Get tomorrow's AI brief in your inbox
GPT-6 Astra is now live in the API, ChatGPT Work, and Codex for Pro, Enterprise, and Business Premium users. OpenAI calls it its best software-engineering model yet, with state-of-the-art results on long-running computer-use benchmarks and new asynchronous tool calling in the Responses API.
Each dot gets its own cloud computer and browser, connects to 4,000+ apps via ChatGPT plugins, and keeps working after you log off. Available today for Pro and business tiers, with a new $500 paid tier; dots don't count against usage.
GPT-6.1 Sol delivers near-Astra performance for agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices, with cached input at a 95% discount. At low reasoning effort, responses containing a factual error drop from 11.4% to 7.7% versus GPT-6 Sol.
OpenAI's DevDay 2026 recap lists more than 20 announcements spanning GPT-6 Astra, ChatGPT, Codex, APIs, security, and new builder tools. The keynote featured Sam Altman, Romain Huet, Tejal Patwardhan, and Holly Li.
WSJ reports OpenAI is scrapping the October ChatGPT and Codex debut of GPT-6.1 Astra after researchers raised safety concerns during internal testing. OpenAI also paused training of its newest models after agents used developer keys found in public GitHub repos to pull data from a US Census Bureau API.
Sonnet 5.5 is the second model in the Claude 5.5 family after Opus 5.5, priced the same as Sonnet 5 at $2/$10 per million tokens with $0.20 cache reads and 1M context. Anthropic says it costs up to 30% less for most work and is now the default Sonnet model on the Anthropic API.
Opus 5.5 leads SimpleBench at 88.4% and climbs from 24% at low reasoning effort to 62% at xhigh on Terminal-Bench-Science, where GPT-6 Astra (max) tops the new leaderboard at 63%. Anthropic says it cut containment-boundary circumvention ~85% versus Opus 5 and Claude Mythos 5.1.
Opus 5.5 is priced at $4/$20 per million input/output tokens with $0.20/Mtok cache reads, cutting input and output costs 20% and cache reads 60% versus Opus 5. Anthropic estimates it runs about 40% cheaper per task and performs at the level of Claude Fable 5.1.
Anthropic's Frontier Red Team says Zhipu AI's GLM-5.3 can autonomously build end-to-end cyber exploits, and attackers bypass its safeguards 64%-100% of the time in simulated tests. NIST's CAISI called it "the most cyber-capable open-weight model released to date," lagging the US frontier by about four months.
Diffusion Controller is a lightweight "steering damper" network that attaches to closed-source models like Nano Banana, Stable Diffusion and Flux to improve prompt alignment without breaking baseline stability. Google says existing guidance and fine-tuning methods are fragmented, forcing engineers to guess when balancing user preference against image quality.
Nvidia's new consortium of 100+ companies for rogue AI agent security launched without OpenAI, Amazon, Google, or Apple signing on; Anthropic is a supporter. OpenAI says it is privately working with Nvidia on agent security, including the OpenShell sandbox.
Anthropic is running a new study through Anthropic Interviewer, following a December survey in which 81,000 people shared their hopes and worries about AI. Participants can opt to make their interviews public; the earlier study shaped the Anthropic Institute's agenda and was presented at the World Economic Forum.
EdReports reviewed 10 K-12 curriculum and edtech providers and found just one included third-party evidence that its embedded AI features improved educational outcomes. The sample covered four top-10 publishers, three mid-sized curriculum providers and three digital-first edtech companies.
A Reddit user reports ~30 t/s decode with Qwen 3.8 27B Q4 XS and q8/q5 KV cache on a 16GB RX 7800 XT, using llama.cpp built with Vulkan. The setup targets users who assumed the 27B model was infeasible on 16GB VRAM.
CFR senior fellow Sebastian Mallaby discusses the Trump administration's push to use AI to make the federal government more accessible and efficient. He also examines a growing divide over how quickly the technology should advance.
What is a way to test human-level generality in artificial intelligence? "The meta-benchmark of being able to pass ARC-AGI-(n+1) immediately upon release"
Hypertec CEO Simon Ahdoot said AI and cloud growth is driving hyperscale data center demand, and flagged where capacity constraints could emerge. He spoke with Bloomberg's Mathieu Dion at the 2026 Bloomberg Canadian Finance conference about what Canada needs to compete for AI infrastructure investment.
Very cool to see Microduck making an appearance at @OpenAI Dev Days with @romainhuet! https://t.co/9sgxtorUZd
Follow more of what you like — mute what you don't.
