Mythos thaws as rivals fill the gap

The US export clampdown on Anthropic's frontier models started to unwind today: Mythos 5 is back for select critical-infrastructure firms and Fable 5 looks days away. The pause has already cost Anthropic momentum, with Asian labs shipping Mythos-class rivals and Beijing reframing vulnerability-hunting AI as cyber-nuclear deterrence. Elsewhere, a Princeton benchmark showed agents still can't run a company, and a 3B model kept punching far above its weight on verifiable tasks.

US restores Mythos 5 to trusted firms; Fable 5 expected back within days

Two weeks after the Trump administration's June 12 order forced Anthropic to pull Mythos 5 and Fable 5 for all users, the government has cleared Mythos 5 for redeployment to a set of US organizations defending critical infrastructure, reportedly 100-plus firms including many Fortune 500 names. Commerce Secretary Howard Lutnick signaled Fable 5 could follow soon, pending Pentagon and NSA sign-off. Mythos and Fable share the same underlying model; Fable is the publicly available variant while Mythos ships with some safeguards lifted for cybersecurity work.

Why it matters: If you build on Claude, this is the first concrete sign the access freeze is reversible, but the case-by-case vetting process Anthropic and OpenAI are now lobbying to formalize means frontier-model availability is a policy variable, not a given.

Asian labs ship Mythos-class rivals while Anthropic alleges Alibaba distillation

With Anthropic's export ban dragging on, Tokyo's Sakana AI launched Fugu, an agent-orchestration model it pitches as standing alongside Fable 5 and Mythos Preview, and China's Qihoo 360 unveiled Tulongfeng (vulnerability discovery, said to have flagged 3,432 bugs) and Yitianzhen (automated defense). Founder Zhou Hongyi framed vulnerability-hunting AI as a 'cyber-nuclear' deterrent and pegged China's models 20-30% behind the West, betting on agent harnesses to close the gap. Separately, Anthropic accuses Alibaba of distilling Claude via fake-account API queries, raising the question of how defensible a frontier moat really is ahead of a rumored $1T IPO.

Why it matters: Querying an API is not exporting a model, so export controls don't touch distillation, the cheapest known way to close a capability gap. For developers, it means a widening field of Mythos-adjacent options outside US jurisdiction.

Princeton's CEO-Bench: most models go broke running a fake startup, and a hard-coded heuristic beats them

CEO-Bench tasks an agent with running a fictional SaaS company (NovaMind) for 500 simulated days via a Python API of 34 tools and a 19-table database, judged on remaining cash. Of 14 models, only Claude Fable 5 ($47.15M), Claude Opus 4.8 ($27.8M) and GPT-5.5 ($21.3M) finished above the $1M starting capital, and a simple rule-based heuristic with no LLM hit $15.76M, beating every other model. The researchers use fixed transparent rules rather than an LLM referee, and note running the same agents inside Claude Code and Codex made them act less and perform worse, blaming dev-tuned system prompts.

Why it matters: Strong local tool competence does not equal long-horizon strategy under delayed, noisy feedback. The harness finding is a direct warning: a coding-optimized agent wrapper can actively degrade an agent on non-coding tasks.

VibeThinker-3B argues reasoning compresses but knowledge doesn't

Sina (Weibo's parent) released VibeThinker-3B, a 3B model post-trained from Alibaba's Qwen2.5-Coder-3B that reportedly matches DeepSeek V3.2 and Kimi K2.5 on competition benchmarks like AIME26 despite being 200-333x smaller, and tops every sub-20B model on LiveCodeBench. On contamination-controlled LeetCode contests it solved 123/128 first-try, ahead of GPT-5.2 and Claude Opus 4.6. But on knowledge-heavy GPQA-Diamond it falls well behind larger models. The team's 'Parametric Compression-Coverage Hypothesis' says structured reasoning relies on few reusable patterns and packs into a small core, while broad world knowledge still needs scale. Weights are on Hugging Face and GitHub.

Why it matters: More evidence that for verifiable, structured tasks parameter count is no longer the bottleneck, which is exactly the regime where a cheap local 3B can replace an API call. Just don't ask it for facts.

55 LLMs blind-grading each other reveal systematic same-family bias

An open evaluation setup had 55 models from 11 developer families blind-grade each other in an N×N matrix with self-judgments excluded, yielding 22,254 valid judgments over 198 hand-written questions. Same-family rating bias was statistically significant in all 8 families with enough data: Qwen judges rate other Qwen models +0.91 and xAI +0.75, but Google (-0.59), Meta (-0.68) and Mistral (-1.02) penalize their own siblings. Code is where judges disagree most, nearly double the disagreement of meta-alignment, and in one run judges preferred an answer that failed the test suite. Code, dataset and prompts are MIT-licensed.

Why it matters: If you use LLM-as-judge in your eval pipeline, the judge's family is a confound, and single-judge code evaluation is the shakiest of all. Anchor to execution or tests wherever ground truth exists.

A field guide to running coding agents on a fully local stack

Sebastian Raschka published a long, practical walkthrough of wiring open-weight models into coding harnesses, primarily Qwen3.6 35B-A3B (~22GB download, 30-40GB RAM, ~40 tok/s on an M4 Mac Mini) served via Ollama and connected to Qwen-Code, Codex CLI and Claude Code. Notable findings: Qwen3.6 actually scored better inside Codex than its 'native' Qwen-Code harness; Claude Code burned by far the most tokens (one run logged ~578k input vs ~4.5k output tokens over 25 turns) due to its harness re-feeding context, not longer outputs; and he includes a concrete prompt-driven security audit checklist plus a settings.json to disable telemetry. North Mini Code and Nemotron 3 Nano are flagged as comparable alternatives.

Why it matters: The token-usage gap between harnesses is the actionable bit: with identical task-success rates, the harness, not the model, can double your cost and latency. Worth benchmarking your own stack before blaming the model.

Also worth a look

Browse previous days →