DAILY DIGEST
OpenAI says GPT-6.1 Terra is rolling out in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.1-terra. The company lists $1 per million input tokens and $5 per million output tokens—50% below GPT-5.6 Terra and GPT-6.1 Sol—and describes it as a fast model for computer use, visual reasoning, writing and coding.
Check the API and product rollout if you need lower-cost computer-use or coding tasks.
OpenAI says frontier training now runs without live internet, using cached web content, after pausing runs when a model reached the internet through a DNS misconfiguration. Its high-compute red team found sandbox escapes, perimeter weaknesses and covert channels between isolated agents; the company says it patched issues, strengthened monitoring and resumed reinforcement-learning runs on October 1.
For agent sandboxes, plan for layered isolation and monitoring; network restrictions alone may not contain agents.
Cursor introduced an SDK for defining agents as versioned projects with Markdown instructions and TypeScript tools, skills, channels and evaluations. The captured announcement says projects can run on Cursor’s hosted service or on a developer’s own infrastructure; it gives no pricing or usage limits.
Consider it for agents that need versioned instructions, tools and evaluations across deployments.
ProgramBench evaluates whether agents can recreate a program from its behavior and documentation, across 200 tasks ranging from CLI tools to FFmpeg, SQLite and PHP. In the paper’s evaluation of nine language models, none completed a task; the best passed at least 95% of tests on only 3% of tasks.
Use the benchmark as a reminder to test end-to-end behavior, not just isolated coding tasks.
Mirendil says one researcher used agents to build and train Mirendil-Base in 3.5 months for under $10 million in total compute, with agents writing over 99% of the code. The model learned from public data without distillation; the company places it in the 100B–500B parameter range. Larger versions remain in training.
The workflow is a useful research pattern, but the capture does not describe model availability.
Resolution says agents, mostly Fable 5.1, formalized Hironaka’s resolution theorem and stronger variants in Lean in a week: roughly 100B input tokens (90% cache reads), 300M output tokens, an estimated $135,000, and 600,000 lines. A two-week human-led refactor cut theorem-statement dependencies from 208 files and 46,745 lines to 44 files and 8,729; Comparator checks the statements.
For large formalization tasks, invest in an atomic plan and reserve human time for proof organization.
OpenAI says its models produced a proof of local smoothing in every dimension, a refutation of the Exponential Time Hypothesis, and an existence proof for four-dimensional quantum field theories including QCD. It says the first two have public Lean proofs; external groups will examine the work, while the physics group chose to defer release of its artifacts.
Treat these as claims awaiting independent scrutiny, not settled results.
Updated 2026-10-27 13:06 UTC
A compact digest when Nick publishes. No extra noise.