Anthropic acquires Stainless, adds KPMG's 276,000-person workforce

Anthropic acquired Stainless on May 18 — an API tooling company that helps developers build and maintain SDKs — the same week it announced a strategic alliance with KPMG to deploy Claude across the firm's full global workforce. The back-to-back moves signal Anthropic is hardening its developer infrastructure while simultaneously landing the kind of professional-services mega-deal that validates enterprise scale.

OpenAI ships content provenance standard across image generation

OpenAI published a content provenance initiative on May 19 aimed at making AI-generated images traceable at the ecosystem level, with cryptographic metadata embedded at generation time. The move comes as the TAKE IT DOWN Act and EU deepfake rules pressure platforms to prove origin — giving OpenAI a compliance story and a trust differentiator ahead of the ChatGPT Images 2.0 rollout.

xAI launches Grok Connectors, bringing third-party apps into Grok

xAI shipped Grok Connectors on May 6, enabling deep integrations with everyday productivity and business apps directly inside Grok, with mobile support for iOS and Android described as coming soon. The launch moves Grok from a standalone chat interface toward an ambient work layer, a direction every major lab is now racing to own.

Mistral ships Medium 3.5 and moves Vibe coding agents fully to the cloud

Mistral released Medium 3.5, a 128B-parameter dense model that folds chat, reasoning, and coding into one system with a configurable reasoning-effort dial, while simultaneously launching remote async agents for its Vibe coding tool — sessions now run in isolated cloud sandboxes, open PRs on GitHub, and notify developers when done. The unified model scores 77.6% on SWE-Bench Verified and replaces three separate prior models in Le Chat and the Vibe CLI.

Four Chinese Open-Weight Models Hit Frontier Coding Ceiling in 12 Days

GLM-5.1, MiniMax M2.7, Kimi K2.6, and DeepSeek V4 all landed within the same agentic engineering capability band as Western frontier models—none costing more than a third of Claude Opus 4.7. Kimi K2.6 independently scored 58.6% on SWE-bench Pro, matching GPT-5.5 and Gemini 3.1 Pro on that benchmark, while its Agent Swarm architecture scales to 300 parallel sub-agents. The sprint confirms that open-weight models have closed the agentic coding gap at a fraction of closed-model inference cost.

KellyBench Shows Every Frontier Model Fails Long-Horizon Decision Tasks

KellyBench tasked agents with managing a bankroll across a 38-week Premier League season using historical betting data—a structured, long-horizon planning problem with an objective verifier. Every frontier model finished in the red on average, with only 3 of 24 model-seed combinations avoiding ruin; even the top performer, Opus 4.6, hit a sophistication score of just 32.6%. The result is a concrete counter-signal to SWE-bench-era optimism: current benchmark suites overstate agent capability by assuming clean specs and deterministic verifiers.

MIT CSAIL's Engram beats human-designed algorithms by chaining fresh-context agents

Engram fixes two failure modes in long-horizon LLM research agents — evolutionary neighborhood bias and the 'coherence ceiling' caused by context rot — by organizing exploration as a sequence of agents that each begin with a fresh context window and read a compact Research Digest distilled from the prior agent's runs. On multi-cloud multicast routing, Engram found a dynamic-programming solution costing $622, beating the reported human state-of-the-art of $626 and all evolutionary baselines; it also cut mean LLM inference request-routing response time to 23.9s versus the 25.7s expert-designed baseline. The architecture requires no weight updates and is fully open-sourced.

NVIDIA's SANA-WM generates coherent one-minute videos with a 2.6B hybrid diffusion transformer

NVIDIA researchers built SANA-WM, a 2.6B-parameter open-source world model that combines a linear diffusion transformer with a hybrid architecture to produce high-fidelity, minute-scale video — a duration that has been a hard wall for prior diffusion models constrained by quadratic attention costs. The open-source release includes weights and code, making it immediately usable for robotics sim-to-real and embodied AI research that requires temporally coherent long-horizon rollouts.