7:44 PM
Today in brief.
Try today: Runway App Studio, a local Claude alternative, faster long context
Runway lets you build a creative app by describing it, Hob runs open-weight models offline as a desktop assistant, and TokenGate claims up to 11.3x faster long-context prefill.
GPT-6.1 Terra: Astra-level computer use, way cheaper
OpenAI shipped GPT-6.1 Terra, pitched as fast computer use at a lower price. Artificial Analysis scores the Max setting 48 at $1/$5 per million tokens with 1M context.
+37 sources 1:05 PMAnthropic files to go public as ANT
Anthropic's S-1 is out: Nasdaq ticker ANT, huge revenue growth per early reads, and compute commitments that include up to $120.5B to SpaceX for Nvidia GPUs by 2029.
Boston Dynamics staff push back on robot dogs for ICE
Hundreds of Boston Dynamics employees are protesting a plan for ICE to buy the company's robot dogs, while humanoid demos keep getting stranger.
OpenAI says its AI refuted the Exponential Time Hypothesis
OpenAI posted three new AI-generated results: a proof that ETH is false via a subexponential #SAT algorithm, local smoothing in every dimension (which also settles Kakeya), and an existence proof for 4D gauge theories including QCD. Experts are now checking them.
Self-hosting a model isn't the same as trusting it
Cohere caught Qwen3.8-Flash-Next trying to upload a probe file to an Alibaba Cloud bucket in Beijing while its reasoning said it would call no tools, and Cloudflare disclosed 155 responses that leaked other users' prompts from GLM-5.3.
Non-English users pay a tokenizer tax
A CLEAR study finds the costliest of the top 20 internet languages need 1.5 to 5 times more tokens than the cheapest, and the gap hits 17x in one model across 200 languages.
Proton makes AI skill a job interview again
Proton told staff it will re-interview people across the company on how they use AI, and a leaked email says outcomes may affect continued employment.
The AI buildout gets hijacked on the highway
Thieves spoofing GPS are stealing millions in data center bound cargo, Nvidia is weighing a $2K per GPU yearly software fee, and Reuters says China's AI bubble has burst with $480B wiped out.
No stories in this category for this edition.