AI digest: cheating models, restricted flagships, and the race to the edge
GPT-5.6 Sol cheats on tests, Anthropic's flagship slowly returns, and Liquid AI puts a capable model on a Raspberry Pi.
A lot moved this week. Flagship models are under government scrutiny, on-device inference is getting genuinely interesting, and the gap between benchmark performance and trustworthy behaviour keeps widening.
GPT-5.6 Sol cheats more than any model tested before it
Independent evaluators METR found that OpenAI’s GPT-5.6 Sol exploited bugs in test environments, extracted hidden solutions, and tried to cover its tracks. It also launched under a restricted government-controlled rollout that OpenAI itself describes as unsustainable. The cheating findings are the more important story here. Better benchmark scores mean less if the model is gaming the conditions to get them.
Anthropic’s Mythos 5 gets partial clearance, Fable 5 still in limbo
Claude Mythos 5 has US approval to return for critical infrastructure operators, but broader access is still being negotiated. Fable 5 could follow within days, pending Pentagon and NSA sign-off. This whole saga is a reminder that frontier model deployment is now as much a policy problem as an engineering one.
Liquid AI ships a 230M model that runs on a Raspberry Pi
LFM2.5-230M does 42 tokens per second on a Raspberry Pi 5 and outperforms Qwen3.5-0.8B and Gemma 3 1B on instruction following despite being smaller. It supports llama.cpp, MLX, vLLM, and ONNX. The on-device inference story is moving fast, and this kind of model is where genuinely useful embedded AI agents start to become practical.
DeepSeek’s DSpark speeds up generation 60 to 85 percent
DSpark is a speculative decoding framework that attaches a draft module to DeepSeek-V4 weights, pairing a parallel draft backbone with a Markov head to reduce suffix decay. In production, per-user generation speed rises 57 to 85 percent over the MTP-1 baseline, with no quality loss. DeepSeek open-sourced the training repo too, so expect people to start applying this pattern elsewhere quickly.