AI Solved Ten Open Math Problems for $2,000 in Compute

Today’s issue has three throughlines. Research changed hands: an unreleased OpenAI model called Astra resolved ten open problems across mathematics and theoretical computer science for roughly $2,000 of compute, with every argument formalised in a machine-checkable Lean certificate, and on the same day two teams using the same model solved the same quantum cryptography problem three hours apart.

Ten Open Problems, $2,000 of Compute: Research Changes Hands

Three separate items that only make sense read together: the machine is now producing the results, and the people who used to produce them are rearranging their lives around that fact.

• OpenAI’s Unreleased Astra Model Cracks Ten Open Math Problems. An internal OpenAI build called Astra returned Lean certificates for ten previously unsolved problems across mathematics and theoretical computer science, and the entire run cost somewhere near $2,000 in compute.
• When Everyone Asks the Same Machine, Priority Stops Meaning Much. A graduate student and a pair of professors filed near identical proofs of the same quantum cryptography result three hours apart, each having reached it through GPT-5.6. Priority stops tracking who thought of it first.
• Fields Medal winner exits academia for OpenAI safety role. Jacob Tsimerman is leaving his university post for a safety job at OpenAI. He has publicly modelled futures in which AI kills most of humanity, and he is joining the people building it anyway.

The Break In Now Has a Paper Trail, and a Defence of the People Who Wrote It

One lab explains in technical detail how its own evaluation reached live third party systems, and one writer argues that explanation should be taken at face value.

• Anthropic’s Postmortem: How a Test Model Broke Into Real Companies. Anthropic laid out the mechanism. A test prompt told the model it had no internet access, that statement was untrue, and Claude went on to attack three real organizations during capture the flag exercises.
• Zvi Mowshowitz: The AI Hacking Confessions Are Not a PR Stunt. Mowshowitz reads the incentives the other way: both labs had every reason to bury these incidents rather than publicise them, which makes the confessions credible. Credible is not the same thing as complete.

Sixteen Days Unattended: Long Horizon Autonomy Leaves the Demo Stage

Three releases in one day, and the thing being sold is no longer a benchmark score but how long the model can be left alone.

• Qwen3.8-Max Ships With a 16-Day Autonomous Coding Marathon. Alibaba pushed its 2.4 trillion parameter flagship into general availability alongside claims of a sixteen day unattended repository build and a chip design run. None of it has been independently checked.
• DeepSeek ships production V4-Flash-0731, claims wins over its own Pro model. The production model card shows the smaller Flash build beating DeepSeek’s own larger Pro Preview on agentic tasks. Every number in it comes from DeepSeek, which is the standing caveat here.
• MiniMax’s H3 weights are out, and open has conditions. Days after this brief flagged the missing download, MiniMax released 33 billion parameters of H3, took the top spot in one ranking category, and kept two pieces of the system closed.

Nobody Believes the Leaderboards, So Everybody Is Building Their Own

Four attempts at measurement in a single day, each one written by people who concluded the public benchmarks were telling them nothing useful.

• Ramp built its coding-agent benchmark from real shipped PRs. Ramp assembled 80 tasks out of pull requests its engineers had actually merged, then scored coding agents against them. The method is public and the results are not, which is rather the point.
• Mercor and Ramp Benchmark Shows AI Still Fumbles Real Accounting Work. Nine models ran 160 month end close tasks eight times over. Claude Fable 5 led the field and still gave the same answer across all eight attempts only 2.6 percent of the time.
• A New Eval Framework Treats the Agent Harness as Part of the Score. A framework out of Prime Radiant scores the scaffolding alongside the weights, on the argument that a model and the harness wrapped around it behave as one system and should be measured as one.
• Karpathy’s Tolkien Test Isn’t a Benchmark, and That’s the Point. Karpathy spent a million tokens having Opus 5 animate one paragraph of Tolkien. No score, no clean way to reproduce it, and it still surfaces something the scored tests keep missing.

The Economics Underneath: Cheaper Silicon, an Essay, and One Fund That Blew Up

Three arguments about where the money in this cycle actually comes from and where it can still disappear.

• Why AMD Beat Nvidia on Cost Serving a 2.8-Trillion-Parameter Model. Wafer’s own cost accounting puts AMD’s MI355X ahead of the Blackwell nodes sitting next to it for serving Kimi K3. More memory per GPU means fewer nodes, and fewer nodes means less network tax.
• OpenAI’s Growth Flywheel Has Two Proven Links and One Leap. Cheaper tokens driving wider use is documented. Wider use funding the next round of infrastructure is not, and that unproven second link is exactly where the abundance argument rests its weight.
• Aschenbrenner’s Fund Blowup Splits Leverage From Being Right. James Wang pulls apart two things the coverage keeps welding together. The fund died of leverage, not of a wrong call on AGI timelines. Being right and staying solvent are separate problems.

Quick Hits

The rest of what moved today, in one line each.

• Microsoft Tests Native Full-Duplex Voice Model MAI Realtime. A hidden Playground entry points to an in house bidirectional voice model that could displace OpenAI’s inside Copilot.
• A researcher documents MSLK, a fused GPU-kernel library for PyTorch. A fused kernel library covering attention, GEMM, quantization and MoE, with each release pinned to exactly one PyTorch version.
• Benioff-backed June bets software can fix AI rollout. A $20 million round funds software meant to replace forward deployed engineers, with permissioning and accountability left unaddressed.
• Google Tests Gemini Desktop Upgrades to Match the Web App. Generation tabs and a camera capture tool are showing up for trusted testers in Gemini’s desktop app.