Google lets every engineer use Claude Opus 5 inside Antigravity
Rivals keep trading places today: Google opens its doors to Anthropic's best coding model, a Chinese lab sweeps the voice leaderboards, and SemiAnalysis maps where a robot's brain should live.
Google has given staff across the company access to Anthropic's Claude Opus 5 for internal work — but only inside Antigravity, Google's internal version of its development platform, according to Business Insider's report. A Google spokesperson said "Gemini remains our primary and foundational model for internal development, with third-party models available on a quota to support specialized use cases," adding that Claude comes with per-user quotas and is meant to be supplementary. DeepMind teams have had Claude access for some time and some employees had secured it for high-priority projects, per three people with knowledge of the matter. The concession lands where it stings: despite recent gains from its Flash 3.8 model, Google still trails Anthropic and OpenAI on coding, and Googlers had grown frustrated with Gemini's limits. The direction of travel is the story — Palantir and Nvidia moved the opposite way hours ago (restricting Claude over data fears), while Google just widened the door. After spending billions pushing its workforce onto Gemini, the company has conceded that the best coder in the building isn't its own model — quota or no quota.
StepFun shipped five StepAudio 3 voice models at once — Realtime, ASR, TTS, Gen and Music — and the Shanghai lab says several topped Artificial Analysis leaderboards globally. The company claims the ASR model ties for first on non-streaming accuracy with a 1.7% word error rate, while Realtime scores 98.9% on conversational dynamics and 99.7% on speech reasoning. Realtime runs native full-duplex conversation — listening, reasoning and calling tools at the same time rather than pausing to think — and Gen unifies voice, sound effects, ambient audio and music under one audio representation. All five are live on StepFun's open platform today. The signal isn't another TTS endpoint; it's a Chinese lab bundling recognition, dialogue, production and composition into one workflow and claiming global firsts in the cheapest category to win.
SemiAnalysis published a deep dive on where a robot thinks — on-device or in the data center — covering robot model types, silicon and DRAM efficiency, Jetson Thor versus B300 total cost of ownership, deployments, and what the authors call "the network wall." The framing borrows Boston Dynamics' split: a System 2 planner breaks a factory work order into steps simple enough to act on, while a System 1 vision-language-action model turns a camera view plus one short instruction into motor actions, onboard, because motion carries hard latency and determinism requirements. The economics fall out of that split: the planner is a data-center workload, the reflex is a chip problem, and the network between them decides which robots ship.
What to watch: whether Google's Claude quota spreads beyond Antigravity — or Gemini 4 closes the coding gap first.
Would you let a rival's model write your company's code if it were measurably better? Tell us in the comments.
