AI Model Roundup: June 9–15, 2026
One story dominated the week: Anthropic launched its most capable public model on Tuesday and had to switch it off on Friday under a US export-control directive. Around it, Google released an experimental text-diffusion model and Moonshot shipped a coding-focused Kimi with only self-made benchmarks.
Fable 5: launched June 9, offline June 12
On June 9, Anthropic released Claude Fable 5, the first public model in its new Mythos class, positioned above Opus. Launch numbers included 80.3% on SWE-bench Pro, a 1M-token context window, and $10/$50 per million tokens, twice the price of Claude Opus 4.8. A research sibling, Mythos 5, shipped as a restricted preview.
At 5:21 p.m. ET on June 12, a letter from Commerce Secretary Howard Lutnick directed Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national, inside or outside the US. Anthropic can't verify nationality in real time, so it disabled both models for all customers. Other Claude models were unaffected. The directive followed a public claim by the red-teamer Pliny the Liberator that he had bypassed Fable 5's safeguards; Anthropic called it a "narrow, non-universal jailbreak" and argued that applying the same standard across the industry "would essentially halt all new model deployments for all frontier model providers."
It is the first time the US government has used export controls to take a live, broadly available model offline. Full account: The US Government Shut Down Claude Fable 5 in 3 Days.
Google DiffusionGemma
On June 10, Google released DiffusionGemma, an experimental open model that generates text by diffusion instead of token-by-token decoding. It is a 26B mixture-of-experts with 3.8B active parameters, under Apache 2.0. Google claims up to 4x faster output on dedicated GPUs: more than 1,000 tokens per second on a single H100 and more than 700 on an RTX 5090.
Diffusion language models trade some quality for large speed gains, and this is the first one from a major lab with open weights at this size. It is worth testing for latency-bound tasks such as autocomplete, classification, and drafting, not as a general replacement.
Kimi K2.7 Code
On June 12, Moonshot released Kimi K2.7 Code, a coding-specialized variant of Kimi K2.6: 1T total parameters, 32B active, 256K context, open weights under a modified MIT license, and $0.95/$4.00 per million tokens at launch. Every launch benchmark came from Moonshot's own suites, with no SWE-bench, Terminal-Bench, or LiveCodeBench numbers. Analysis: Kimi K2.7 Code and its self-made benchmarks.
Score changes
No LMRank scores changed this week. Fable 5 will be scored once it is available and independent results exist.
Models mentioned: Claude Fable 5 · Claude Opus 4.8 · Kimi K2.7 Code · Kimi K2.6
See also: Best Coding Models · Best Open-Weight Models
Takeaway
Regulatory risk is now a deployment risk. If a product depends on a single frontier model, keep a tested fallback from a different provider behind the same interface. Teams that could switch from Fable 5 to Opus 4.8 on June 12 lost an afternoon; teams that couldn't lost their product until the model came back.
Update: Commerce lifted the directive on June 30, and Anthropic restored Fable 5 on July 1.