GLM, DeepSeek, MiniMax and Step alongside Qwen, with access set by department or user.
- NVIDIA announced Nemotron for Science at SC26 on 16 November, pairing the Nemotron 3.5 Super VL open model with an open chemistry dataset and post-training recipes.
- Cloudflare introduced two models of its own, a decision model called Clef and an extract model called Clairic, according to an interview on Latent Space.
- ByteDance opened the API for Seedance 2.5 4K, its video model for film production, to developers, TechNode reports.
- Genesis AI, a French robot maker, raised $500 million at a $3.5 billion valuation, Reuters reports.
- Moonshot, Tencent and Zhipu reportedly limited the cyberattack capabilities of their new models.