OpenAI announced 722 math manuscripts in October 2026—claiming solutions to Fields Medal-level problems and Millennium Prize conjectures. Here's the catch: only 42% are formally verified. For founders deciding whether to build math-heavy products on OpenAI capabilities, that gap between announcement and verification is the central question.
The model: GPT-5.2 and GPT-5.2 Thinking—a specialized variant trained explicitly for proof generation across algebraic geometry, number theory, combinatorics, topology, and applied math. On expert-level problems (FrontierMath benchmark), the model solves 40.3% correctly. On graduate-level questions (GPQA Diamond), it hits 93.2%. But "correct solution" in OpenAI's testing means peer-reviewed, not formally verified.
Here's what OpenAI's math actually enables, what remains unverified, and when to use this for production.
The Benchmark: FrontierMath and Real Verification Gaps
FrontierMath is a 222-problem benchmark of unsolved or extremely difficult research problems. OpenAI's 40.3% success rate means GPT-5.2 solves 90 of those problems—a significant jump from earlier models. GPQA Diamond, a public multiple-choice benchmark of grad-level questions, shows 93.2% accuracy (up from GPT-4's 88%).
But here's where the story changes. Of the 722 manuscripts submitted for formalization in Lean (a formal proof language), only 42% verified (roughly 303) have been formally verified. The remaining 58% remain as natural language proofs awaiting formalization or showing issues during the process.
Translation errors exist between reasoning and formalization. The most cited example: a claimed partial solution to Navier-Stokes stability (one of the Millennium Prize problems). The model's natural language proof appears sound to mathematicians, but Lean formalization revealed missing lemmas and unstated assumptions about boundary conditions. When formalized, it covers only a special case, not the general problem.
This doesn't invalidate the work—special cases are valuable to research. But it proves the gap is real and consequential.
What Changed: GPT-5.2 Thinking's Extended Reasoning
Before October 2026, OpenAI's math was strong on computation but limited on proof generation. GPT-5.2 Thinking uses extended reasoning—the model generates internal reasoning chains before outputting a proof, mimicking how human mathematicians work: explore approaches, dead-end, backtrack, try again.
The cost: 3-5 minutes of inference per problem at ChatGPT Pro compute levels. This extended reasoning produces more rigorous proofs than base GPT-5.2, but it's expensive at scale.
Why the 722 manuscripts matter: They represent outputs tested against curated open problems—some published competition problems, others from active research. The model generates natural language proofs, which are then independently formalized in Lean by humans or automated tools.
Only 10 of 722 reasoning chains were published. Despite advisory standards from Princeton's Institute for Advanced Study calling for transparency on all reasoning, OpenAI shared internal reasoning for fewer than 2% of manuscripts. Researchers using GPT-5.2 see a proof but not how the model arrived at it.
Real-World Performance: Where It Works, Where It Fails
On multi-step reasoning problems, the model excels. Optimization, combinatorics, applied calculus—GPT-5.2 handles the mechanical steps that humans find tedious. A structural engineer using your tool to optimize beam design; the model generates parametric solutions; the engineer verifies feasibility. The model excels at intermediate reasoning.
On novel problems outside the training distribution, success rates drop sharply. The 40.3% rate applies to FrontierMath's curated set. Real-world problems may be messier, less structured, and harder. Success rates on your domain could be 20-30%, not 40%.
Text watermarking remains weak. Short content, factual responses (where there's less room to alter generation), heavily rewritten passages, and translated material show degraded detection. If your product analyzes AI-generated writing, text verification is unreliable.
Problem interpretation remains human work. The model solves well-formulated mathematical problems. Translating customer requests into correct mathematical formulations still requires human expertise. If your product is customer-facing, interpretation errors upstream make perfect math downstream useless.
Use Cases: When OpenAI Math Wins
Proof generation with verification downstream. You're research-facing (an academic tool, publishing platform, or mathematician knowledge base). Generate candidate proofs using GPT-5.2, then route them to formalization or expert review. The model becomes a starting point, not a finished product. This workflow works because peer-review and formalization are already built into research validation.
Multi-step reasoning without human derivation. You're solving moderately complex problems—optimization, combinatorics, applied calculus—where user time is valuable. GPT-5.2 generates step-by-step solutions explaining intermediate reasoning, saving hours of derivation work.
Exploring solution space for known problems. You're tackling a problem with a known answer but unknown derivation. The model generates multiple approaches, and you verify correctness against the answer. Example: An AI tutoring product where students see multiple proof strategies for the same theorem; the model generates approaches; correctness is automatically verified.
