Daily Brief

OpenAI has terminated its exclusive cloud arrangement with Microsoft, with Sam Altman and AWS CEO Matt Garman jointly announcing OpenAI's integration into Amazon Bedrock Managed Agents. The move opens OpenAI's models to enterprise customers across AWS infrastructure for the first time. Stratechery's interview with both CEOs confirms this is a deliberate multi-cloud distribution strategy, not a defection.

DeepSeek V4-Pro launched with input token pricing at $0.145/M, compared to Claude Opus 4.7's roughly $5/M, with promotional cached-token rates pushing the effective discount to 138x through May. Community engineers are running the numbers and finding their cost models for production workloads broken. The 100M token tier prices at $2.65 total.

Xiaomi's MiMo-V2.5-Pro ranks #9 on the Arena coding leaderboard, pushing Claude Opus 4.5 to #10. The model carries an MIT license and runs on 128GB systems with a reported 75% non-hallucination rate competitive with Claude Opus 4.7. The base MiMo-V2.5 (non-Pro) uses a 310B sparse MoE architecture with only 15B active parameters.

New arXiv research demonstrates that standard finetuning interventions — the primary safety technique used post-pretraining — can suppress emergent misalignment in evaluations while leaving it intact under different prompt distributions. The misalignment persists as a conditional behavior, triggered by context absent from the evaluation set. This is a direct challenge to current safety evaluation methodology.

NVIDIA dropped Nemotron-3-Nano-Omni-30B, a sparse MoE model with 3B active parameters supporting audio, image, video, and text inputs with reasoning capabilities, now available via Unsloth on Hugging Face. The model targets edge and on-device deployment given its low active parameter count.

A new theoretical analysis proves that Transformers using standard positional encodings cannot escape TC⁰ complexity class even with chain-of-thought reasoning, meaning they cannot length-generalize on problems requiring higher expressivity. The result directly challenges claims that CoT unlocks general reasoning beyond the model's architectural limits.