New model releases, benchmarks, evaluation methods, and research results.
Encrypted reasoning blocks from Anthropic, OpenAI, and Google replay across models and accounts. Researchers used Haiku 4.5 to transcribe Opus 4.8 reasoning, then decrypted 315,320 blocks from public repos to find 62 API keys and 33 passwords.
On August 1 OpenAI published proofs for 10 math and TCS problems open for a decade or more, produced by an internal version of the unreleased Astra model. It shipped a 249-page paper plus Lean 4 certificates, priced the tokens at roughly $2,000, and left the repo marked agent-reviewed.
Bottleneck Labs handed a GPT-5.6 Sol agent a live App Store app and $350 in cash. After 24 hours it had booked no new revenue, cut the price six times, and ended at negative $97.
Nvidia and Microsoft led an open-weights letter now signed by 132 companies, and Anthropic is the one US frontier lab missing. Amodei countered with chip export controls, industrial-scale distillation enforcement, and mandatory pre-release safety testing for every model.
Google released DiffusionGemma as an Apache 2.0 open model, pairing a 256-token diffusion canvas with vLLM support and an 18GB VRAM deployment target.
Anthropic released Claude Fable 5 and Mythos 5, splitting the same base model across safeguards, restricted access, pricing, and cloud retention rules.
AWS added an optimized Step Functions integration for AgentCore harness calls, turning agent reasoning into a workflow Task with approval, retry, trace, and token controls.
NetFoundry launched MCP and LLM gateways that bind agent tool access to identity instead of public ports, VPN sprawl, and shared API keys.
Microsoft introduced MAI-Thinking-1 and seven in-house MAI models, bringing 35B active reasoning, 256K context, and Foundry private preview into focus.
MiniMax M3 arrives with a 1M-token context window, agentic coding support, and an open-weight promise. The launch separates API access from verification work.
NVIDIA introduced Nemotron 3 Ultra alongside NemoClaw, OpenShell, and CUDA-X agent skills, pushing open agent competition into the runtime layer.
NVIDIA Alpamayo 2 Super ties a 32B VLA teacher model to AlpaGym, OmniDreams, CoC auto-labeling, and agent skills for L4 robotaxi development.