Anthropic said this week that an unreleased research version of its Claude model pushed forward one of mathematics’ oldest open problems, raising a long-standing lower bound tied to the Riemann hypothesis from 41.6% to 67.2%. The company published the result on its research site on October 5, 2026, and it was picked up by aggregator Ground News two days later, turning what started as an internal experiment into the biggest AI-and-math story of the week.
The announcement lands less than two weeks after Anthropic said it had produced the first complete, computer-checked formalization of Fermat’s Last Theorem, and just days after OpenAI published 722 papers covering 377 open mathematical problems from its own unnamed internal model. Together, the two announcements have turned math into the new proving ground for frontier AI labs, replacing coding leaderboards as the place where OpenAI and Anthropic are racing to show whose models can do something a human mathematician would recognize as real work.
What makes the Anthropic result different is the framing. The company did not claim to have solved the Riemann hypothesis, and it was explicit about that limit. It claimed something narrower and, by its own account, more verifiable: a specific numeric improvement to a bound that mathematicians have chipped away at for decades, checked against the formal proof language Lean and reviewed by in-house mathematicians before publication.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Anthropic’s Unreleased Model Actually Did
According to Anthropic’s own account, a staff researcher gave an unreleased research build of Claude what the company called “an unreasonable challenge”: take a real stab at the Riemann hypothesis, a conjecture dating to 1859 that carries a $1 million bounty from the Clay Mathematics Institute. Anthropic wrote that Claude “did take a real stab, but as you might have expected if you’re familiar with the difficulty of the task… it didn’t succeed,” a candid admission that is rare in an industry more accustomed to triumphant launch posts.
What the model did produce, Anthropic said, was progress on a related and still-significant problem: the fraction of nontrivial zeros of the Riemann zeta function that can be proven to lie on the critical line, the condition at the heart of the hypothesis itself. “An unreleased research version of Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis,” the company said in its research post, adding that the system was “drawing on extensive prior research by mathematicians over the past decades” rather than starting from scratch.
Reports describing the work, including the Ground News aggregation published October 7, say the model tested roughly 650 distinct mathematical approaches and coordinated the search across 60 subagents before the result was formalized in Lean, the proof-assistant language increasingly used across the industry to machine-check advanced mathematics. Anthropic has said two of its in-house mathematicians reviewed the result internally, and outside experts were brought in to check the work on short notice before publication.
The Riemann Hypothesis, Explained for Non-Mathematicians
The Riemann hypothesis is arguably the most famous unsolved problem in pure mathematics. Proposed by Bernhard Riemann in 1859, it makes a precise claim about where the zeros of a particular function, the Riemann zeta function, are allowed to sit on the complex plane. If true, it would lock in extremely tight guarantees about how prime numbers are distributed, with knock-on effects across number theory, cryptography and physics. It is one of the seven Clay Millennium Prize Problems announced in 2000, each carrying a $1 million reward, and to date only one of the seven, the Poincaré conjecture, has been resolved.
Because a full proof has eluded mathematicians for 167 years, researchers have spent decades instead trying to prove weaker, partial versions of the claim. One of the most useful partial results is a lower bound: a guarantee that at least some percentage of the zeros definitely satisfy the hypothesis, even if not every single one can yet be confirmed. Raising that percentage does not prove the hypothesis, but it narrows the space in which a counterexample could hide, and it is the kind of incremental, technical progress that professional number theorists treat as genuine contribution rather than a stunt.
From 41.6% to 67.2%: What the Jump Actually Means
The specific number at the center of this week’s news is the move from a 41.6% lower bound to 67.2%. In plain terms, before this result, mathematicians could only formally guarantee that 41.6% of the Riemann zeta function’s zeros sit where the hypothesis says they should. Anthropic’s unreleased model pushed that guaranteed floor to 67.2%, a jump of more than 25 percentage points on a figure that typically moves in much smaller increments when human researchers publish incremental advances.
Anthropic has been careful about how it frames the achievement. The company’s own language stresses that the system built on “extensive prior research by mathematicians over the past decades,” positioning Claude’s contribution as an extension of existing techniques rather than an independent breakthrough invented from nothing. That framing matters, because it is the opposite of the narrative that follows OpenAI’s math announcements, where the emphasis has tended to land on the sheer volume of output rather than on how the work sits inside the existing literature.
It is also worth noting what the result is not. It is not a proof of the Riemann hypothesis, it does not collect the Clay Institute’s bounty, and Anthropic has not claimed otherwise. The company’s own post is unusually blunt that Claude “didn’t succeed” at the actual challenge it was given. The 67.2% figure is the consolation prize, and by mathematical standards it is a real one.
Why the Lean Formalization Step Matters
Both this result and Anthropic’s Fermat’s Last Theorem work were formalized in Lean, an open-source proof assistant that forces every logical step of an argument to be checked by a computer rather than trusted on a mathematician’s say-so. That distinction has become the dividing line in how seriously the mathematical community treats AI-generated claims this year. A natural-language proof sketch can contain subtle gaps that are easy to miss on a first read, something mathematicians scrutinizing OpenAI’s 722-paper release have pointed to directly. A Lean-checked proof either compiles against the formal rules of logic or it does not.
That is also why the subagent architecture behind the Riemann result is getting attention on its own. Coordinating 60 subagents across 650 candidate approaches is a very different workflow from a single model generating one long chain of reasoning. It looks more like a distributed search process, where many smaller explorations get pruned and the surviving threads get stitched into a verified result. Anthropic has used a similar multi-agent structure in its other recent mathematics work, and the approach appears to be becoming the company’s default method for tackling problems too large for a single context window.
