On OpenAI's new MultiLangBench, where a model builds a SQLite-compatible engine in 20 languages, GPT-6 Astra averages an 87.4% test pass rate (COBOL jumps from 10.9% to 72.1%) versus 26.8% for GPT-5.6 Sol. On its COBOL migration eval, GPT-6.1 Sol hits 86.6% at low reasoning effort and 97.9% at high, though both benchmarks are OpenAI's own and not independently verified yet.
Foundry agents now get a managed cloud browser (via Playwright Workspaces, provisioned as a separate Azure resource) that can reach public sites and private VNet apps, exposed to prompt or hosted agents through a toolbox's MCP compatible endpoint. Operators can watch sessions live, take over mid run, and require approval before clicks, typing, or uploads, which you enable only when a task needs them…
Built on MCP, the connectors let an agent carry engineering work from CAD (CATIA, NX) through meshing (ANSA, HyperMesh) to simulation (Abaqus) across Windows workstations and Linux clusters, with engineers setting goals and validating results; one customer cut a wheel rim geometry cleanup from a day to about three minutes. It's available now, but only to existing Neural Concept customers (SiliconANGLE)…
The filesystem, memory, sequential thinking, and everything servers move to TypeScript SDK v2 and the 2026 07 28 spec while still answering older clients, and the Python fetch, git, and time servers ( 2026.10.19 ) now return a JSON RPC 32602 error for unknown tools instead of an isError result. The TypeScript servers require Node 24 or later, and v1.0.0 (Oct 11) was the final legacy era release, so pin it if you can't upgrade yet…
Workspace chats can edit the dev checkout, run commands, and read service logs (asking before Git pushes by default, with optional approvals for shell, web, and platform changes), while organization chats propose stacks, flag EOL service versions, and dig into failed deploys. It bills against one time org AI credits ($5 on the Developer plan, $10 with a paid subscription), and Wodby also opened the same dev workspaces to Codex and Claude Code (docs)…
MiniMax says the agent and coding model, which ran anonymously as "Space Bunny" and served over 200 trillion tokens, is nearly 3x faster than its predecessor (170+ tokens/sec median) and cuts agentic task cost by 70%, with a 1M token context and text, image, and video input. It's live now on the MiniMax API (OpenAI and Anthropic compatible), M Plan, and MiniMax Code, with weights and a bigger Pro model due later…
Spotted something we missed? Start a thread.