OpenAI claims a math proof, walks back its launch, and the model race turns to cost

OpenAI says GPT-5.6 Sol Ultra proved the Cycle Double Cover Conjecture, a graph-theory problem open since the 1970s, by running 64 subagents in parallel for just under an hour, then posted the prompt and the proof for anyone to check. Whether that proof holds is now a job for mathematicians rather than marketers, and that verification is the thing to watch. The claim surfaced during the same day OpenAI spent walking back how it shipped the model. Tibor Sottiaux posted an unusually plain apology: the ChatGPT Work and Codex superapp reorg, covered here Friday, buried familiar chats and projects, nudged people toward the most expensive reasoning settings without making the cost clear, and broke some existing multi-agent workflows. The fix that shipped hours later was blunt. OpenAI reset usage limits twice in one day, changed the defaults that steered users to high compute, and reassured Codex holdouts that Codex “is here to stay.” The pitch underneath the mess is real, though: OpenAI claims GPT-5.6 Luna beats GPT-5.5 at its top reasoning setting while costing 25x less. On NLW’s read in the AI Daily Brief, efficiency, not raw capability, is now the axis the model race is being fought on.

Cursor shipped into the same day and aimed at the problem the superapp reorg fumbled: keeping long agent work legible. It added side chats, durable side conversations you can start without derailing the main thread and then @-mention to pull their context back in, plus local full-text search across thousands of past agent transcripts and new cloud-agent hooks. Across both companies the through-line is the same, and it is not the model: context management is becoming the product surface.

GitHub published the sharpest agent-engineering writeup of the day, and the finding runs against instinct. Swapping Copilot code review onto better, shared tools made it worse. The team replaced code review’s bespoke file tools with the same grep, glob, and view that power the Copilot CLI, expecting a clean upgrade, and instead measured higher cost and fewer issues caught. The tools were fine. The instructions carried the CLI’s instincts: browse broadly, guess paths, accumulate context, all reasonable for an interactive coding assistant and wrong for a reviewer, who should start from the diff and read only the narrowest evidence that confirms or kills a concern. Rewriting the tool instructions for a reviewer’s workflow, narrow with grep and glob first and call view only once the file and line range are known, flipped the regression into roughly 20% lower average review cost at the same quality. The transferable lesson: for an agent, tool descriptions are API docs, and a tool result is not a printout, it is tokens that stay in the context window and shape every later decision.

The agent-security surface got a little more instrumented in the same window, against a worse backdrop. GitHub’s CodeQL 2.26.0 added detection for AI prompt injection alongside Kotlin 2.4.0 support, so the static analyzer behind code scanning can now flag injection paths in agent code. New VentureBeat research put the exposure in numbers: 69% of enterprises are running agent fleets on shared API keys, one leaked credential away from a fleet-wide compromise. And Semgrep argued you can’t reverse-engineer trust into an open-weight model the way you can audit source, so provenance and independent auditing, not inspection, are the levers that matter. Three companies, one message: the agent supply chain is under-secured at exactly the moment fleets are scaling.

Against the “AI employees” framing that ChatGPT Work leaned on, Simon Willison pushed back: calling these tools employees is “both disrespectful to humans and a complete misunderstanding of what these tools can do,” and you “may as well start adding Excel spreadsheets to your org chart.” It reads as a quip, but the point is load-bearing. Org-chart framing sets the wrong defaults for how you scope, supervise, and grant credentials to an agent, which is the exact gap the day’s security numbers expose.

The model-loyalty churn cut the other way for Anthropic. Claude Code users spent the day working out that Fable 5 leaves after July 13 and weekly limits get cut 50%, after days of complaints that Fable and Opus 4.8 had degraded and a temporary access extension that ran through today. Part of that thread is people already pricing a move to Sol. That is the pressure to watch into next week: with Luna claiming 25x cheaper and Fable’s limits tightening, near-term switching is being decided by cost per token and usage caps, not benchmark deltas, and whether OpenAI’s proof claim survives outside review will say more about the ceiling than any of it.