Released Oct 20 with FP8, NVFP4, GGUF, and MLX variants, the reasoning model has a 262K context and serves via vLLM or SGLang with tool calling, and Ornith reports 0.845 on SWE-bench Verified and 0.813 on Terminal-Bench 2.1 (its own numbers, per HuggingNews). Community quants from AtomicChat bring it down to about 35GB for 64GB Macs.
Since Oct 20, GPT 5.6 and later models in the Responses API accept prompt cache options.ttl: "12h" (or a per breakpoint ttl ), with 12 hour cache writes billed at 2.5x input versus 1.25x for the default 30 minutes and reads unchanged (changelog). OpenAI's Tibo (@thsottiaux) added that caches for logged in Codex and Work users will be kept permanently, so resuming an old conversation hits the cache (PANews)…
Per The Information, a tester said the unreleased model spotted broader related problems when asked for a simple bug fix, but it's unclear whether it can tell an incomplete request from work you deliberately left out. That gap matters: the Overeager Coding Agents study found Claude Code, Codex CLI, and Gemini CLI took out of scope actions in 5.4% to 27.7% of runs, versus 0.2% to 4.5% for ask first OpenHands…
Calling /create formal verification skill makes the agent pick a high stakes property (races, retries, parsers, unsafe code), choose a tool such as TLA+, CBMC, or Kani, show the check failing on a deliberately broken variant, then leave a project local skill so the next agent can rerun it with one command. pstack also ships into Claude Code's plugin marketplace from the same repo, and Cursor's own eval behind the skill is small (one run per repo on six open source projects)…
The open eval (GitHub) checks a from scratch SQLite 3.40.1 engine against 272 tests adapted from SQLite's own suite and scores pass rate alongside API cost; OpenAI says GPT 6 Astra averages 87.4% (COBOL hardest at 72.1%) vs 26.8% for GPT 5.6 Sol, while cheaper GPT 6 Luna hits 40.3% at about 78% lower cost. A separate COBOL migration eval has GPT 6.1 Sol at 97.9% on high reasoning, though all figures are OpenAI's own runs…
The single file, stdlib only Python tool (macOS only) confines a run to the folders you grant, blocks your home folder, and routes every API call through a loopback proxy that holds the real key, prices each response, and kills the run once the budget is spent. Its Codex CLI adapter is weaker: per the README it hasn't been tested against a live CLI and needs the key in the run's environment (write up)…
Spotted something we missed? Start a thread.