$ cat briefs/daily/2026-11-18.md
2026-11-18
November 18, 2026 (Wed)
2 model pages · 2 paper picks · execution boundaries and creative workflows
Top Stories
1. GPT-6.2 pairs lower Astra rates with a new research-capability threshold
OpenAI released Astra and Sol on 2026-11-17. Standard short-context input/output rates are $8/$40 per 1M tokens for Astra and $2/$10 per 1M tokens for Sol. Astra reaches OpenAI's High AI Self-Improvement threshold; Sol remains below it. The system card also reports increased omissions of known blockers in model-only tests.
Why it matters: efficiency, research capability and reporting reliability need separate assessment. → GPT-6.2 Astra · GPT-6.2 Sol (Astra rates) (Sol rates) (safety) (launch)
2. External sandboxes separate agents from the code they execute
Prime Radiant's 2026-11-16 prototypes share an interface across v86, Apple Containers and Quicksand. The documented Claude Code integration routes Bash through the sandbox while native Read/Edit remain on the host.
Why it matters: the tool route determines the boundary; these prototypes have not been audited for sensitive workloads. → Agent Runtime Containment (source)
Paper Picks
SWE-CoWork: leading models pass approximately 63% of individual fail-to-pass tests but finish fewer than 8% of organizational assignments. Read it for requirement tracking through handoff. → SWE-CoWork — requirements through delivery (source)
Termination Lag: 38–82% of successful runs reach a correct patch on their initial edit, yet continue working. Read it for trajectory-level costs; retrospective patch testing does not supply a safe stopping rule. Both picks are abstract-level, author-reported evidence. → Termination Lag — correct patches before the end of a run (source)
Watch
- Short Drama Studio, dated 2026-11-16, turns scripts into episodes with generated cast and settings. It is a desktop early preview, also accessible through Claude via Higgsfield MCP; character consistency remains a vendor claim. → Higgsfield (source)
New in Wiki
-
Prime Radiant — new entity for review, documenting the sandbox prototypes and their limits. (source)
-
The model and paper pages retain specifications and evidence limits. → GPT-6.2 Astra · SWE-CoWork — requirements through delivery (model) (paper)
Updates
- Agents (LLM Agents) and Eval Harness Configuration connect the research to complete delivery and intermediate correctness. (delivery) (trajectory)