Problem decomposition and interpretation. Your product ingests informal problem statements (customer requests, research abstracts, bug reports) and needs to structure them mathematically. GPT-5.2 Thinking excels at breaking a vague problem into formal definitions and subproblems.
Internal QA for AI-generated proofs. Your product uses Gemini or another AI internally. You want to prevent hallucinations from reaching users. Checking for a clean watermark can be part of a QA workflow—if output lacks the expected structure, something went wrong.
Pricing: What It Actually Costs at Scale
ChatGPT Pro access: $20/month; includes GPT-5.2 Thinking for extended reasoning. Good for exploration and low-volume use.
OpenAI API (GPT-5.2): $50-100 per 1M input tokens (price not yet publicly finalized for math variant). Scaling linearly with problem volume.
OpenAI API (GPT-5.2 Thinking): Not yet available on API; enterprise early access only, pricing TBD.
Formal verification (Lean integration): Free (Lean is open-source), but human time for formalization costs $50-200 per proof at expert rates.
Real scenario: 1000 proofs/day × 3-5 min reasoning per proof at ChatGPT Pro rates (~$0.01-0.05 per minute) = $30-150/day for basic proofs, higher for complex problems.
Cost comparison:
OpenAI GPT-5.2 Math ($0 for ChatGPT Pro users; high API costs for scale): Generate proofs at human-readable quality. Requires downstream verification. Best for low-volume, expert-reviewed use cases.
Wolfram Language ($315-350/year personal; $2500+/year commercial): Compute-first, symbolic math. Exact solutions to a narrower class of problems. No proof generation. Best for well-defined computational problems.
Manual expert consultation ($150-500/hour): Best for critical, novel problems. Expensive but provides certainty.
Formal verification infrastructure (Coq, Isabelle): Free tools, steep learning curve. Requires expert users. Complements OpenAI generation.
When OpenAI Math Dominates vs. When It Doesn't
Dominates: - Research teams that already have peer-review and formalization workflows. Upstream generation saves weeks of exploration. - Educational products. Users see multiple proof strategies; correctness is automatically verified against known answers. - Internal workflows. Teams that need candidate proofs explored quickly, with expert review downstream. Fails: - High-stakes systems where rigor is mandatory. Financial products, safety systems, publishable research. Unverified proofs can hide subtle errors. 58% of manuscripts remain unformalized—production systems can't tolerate that. - Products serving non-experts. If users trust the proof as-is, OpenAI Math is risky. If users understand proofs are candidates requiring verification, you're fine. - Adversarial environments. The model can be manipulated. Motivated adversaries can strip watermarks or spoof solutions. For systems defending against attackers, this is disqualifying. - Latency-critical applications. Extended reasoning takes 3-5 minutes per proof. Sub-second latency is impossible. - Massive concurrent usage. Compute costs explode. Thousands of concurrent users without proportional infrastructure is unaffordable.
Practical Recommendation: Choose Your Path
Choose OpenAI Math for exploration if: - Your product includes verification downstream (peer review, automated answer-checking, expert review, formal proofs). - 40-50% success rates are acceptable. You treat OpenAI as one input source among many. - Your users understand proofs are candidates, not finished products. Add formal verification if: - Publication or regulatory compliance is involved. Budget for Lean formalization overhead ($50-200 per proof). - Your use case requires mathematical rigor (financial models, safety-critical systems). - You're publishing research that others will build on. Skip OpenAI Math if: - Your product is entirely human-made content or uses only non-adopting generators. - Users must trust outputs as-is with no verification step. - You need sub-second latency or must serve thousands of concurrent users.
Bottom Line
OpenAI's 722 manuscripts prove the model can reason through expert-level math—a genuine breakthrough. But verification is the bottleneck. Only 42% formally verified. Translation errors exist between natural language and formalization. Compute costs are high. Unverified proofs can hide subtle errors.
For founders, the decision is simple: Build with OpenAI Math if verification is downstream from generation. Pair it with Lean formalization if publication or regulatory compliance matters. Skip it if users must trust outputs as-is.
Start by testing GPT-5.2 on your core problem. Map where you'll need formal verification. Build verification into your workflow from day one—don't assume unverified proofs are production-ready. As advisory groups note, standards for AI-assisted math work are still emerging; align with field expectations early.
Next step: Head to Bitroot to explore frameworks, guides, and decision tools for founders building with AI.
