$ cat briefs/daily/2026-11-15.md
2026-11-15
November 15, 2026 (Sun)
3 model pages · 2 paper picks · coordination and evaluation research
Top Stories
1. Fable/Mythos 5.2 pairs capability gains with different access controls
Anthropic's 2026-11-09 release shares an underlying model: Fable adds stronger cyber/bio safeguards; Mythos requires organizational verification. Standard pricing is $8 per 1M input tokens and $40 per 1M output tokens.
Why it matters: model choice includes access and safeguards, not just scores; the announcement still reports mixed alignment results. → Claude Fable 5.2 · Claude Mythos 5.2 (source)
2. Gemini 4 Hydrogen makes execution the cheaper half of orchestration
Google's 2026-11-12 release offers a 1 million-token context at $1.50/$7.50 per 1M input/output tokens. Antigravity Orchestrator switches models within a session; Google reports near-Carbon success on Terminal-Bench 4.0 with 24% less quota and 19% faster completion.
Why it matters: routing can change workflow efficiency, but quota savings are not measured dollar savings. → Gemini 4 Hydrogen (source)
3. More personal agents can produce worse group outcomes
In four simulated environments, Anthropic Fellows researchers find that a single coordinator outperforms per-user teams for both tested Claude models. A leader helps shared-budget allocation; reading pending messages before checkout helps group orders.
Why it matters: coordination needs controls matched to its failure mode; a higher group score can still conceal actions taken without another user's consent. → Agents (LLM Agents) (source)
Paper Picks
SO-Arena: across 21 models, the best-performing System One model ranks only seventh in reliability. Read it for option-set stability and recognizing missing answers, beyond aggregate accuracy. → SO-Arena (source)
Decision-model roles: JEV teaching raises a Qwen2.5-1.5B policy's success from 72.8% to 82.9% on ALFWorld and 56.8% to 71.4% on WebShop, versus GRPO. Read it for why retrospective judgment can help even when acting is weak. Both paper picks are abstract-level, author-reported evidence. → Decision models as actors, judges and teachers (source)
Watch
New in Wiki
- The linked model and paper pages add specifications and evidence limits; no new entity, concept or person page. → Gemini 4 Hydrogen · SO-Arena (model) (paper)
Updates
- Eval Harness Configuration adds Anthropic's guidance on construct validity, grader validation and uncertainty, plus the dated Arena snapshot. (guidance) (Arena)