Asana makes its browser agent 76x cheaper and 5x faster using GPT-6 Astra to tune GPT-6.1 Sol
Asana's StackAI team had GPT-6 Astra in Codex run a 144-run study of its browser agent and found the agent never cached its growing page and screenshot history, and edited that history at nearly every step, so each request was resent at full price. Caching the history, keeping more text, and removing screenshots in batches of 20 cut cost on GPT-6.1 Sol from $1.97 to $0.47 per run, with 89% of input served from cache; against the original production setup on another lab's model, OpenAI reports 76x cheaper and 5x faster. The numbers come from one task (collecting fields for 32 books from a demo catalog) averaged over three runs, and the comparison models are unnamed, so treat the headline ratio as an upper bound; the transferable finding is that cache-friendly history management mattered more than the model choice.
Sources & depth
OpenAI fires three safety researchers; they dispute the misconduct claim
OpenAI dismissed safety researchers Jasmine Wang, Tomek Korbak and Mikita Balesni last week, saying they violated policy by accessing and handling sensitive company information and citing a pattern of misconduct beyond sharing material with an outside AI evaluation group. In an open letter the three deny mishandling information, say they acted in good faith, and warn the firings chill open safety discussion at the company; Wang says the access was to an executive's email that was accidentally visible in a recruiting inbox. The accounts conflict and nothing independent has confirmed either side.
Sources & depth
Sophos says OpenAI Daybreak agents now resolve about half of its managed-detection cases end to end
Sophos reports that agents built through OpenAI's Daybreak programme cut average case handling from about 38 minutes to 89 seconds (a 96% reduction) and resolve 52% of managed detection and response cases end to end, across 1,000 to 2,000 cases a day. An investigation agent gathers context and indicators, a planning model runs a plan-execute-review loop, and customers pick between notify, collaborate and automated response modes. These are vendor-reported figures from a partner case study.
Sources & depth
TypeSafe AI raises $870 million at a $7.5 billion valuation
The AI lab announced a Series A led by Andreessen Horowitz with Sequoia Capital, existing investor DCVC and angels participating; a16z's Martin Casado joins the board. TypeSafe builds models that let software make automated decisions, and its first model, Jev, is in early access.
Sources & depth
Ai2 replaces its GPU queue with time budgets and fair-share scheduling
Ai2 describes moving its research clusters from priority-based scheduling to hierarchical fair-share time budgets plus a "scheduling contract" in which a job declares a minimum runtime (up to 8 hours) before it can be preempted. Over a 30-day test, teams received 98% of the GPU hours they were owed, occupancy held at 98%, p90 wait for debug jobs fell from about 2 hours to 30 seconds, and repairs needing human intervention dropped 74%.
Sources & depth
Nathan Lambert: rapid AI progress, but not toward general superintelligence
Lambert argues gains will come mainly from engineering in training and inference infrastructure, a grounded form of AI-assisted language modeling rather than runaway recursive self-improvement. He expects coding agents to make research easier, so good ideas count for more than execution, and the cost of model intelligence to fall quickly as agents automate efficiency and pretraining research over the next few years.
Sources & depth