GPT-6.2 Astra and Sol, Microsoft's MAI Models, and More
Tuesday's big launch was OpenAI's GPT-6.2, which comes in two flavors: Astra, the new frontier model for hard projects, and Sol, the efficient everyday model. Both are live in ChatGPT, Codex, and the API, and Astra ranks #2 on the Artificial Analysis Intelligence Index.
The reception was the interesting part. Early power users are split, and VentureBeat rounded up the "essentially unusable" complaints. OpenAI's Tibo says some influencers got different early access builds, which explains part of the gap between the benchmarks and the vibes. I'd give it a few days of real use before drawing conclusions either way. (OpenAI, OpenAI on X, VentureBeat, Tibo on X)
What Else Happened
Microsoft launched seven MAI models at Ignite. The lineup is MAI Code 2 and Code 2 Flash (now in GitHub Copilot, with Microsoft claiming about 98% of GPT-6.1 Sol's coding performance at 40% of the price), MAI Cyber 2 Flash, MAI Image 2.7 and Flash (#2 lab across the Image Arenas), MAI Realtime 1, and MAI Transcribe 2. Microsoft is now openly competing with OpenAI and Anthropic on price-performance. (Microsoft AI, Microsoft AI on X, Arena on X)
Nvidia posted 115.8 billion dollars in Q3 FY27 revenue, up 103% year over year. Data Center brought in 107.6 billion, and Q4 guidance of 130 billion (plus or minus 3%) came in ahead of the street. It was the first report since the CEO-transition news, and Nvidia didn't take questions on the resignation. (Nvidia, CNBC)
"Java4Shell": Codex found a critical OpenJDK RCE in about an hour. A Calif researcher used Codex to find a memory-corruption bug in the AVX-512 Base64 decoder (CVE-2026-114938, affecting OpenJDK 17.0.6+ on supported x86-64). A working exploit took about two more hours and under 200 dollars in tokens. If you run Java services, check your JDK version today. (Calif, BleepingComputer)
OpenAI is working with NATO on cyber defense. NATO's Communications and Information Agency gets OpenAI models for threat detection, finding vulnerabilities, and testing fixes, and it already uses about 300B tokens a month for this work. (OpenAI, Greg Brockman on X)
Gemini 4 Neon looks close. It showed up in the Gemini web app code and in Google Cloud SKUs priced at 3 and 15 dollars per million input and output tokens, twice Hydrogen's price. Google hasn't announced it, though Logan Kilpatrick posted a lone "Gemini" last night. (TestingCatalog, Logan on X)
Launches Worth a Look
Claude Code has an experimental Code Mode. Claude writes JS scripts that chain Bash, Read, and Edit calls, some in parallel, which cuts token use on heavy workflows. (ClaudeDevs on X)
Claude Platform added linked billing in beta. Tools like Linear, Ramp, Factory, and Mintlify can bill Claude API usage to your org without you sharing API keys, with spend caps and the option to unlink anytime. (Claude)
Cognition launched Devin Bots. They're persistent agents with their own identity, memory, and permissions that run teams of sub-agents. Cognition says one game studio went from 50 to 1,900 PRs. (Cognition on X)
SID-2 is a search model built on compute-scaling agent swarms. SID says it beats Opus and Astra on results while running 8x faster and 20x cheaper. (Max Rumpf on X)
Details MCP showed an agent "unslopping" a vibecoded website. The 29-second demo has an agent pulling from curated real-site designs to fix spacing, type, and layout. (Kevin on X, Details MCP)
Google AI Studio can now ship Android apps. You can build an Android app and install it through Google Play (US), and buy and attach custom domains inside AI Studio. (Logan on X, Google AI Studio on X)
Gemini Tasks is now free for hundreds of millions of users in the US, Japan, Korea, Canada, Australia, and India. It handles longer projects and recurring schedules. (Gemini on X)
ChatGPT for Sales pairs a Sales plugin with 30+ partner plugins, including Salesforce, HubSpot, Gong, Clay, and Attio. (ChatGPT on X)
Kling Image 4.0 generates native 8K photoreal images and adds style-consistency tools. (Kling on X)
Higgsfield Miracle splits video into editable layers so you can recolor, relight, or move one element without regenerating the shot. It's in private beta. (Higgsfield on X)
Karpo Decision is six open-weight models (0.8B to 31B, built on Gemma 4) for agent tool selection and browser navigation. The 31B tops the Decision Index. (Leah on X)
Bonbon, a NeurIPS Oral, is a 1.6B biology foundation model with a 61.5% wet-lab hit rate across 9 disease targets. (Adi Mashiach on X)
What I'm Reading
- Paul Jun, "Brand as Software" part 2 – how Ramp moved senior designers from one-off assets to building brand tools and agents.
- femke on PMs shipping AI prototypes – on designers getting stuck as blocker or clean-up crew, with good replies on how to get ahead of it.
Accounts Worth Following
- @maxrumpf – Max Rumpf, co-founder and CEO of SID, building search swarms that hold their own against frontier models.
- @PaulJun_ – Paul Jun leads brand at Ramp and writes "Brand as Software".
- @femkesvs – femke, design lead at Gusto, writing on designers' strategic influence.
That's it for today. 👋