Skip to content
← AI by field
Mathematicians, researchers, teachers and students

AI for math, right now

Solving competition and research problems, writing and checking proofs, and working on open problems.

Moving: very fastUpdated Nov 12, 2026 from 9 sources
Open math in Spezia Free account. No card.
Where AI is in math now

AI models now solve research-level math problems and formalize proofs, with Claude Fable 5.2 leading benchmarks.

What models can do now

  • Claude Fable 5.2 achieves 95.1% on FrontierMath, up from 55.8% a year ago.[1]
  • GPT-6.1 Sol scores 100% on FrontierMath Tier 4, research-level problems that take mathematicians days.[2]
  • Gemini 4 Carbon and AlphaProof Code formally verified libsignal, generating over 6.5 million lines of Lean in four weeks.[3]
  • An unreleased Anthropic model proved the 20-year-old Aaronson–Ambainis quantum conjecture, verified by Lean.[5]
  • AI error-checking found major flaws in about 6% of pre-August math papers, with major-error rate falling to 1.5% recently.[7]

Where they still fall short

  • AI-generated proofs are being produced faster than humans can verify and understand them.[8]
  • OpenAI did not release the 1147-page proof of the 4D Yang-Mills existence problem, instead appointing a committee to shepherd it.[9]

What is expected next

  • Altman reportedly said the AI behind OpenAI's Navier-Stokes result is the next GPT and will be released soon.[6]
  • Anthropic says Claude helped solve two open physics problems and formalize proofs with 1,000 agents.[4]

How fast it is moving

The record on this field's benchmarks over the last year, from Spezia's model leaderboard.

FrontierMath

Hundreds of new, unpublished math problems written by mathematicians, from hard undergraduate level up to research, with answers a computer can check.

Record: 95.1%, up from 55.8% a year ago (1.7×) (Claude Fable 5.2)

0%25%50%75%100%Claude Fable 5.2GPT-5.5 Pro
Oct 25Jan 26Apr 26Jul 26Oct 26

the record so far set a record other models

FrontierMath Tier 4

Research-level problems that take professional mathematicians days.

Record: 100%, up from 21.9% a year ago (4.6×) (GPT-6.1 Sol)

0%25%50%75%100%Claude Fable 5GPT-6.1 Sol
Oct 25Jan 26Apr 26Jul 26Oct 26

the record so far set a record other models

The best models for math now

On this field's benchmarks alone.

  1. 1Claude Fable 5.2
  2. 2GPT-6.1 Sol
  3. 3GPT-6 Astra
  4. 4Claude Opus 5.5
  5. 5Gemini 4 Carbon
  6. 6GPT-6.1 Terra

Milestones

What became possible, newest first, each marked by how firm it is. Open one for the words it rests on.

Showndemonstrated with published results, a paper or an independent checkClaimedsaid by the company or people who built it, not yet checked by othersReportedreported by others, not confirmed by the company or a published resultNot publicdone by a model that is not released to the public

  1. Nov 10, 2026ClaimedClaude
    Anthropic says Claude helped solve two open physics problems and formalize proofs with 1,000 agents.
    “Anthropic says Claude helped solve two open physics problems and formalize proofs with 1,000 agents.” Anthropic says Claude helped solve two open physics problems and formalize proof
  2. Nov 10, 2026ClaimedClaude Fable 5.2
    Unreleased Claude Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours.
    “a team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours with minimal human intervention” Unreleased Claude Fable 5.2 agents fully formalized Fermat's Last Theorem in 30
  3. Nov 8, 2026Claimedunreleased Anthropic model
    An unreleased Anthropic model proved the 20-year-old Aaronson–Ambainis quantum conjecture, verified by Lean.
    “with the help of an unreleased Anthropic model, it proved the Aaronson–Ambainis conjecture, which has puzzled the mathematics community for nearly 20 years” An unreleased Anthropic model proved the 20-year-old Aaronson–Ambainis quantum c
  4. Nov 6, 2026ReportedOpenAI agents
    OpenAI reportedly used 10,000 AI agents to solve a Navier-Stokes problem.
    “OpenAI reportedly used 10,000 AI agents to solve a Navier-Stokes problem” OpenAI reportedly used 10,000 AI agents to solve a Navier-Stokes problem, and Al
  5. Nov 6, 2026ClaimedOpenAI model
    Altman says an OpenAI model helped settle a Clay Institute math problem, costing about $10 million.
    “Altman says an OpenAI model helped settle a Clay Institute math problem, costing about $10 million.” Altman says an OpenAI model helped settle a Clay Institute math problem, costing
  6. Nov 5, 2026Reported
    An AI agent edited 0.016% of another model's weights, lifting MATH-500 from 7.9% to 48.1%.
    “An AI agent edited 0.016% of another model's weights, lifting MATH-500 from 7.9% to 48.1%.” An AI agent edited 0.016% of another model's weights, lifting MATH-500 from 7.9%
Every milestone, story and model Pick a model to see what it can do in math.

Sources

  1. [1]FrontierMath: best result 95.1%, up from 55.8% a year ago (1.7×)Leaderboard · 2026-11-09
  2. [2]FrontierMath Tier 4: best result 100%, up from 21.9% a year ago (4.6×)Leaderboard · 2026-09-29
  3. [3]Gemini 4 Carbon and AlphaProof Code formally verified libsignal, generating over 6.5 million lines of Lean in four weeks.Spezia news · 2026-11-11
  4. [4]Anthropic says Claude helped solve two open physics problems and formalize proofs with 1,000 agents.Spezia news · 2026-11-10
  5. [5]An unreleased Anthropic model proved the 20-year-old Aaronson–Ambainis quantum conjecture, verified by Lean.Spezia news · 2026-11-08
  6. [6]OpenAI reportedly used 10,000 AI agents to solve a Navier-Stokes problem, and Altman says the next GPT is coming soon.Spezia news · 2026-11-06
  7. [7]AI error-checking found major flaws in about 6% of pre-August math papers, with major-error rate falling to 1.5% recently.Spezia news · 2026-11-03
  8. [8]AI Solves the Toughest Math Puzzles—What Does That Mean for Us, Specifically?Web · 2026-11-10
  9. [9]AI rivals tackle one of physics’ deepest mathematical challengesWeb · 2026-11-10

Other fields