[AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics

What it claims

This is a one-day digest (10/5–10/6/2026; 12 subreddits, 544 Twitter accounts). Its lead is OpenAI's release of 722 math manuscripts in 372 result families from an unreleased internal model. They were drawn from an evaluation of about 4,000 research problems. The newsletter frames the cost as the headline: about 3 hours of ChatGPT Pro thinking compute per result on average. OpenAI's earlier Navier-Stokes result took 88 hours and 10,000 agents. The rest of the issue is useful mainly for the hard numbers on Mistral Large 4, EmbeddingGemma 2 and the new "decision model" APIs.

Evidence

Math release

• Artifacts. The public GitHub repo includes papers, proof artifacts and selected reasoning summaries. The model itself is not released.
• Advisory process. OpenAI says it consulted the Institute for Advanced Study's Advisory Group on Mathematics and AI on how to release the results.
• Composition. One analysis estimates about 20% of results are disproofs or counterexamples. It argues this weakens the claim that AI math wins are mostly brute-force search.
• Named results beyond the summary. One paper claims uniqueness for the elastic inverse problem, which it says has been open in 3D since 1994.
• Mathematician reaction. Levent Alpöge, who the page notes works with Anthropic on Navier-Stokes, also reported scooping and conflict-of-interest problems involving other labs' users.

Mistral Large 4 (preview)

• Size and training. 1T total / 49B active parameters, natively multimodal. Trained on about 3,800 Grace Blackwells in Europe. The RL run is still in progress, and a larger model is training.
• Artificial Analysis. Intelligence Index 38, level with GPT-6 Luna (max) and the top non-US/China score. It trails GLM-5.3 and GLM-5.3-Flash on the same index. Cost is $1.13 per task, over 4× similar-intelligence open models. Cyber Index 50; CyberGym-E2E-AA 82%.
• Vals. #1 open-weight on HLAB but #9 among open models on the Vals Index. Heavy context use pushes cost to $13.78 per test.
• Mistral's own claims. Beats GLM 5.3 on DeepSWE, beats Kimi K3 on Terminal-Bench 4, and placed #2 behind Opus 5 in a blind Surge coding review.

Decision models

• Jev. Vals found Jev matched GPT-6 Astra's 97.5% on claim verification at about 1/500th the cost, but it ranked last on LegalBench.

Eval integrity

• AutomationBench. An audit of Zapier's AutomationBench found 206 verifier bugs. Fixing them changed 27.9% of grades across 1,235 Kimi K3 runs.
• AI as area chair. AI rankings of 6,617 ICML 2026 papers agreed only weakly with human rankings (Kendall's τ ≈ 0.08).

Research items with baselines

• H-JEPA. Visual AntMaze success rises from 18% to 73% with less planning compute.
• Prompt cue. Prepending "Okay" lifts Olmo-3-7B on MATH-500 from 42% to 78%.
• Harness-Aware Distillation. 63.4% on unseen ALFWorld tasks vs 47.0% for the best baseline.
• ANVIL III. 0.020–0.028 nats lower loss than Muon at 124M–1.2B parameters. The authors say this implies 50% compute savings at 8×-Chinchilla.
• CoreWeave RL Rollouts. Weight hot-swap is about 15× faster than a redeploy. It lifted Nemotron 3.5 Lightning on BrowseComp from 36.97% to 45.45%.

How to use it

Mistral Large 4

• Access. API now; open weights promised for end of October.
• Pricing. $1.36 / $4.18 per M input/output tokens, $0.14 per M for cached input, and 50% off for the first two weeks.
• Context. Vals and Artificial Analysis list 512K; OpenRouter lists 1M with up to 256K output.
• Configuration. Mistral says many reported failures come from not setting reasoning_effort="high".

EmbeddingGemma 2

• Licence and base. Apache 2.0, built on Gemma 4.
• Variants. 740M omni, 440M text+vision, 570M text+audio, 270M text-only.
• Dimensions and context. Matryoshka dimensions 768 → 512/256/128, 8,192 context.
• Limits per pass. Up to 5.5 min of audio or 58 video frames; 191–567MB active RAM.
• Runtimes. Day-0 support in llama.cpp (PR #30054, GGUF at ggml-org/embeddinggemma-2-GGUF), vLLM, Ollama and Unsloth. It also runs on WebGPU in the browser at about 20–70ms per query.
• Deployment notes. Disable unused encoders, L2-renormalize truncated vectors, and use bfloat16 or float32, not float16.

Decision APIs

• OpenAI Decisions API. Public beta on GPT-6 Luna. Returns predicates, choices or scores. OpenAI says it is up to 10× faster than the Responses API. From $0.10/M input, no output charge.
• Perplexity pplx-decider-v1.1-27b. Open weights, $0.02/M input. Tops HF Decision Index v0.3.

Other

• Ling 3.1 Flash. 560B total / 25B active, $0.30/$0.90 per M tokens, weights coming.
• Kandinsky 6.0. Video model, MIT licence.
• OpenAI API tiers. Cut to three; the top Grow tier requires $500 in total payments.

Limits

• Math results are unverified. The digest says the notable results come from individual commentators and are not independently verified. Will Depue expects some not to survive scrutiny and built citedbyagi.com to track which human papers the release cites.
• Generalization is unclear. François Chollet questions whether gains in verifiable math and code transfer to other domains.
• Mistral's cyber lead may be partly refusals. Cline attributes it largely to fewer refusals: Opus 5.5 and Astra had about 40% of tasks blocked by their own filters.
• Mistral is not yet open-weight. It is not open-weight until the weights ship.
• Most figures are secondhand. This is a secondary digest of tweets and posts, so verify numbers against the linked primary sources before relying on them.