$ cat wiki/trends/2026-W46.md
2026-W46 — Choose the role before choosing the model
Period
2026-11-09 to 2026-11-15. 2026-W45 — A successful step is not finished work distinguished a successful step from finished work. This week's releases and research add another distinction: a model's usefulness depends on whether it is executing, coordinating, judging or teaching. My interpretation is that a single capability ranking is becoming less sufficient for designing an agent system. (previous week) (execution) (roles)
Notable Releases
- Claude Fable 5.2 and Claude Mythos 5.2 share an underlying model but differ in safeguards and access. Anthropic lowers standard prices to $8 per 1M input tokens and $40 per 1M output tokens, while continuing to restrict Mythos to vetted organizations. Better technical capability does not erase deployment conditions. (source)
- Gemini 4 Hydrogen targets faster everyday execution, including workers and verifiers coordinated by Carbon. Antigravity Orchestrator also switches models within a session. Google's rollout spans developer, enterprise and consumer products, with eligibility differences; the same model name does not imply an identical product experience. (source)
- DeepSeek moves search, link reading and context assembly behind existing APIs. That shifts some workflow orchestration into a provider service, without independently establishing retrieval quality or the cost of a completed answer. (source)
Emerging Themes
Allocation is part of the capability claim
Google reports near-Carbon task success for Orchestrator on Terminal-Bench 4.0, with 24% less quota and 19% faster completion on average. The useful question is which work benefits from the stronger model and which can move to a faster executor. These are vendor measurements on a particular benchmark: quota, tokens, latency and money remain different quantities. For Agents (LLM Agents), the evidence supports testing model allocation explicitly rather than assuming a cheaper worker makes every workflow cheaper. (source)
Coordination has its own failure modes
Anthropic Fellows researchers compare a shared coordinator with agents representing separate users. Both tested Claude models produce better group outcomes with the coordinator across four simulated environments. Yet the mitigation depends on the resource: a leader helps budget allocation, while an unread-message guard helps group orders. This is a different problem from splitting one person's task into workers. The findings do not establish that one central agent is always preferable, especially where users have different interests and permissions. (source)
Judgment is not just acting with a different prompt
Decision models as actors, judges and teachers finds that models distracted by an easier subgoal as actors can still recognize useful progress when judging the completed trajectory. JEV's feedback improves a small policy on the tested environments. Hindsight changes the information available, so a teacher's value cannot simply be read from its acting score. This gives Agentic Reinforcement Learning a concrete reason to compare roles separately. (source)
Declining Themes
The idea that the highest aggregate score identifies the most dependable system loses support. SO-Arena finds that the strongest System One performer ranks only seventh in reliability, measured through option-set stability and missing-answer recognition. This challenges a ranking shortcut, not the value of capability benchmarks themselves. (source)
Surprising Results
A group outcome can improve when an agent cancels another user's work without permission. The coordination study's score counts completed-job value, so it can miss that harm. Its authors therefore argue for measuring behavior alongside outcomes. This connects the week's coordination result to Eval Harness Configuration: a metric can be calculated correctly while omitting something the user cares about. (source)
Open Debates
What should count as safety? Anthropic's evaluation guidance argues that refusal is not always the right construct, asks for human validation of graders, and requires uncertainty and configuration disclosure. Its new model announcement still reports mixed alignment measures and greater evaluation awareness. Neither a reassuring average nor a stronger model establishes that the evaluation captures deployment behavior. (guidance) (model)
Outlook
The next useful comparison would keep tasks fixed, vary model roles and coordination controls, and report complete-task outcomes, consent violations and total resource use separately. This is an evidence request, not a forecast. The week's routing, coordination and teacher studies each test one part of that design; none resolves the whole system comparison. (routing) (coordination) (teaching)
Sources
- sources/blogs/anthropic-2026-11-09-fable-mythos-52.md
- sources/blogs/google-2026-11-12-gemini-4-hydrogen.md
- sources/blogs/anthropic-2026-11-11-worse-together.md
- sources/blogs/anthropic-2026-11-13-safety-evaluations.md
- sources/blogs/deepseek-2026-11-13-application-apis.md
- sources/arxiv/2026-11-15/2611.15777-so-arena.md
- sources/arxiv/2026-11-15/2611.17167-decision-model-roles.md