4B model hits 82.9% mean pass@1 on AIME 2024–2026 with constant memory

An internal Amazon memo pushes back on "tokenmaxxing," urging engineers to be more selective about when they use AI coding tools. The guidance lands as Amazon has rapidly spread AI across its business under CEO Andy Jassy.

Katelyn Lesse and Angela Jiang, Anthropic's heads of Engineering and Product, explain the Claude Platform: launching Opus 5.5, decoupling the model's "brain" from the agent's "hands," and rebuilding the Messages API in Rust. Anthropic credits a -126% loss to 19% GAAP profit turnaround to its improvement and investment loops.

Vercel rebuilt how Agent manages context, cutting tokens per session and improving prompt-cache reuse across dashboard and Slack sessions. Billing rates are unchanged: paid sessions cost the model provider's inference rates plus a $0.25 per million token Vercel Agent fee.

The community-built ComfyUI text generator adds context memory so prompts can be refined across runs without re-pasting everything, plus new prompt guides and a cleaner UI.

The AI SDK HarnessAgent layer now accepts images alongside text in standard AI SDK message format, mixing text and file parts. Supported across all adapters: Claude Code, Cline, Codex, Cursor, Deep Agents, fx, GitHub Copilot, Grok Build, OpenCode, and Pi.