$ cat briefs/daily/2026-10-27.md
2026-10-27
October 27, 2026 (Tue)
1 model release · 2 paper picks · 3 new pages
Top Stories
1. GPT-6.1 Terra lowers the price of computer-use work
OpenAI's new model lists $1/M input · $0.05/M cached input · $5/M output. At maximum reasoning effort, OpenAI reports 50.3% on Terminal-Bench 4.0, versus GPT-5.6 Terra's published 21.5%. Access includes the API, ChatGPT Work and Codex; context capacity is not specified in the captured announcement.
Why it matters: developers gain a cheaper option for visual and desktop workflows, but vendor scores and token rates do not establish cost per successful task in a particular deployment. → GPT-6.1 Terra (source)
Paper Picks
RSIHub: shared infrastructure compares agent-evolution methods while keeping evaluators fixed. The largest reported overall pass@1 improvement is 13.5 percentage points; gains do not always transfer to held-out tasks. Read it for separating editable resources from the test. First submitted 2026-10-23. → RSIHub: comparing agent evolution under fixed evaluation (source)
UserArena: shopping simulations are checked against real trajectories and a controlled production A/B test. Authors say OPTD best reproduces the policy effect among tested simulators, without quantifying that effect in the abstract. Read it for external validation of simulated users. First submitted 2026-10-22. → UserArena: testing shopping simulators against production outcomes (source)
Watch
- OpenAI classifies Terra at the Critical cybersecurity threshold and says it retains GPT-6 Astra's protections. The full linked system card was not reviewed in this capture. → GPT-6.1 Terra (source)
New in Wiki
- GPT-6.1 Terra and both paper notes add specifications, methods and evidence limits. (model) (RSIHub) (UserArena)
Updates
- Eval Harness Configuration connects fixed evaluation with production validation, distinguishing what each can establish. (evolution) (simulation)