OpenAI shipped a voice model built to handle interruptions, then admitted nearly a third of its own coding benchmark is broken. Anthropic tested deleting dangerous knowledge from a trained model, and two funding stories show where the capital is actually going.
Two New Models, One Shared Bet on Speed: Voice and Code Get Cheaper Frontiers
OpenAI and Cognition each shipped models that trade raw benchmark supremacy for something more useful in production.
- OpenAI Launches GPT-Live, a Voice Model That Talks and Listens at Once. The model handles interruptions and backchannel cues mid call, then quietly routes anything that needs real reasoning to GPT-5.5 without the caller noticing a handoff.
- Cognition’s SWE-1.7 Chases Frontier Coding at Lower Cost. SWE-1.7 trails Opus 4.8 on every benchmark Cognition published but lands close to GPT-5.5, turning the coding model race into a fight over price per token rather than raw capability.
When the Yardstick Breaks: Benchmark Trust and Knowledge Control
Two research efforts this week attacked the infrastructure underneath AI progress: what models get measured against, and what they are allowed to know.
- OpenAI Finds Roughly 30% of SWE-Bench Pro Tasks Are Broken. OpenAI’s own audit found close to a third of the coding benchmark’s public tasks unsolvable, mislabeled, or broken by environment errors, a problem for every leaderboard built on top of it.
- Anthropic tests a way to delete dangerous knowledge from AI models. GRAM lets one trained model have its virology, cybersecurity, or nuclear physics knowledge surgically removed after the fact, no retraining required, a step toward selective forgetting as a safety tool.
Where the Capital Is Going: Chips and Consulting
Two funding stories this week point in different directions: a chip challenger raising toward an IPO, and a lab buying its way into enterprise services.
- SambaNova raises $1 billion, hits $11 billion valuation. General Atlantic led the round, pushing SambaNova to an $11 billion valuation as investors keep funding challengers to Nvidia’s inference chips, with a 2027 IPO now on the table.
- OpenAI buys Northslope to staff its enterprise services arm. The deal is OpenAI’s second applied AI acquisition in two months and adds hundreds of embedded engineers, a clear signal the company wants a Palantir style services business, not just an API.
Open Data and Bigger Asks: How Agents Actually Improve
Away from the model releases, the sharper argument this week was about inputs: what data agents train on, and what people actually ask them to do.
- NVIDIA Pushes Open and Synthetic Data as the Real Agent Bottleneck. NVIDIA’s case is that agent transparency depends on open training data, and synthetic data is the only realistic way labs share it without handing over their competitive edge.
- You’re Not Ambitious Enough With Claude. The real waste in frontier AI use is not cost, it is scope: asking Claude to autocomplete one line when it could run the entire migration.
Today’s Quick Hits
- ByteDance’s Seedream 5.0 Pro Bets on Editing, Not One-Shot Images. ByteDance’s new Seed model prioritizes precise, layer by layer design work over single prompt output, with native text rendering in ten or more languages.
- A New Taxonomy Splits Self-Improving Agents Into Three Real Layers. Researcher Shilong Liu’s framework separates prompt tweaks, infrastructure changes, and actual weight updates into three distinct tiers of agent self improvement.
- Google Photos Gains AI Video Editing With Video Remix. Gemini Omni now powers in app relighting, background swaps, and artistic filters for Google Photos subscribers.
- Mistral’s Robostral Navigate Steers Robots With One Camera. The 8B model beats multi sensor rivals on a room navigation benchmark using only RGB video, no lidar or depth sensor required.