☕️ OpenAI gives GPT-4o users 3 months

• OpenAI will retire GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini from its API on January 14, 2027, giving developers roughly three months to migrate production traffic, snapshots, and every fine-tuned model built on them.
• The recommended replacement, GPT-6 Luna, costs half as much as GPT-5.6 Luna, but once a prompt exceeds 272,000 input tokens the input rate doubles and output rises 50% for the whole request.
• Luna handles reasoning and tool calls differently: function calling via Chat Completions needs reasoning_effort=none, combining both requires the Responses API, and enabling reasoning drops temperature, top_p, and logprobs that GPT-4.1 supported.

• Meta released open weights for Muse Spark, its coding and long-running agent model that stayed proprietary after an April debut, now shipping as Muse Spark Open under a new license permitting free commercial use.
• The model activates just 42 billion of its 816 billion parameters per token, accepts text mixed with images, audio or video, and scored more than twice Thinking Machines' Inkling on the JobBench agent test.
• Chinese models like Kimi K3 still beat it on Terminal-Bench coding, but it topped every rival on the MCP-Atlas tool-use test; selling it as a hosted service past $20M revenue requires a separate agreement, and military uses are banned.

• Anthropic has moved Code Review, its automated pull-request reviewer for Claude Code, out of research preview for Team and Enterprise customers, offering eligible enterprises up to $50,000 in credits that expire 30 days after they're claimed.
• Administrators install a GitHub App, grant read access, pick repositories and set each to Quick, Standard or Deep, with reviews now running automatically on new pull requests in six to ten minutes versus about twenty during the preview.
• Organisations on GitHub Enterprise Cloud with data residency can't claim credits, and connecting such an instance afterwards halts credits for the whole org; Anthropic also warns its reviewer may miss bugs and that a clean review doesn't mean sound code.

• Vercel launched Daemons, a public-beta service for running always-on background workers, letting developers deploy a container and set how many instances run per region while the platform keeps them alive and restarts any that crash.
• At launch it targets pull-based workloads: workers fetch jobs from a queue or database and process them as they arrive, suited to CPU-bound tasks like video transcoding, where an example pairs a Next.js frontend and Python API with an FFmpeg container pulling from Vercel Queues.
• Deploys go through Git or the Vercel CLI, with images built and kept in its Container Registry; the catch versus a managed runtime is that you own patching the container and must redeploy for security updates, whereas Vercel pushes those for Node, Bun, or Python runtimes.

• Theo Browne used AI agents to port Microsoft's TypeScript compiler, type checker, and language server to Rust, producing a working implementation that passed the conformance tests with near-100% accuracy, a project he considered impossible earlier this year.
• Claude Opus 5.5 cracked it for about $24,000 in tokens by porting Microsoft's Go version of TypeScript to Rust rather than translating from JavaScript; Browne claims the result runs twice as fast as the official Go rewrite.
• Because Rust needs no garbage collector, it compiles cleanly to WebAssembly, so Browne argues the port could let agentic TypeScript workloads run inside Google's V8 engine instead of a full VM or Docker container, cutting compute costs.

📚 Tutorials & Advice

🔗 Other News & Articles

• Sequences and other volatiles in Neki LINK

• KDE turns 30 with Plasma 6.8, remote upgrades, and an X11 farewell LINK

• Benchmarking DuckDB v2.0-dev on NVIDIA Vera LINK

• How Meta built Forum with AI-Native Engineering LINK

• Redis now has compressed bitmaps (Roaring) LINK

• YouTuber built 30-frame-per-second grayscale Game Boy, then open sourced it for all e-paper displays LINK

• Pure UI in the Agentic Era LINK

• Svg “linked parameters” will change how we do icon systems LINK

• Why your platform team shouldn't build an AI context layer (even though they probably could) LINK

• Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop LINK

📄 Papers & Resources

🧰 Latest Tools

💬 How did you find today's edition?

We read every reply — just reply to this email and let us know how we can improve!