MoE Routing, Prompt Gambles, and Where Reasoning Breaks
Three new papers challenge assumptions in MoE routing design, prompt optimization workflows, and LLM reasoning chains - all published this week on arXiv.
Reasoning
Arcee Trinity: Open-Source 400B Reasoning Agent
Arcee Trinity-Large-Thinking is a 400B sparse MoE open-source reasoning model that ranks #2 on PinchBench at $0.85/M output tokens, 28x cheaper than Claude Opus 4.6.
Claude Opus 4.7
Anthropic's latest flagship model ships with 3x higher resolution vision, a new xhigh effort level, task budgets for cost control, cyber safeguards, and 13% better coding performance at the same $5/$25 pricing.
Compact Contexts, Smarter Fine-Tuning, and the Solver Trap
Three papers from today's arXiv: a joint fix for KV cache bloat and attention cost, new evidence that fine-tuning belongs in the middle of a transformer, and why stronger reasoning hurts behavioral simulation.
Claude Mythos Preview - Anthropic's Restricted Frontier
Claude Mythos Preview is Anthropic's most capable model - restricted to 50 orgs via Project Glasswing, with 93.9% on SWE-bench Verified and thousands of autonomous zero-day discoveries.
Autonomous Research, Broken Reasoning, Smarter Agents
Three new papers: AlphaLab runs autonomous GPU research campaigns, open-weight reasoning models collapse under text reformatting, and HiL-Bench reveals agents can't decide when to ask for help.
Grok 4.20 Review: Four Minds Are Better Than One
xAI's Grok 4.20 replaces the single-model approach with four specialized agents that debate before every answer - a bold architectural bet that pays off in some areas and stumbles in others.
Arcee's Trinity-Large: 398B Open Reasoning at $0.90
Arcee AI ships Trinity-Large-Thinking, a 398B sparse MoE reasoning model under Apache 2.0 that hits 91.9% on PinchBench for $0.85 per million output tokens on OpenRouter.
Clinical AI Harm, Smarter Reasoning, and Safer Agents
Three papers: AI safety measures withhold critical clinical guidance from patients, SAT cuts reasoning tokens by 40%, and conformal prediction blocks wrong multi-agent consensus.
EXAONE 4.5: LG's Open VLM Beats GPT-5-mini on STEM
LG AI Research released EXAONE 4.5, a 33B open-weight vision-language model that posts higher STEM scores than GPT-5-mini and Claude 4.5 Sonnet - but a non-commercial license caps its real-world reach.
Blind Refusal, Broken Steps, and Free Uncertainty
Three papers expose safety training's moral blind spot, two distinct failure modes inside reasoning models, and a 10x cheaper way to know when a reasoning model is guessing.
Muse Spark
Meta's first closed-source frontier model scores 52 on the Artificial Analysis Intelligence Index, leads on HealthBench Hard, and ships free at meta.ai - but has no public API yet.