Not Anthropic’s First Math Swing This Fall
The Riemann announcement is the third major mathematics result Anthropic has published in roughly six weeks, and the pattern is becoming a recognizable cadence rather than a one-off. On September 23, 2026, the company said it had produced the first complete, computer-checked formalization of Fermat’s Last Theorem, the 17th-century conjecture that mathematician Andrew Wiles famously proved in 1994 after a seven-year effort. Anthropic’s version, built by a multi-agent system, reportedly generated approximately 13 million lines of Lean code and established 29,500 intermediate theorems along the way, a project the company said was completed in under two weeks.
That formalization effort connects directly to a market that had already been watching Claude’s math claims closely. A prediction market tracking Fermat’s Last Theorem formalization odds spiked to 99% once the Lean proof surfaced, showing how quickly speculative traders moved to price in Anthropic’s claim once it became public. A separate market tied to the Navier-Stokes existence and smoothness problem, another of the seven Millennium Prize Problems, has also been trading at elevated odds on rumors of a Claude-linked attempt, even though no formal Anthropic claim on Navier-Stokes has been published.
Watson’s 87-Year-Old Problem and “Claude-Shaped Science”
Sandwiched between the Fermat formalization and this week’s Riemann news was an October 1, 2026 post Anthropic titled Claude-shaped science, describing Claude-assisted work spanning 18 scientific fields with 19 human co-authors. The most eye-catching single result in that post was the resolution of what researchers called Watson’s “final problem,” the exact return probability of a three-dimensional random walk with three unequal hopping rates. Mathematician George Watson first worked on this specific family of lattice integrals back in 1939, and the particular case solved by Anthropic’s team had sat unfinished for 87 years.
According to a guest post on Anthropic’s site covered by The Print, a Harvard physicist described Claude completing in weeks a research task that had taken him roughly a year on his own, a comparison he used to raise questions about what the pace of AI-assisted research means for how academic science gets done and credited going forward. That mix of genuine productivity gain and open unease about authorship is becoming a recurring theme every time one of these labs publishes a new math or science result.
Anthropic’s 2026 Math and Science Timeline
| Date | Result | Key detail | Source |
|---|---|---|---|
| Sept 23, 2026 | Fermat’s Last Theorem formalized in Lean | ~13 million lines of Lean, 29,500 intermediate theorems, under two weeks | Anthropic research |
| Oct 1, 2026 | “Claude-shaped science” roundup | 18 scientific fields, 19 human co-authors, Watson’s 1939 problem solved | Anthropic / The Print |
| Oct 5, 2026 | Riemann zeta lower bound raised | 650 ideas tested via 60 subagents, bound moved from 41.6% to 67.2% | Anthropic / Ground News |
How This Compares to OpenAI’s 722-Paper Math Dump
OpenAI took a very different path to the same news cycle. On October 6, 2026, the company released 722 papers covering solutions to 377 open mathematical questions spanning nearly every subdiscipline, all produced by an internal, unnamed frontier model and uploaded to a public GitHub repository, according to Quanta Magazine’s coverage. The scale is the headline. Anthropic published one tightly scoped numeric result with a full methodology writeup. OpenAI published hundreds of papers at once and left much of the verification work to the outside mathematical community.
That difference in approach is exactly what has drawn scrutiny toward OpenAI’s release. Mathematicians, including researchers connected to Terence Tao, have begun combing through the 722-paper batch looking for gaps, duplication, or problems that were already solved elsewhere in the literature, a process that is still ongoing. Separately, OpenAI has also reportedly claimed a proposed solution to the Navier-Stokes existence and smoothness problem, another Millennium Prize question, according to a September 10, 2026 report from Campus Technology, though that claim has not undergone the kind of public, line-by-line formal verification that Anthropic applied to its Riemann and Fermat results.
The contrast gives the current AI-math moment two distinct cultures. OpenAI is optimizing for volume and speed, betting that quantity of output will eventually translate into recognized contributions once the community sorts through it. Anthropic is optimizing for defensibility, publishing fewer results but pairing each one with Lean formalization and named internal review before the public ever sees it.
DeepMind’s Earlier, Narrower Approach
Google DeepMind got here first, just with a narrower target. Its AlphaProof and AlphaGeometry 2 systems jointly solved four of six problems at the 2024 International Mathematical Olympiad, a result formally recognized as equivalent to a silver-medal performance against human competitors, according to reporting from Kingy AI. That was a competition-math result, built on systems tuned specifically for Olympiad-style problems rather than open research questions, and it predates both the current OpenAI and Anthropic pushes by roughly two years.
The three labs are now effectively running three different experiments in parallel: DeepMind’s specialized competition solvers, OpenAI’s general frontier model applied at massive publication scale, and Anthropic’s general frontier model applied narrowly with heavy formal verification. Which approach produces results the professional mathematics community actually accepts as real contributions, rather than impressive-looking output, is the open question the rest of 2026 is likely to start answering.
The Benchmark Race: FrontierMath, Erdős Problems and the New Scoreboard
Underneath the individual announcements, a more systematic benchmarking effort is tracking how fast frontier models are actually improving at mathematics. Epoch AI’s FrontierMath v2 leaderboard, last updated October 6, 2026 according to benchmark tracker BenchLM, shows OpenAI’s GPT-5.6 Sol leading the Tiers 1-3 bracket at 89.0%, ahead of GPT-5.6 Terra at 84.9% and GPT-5.6 Luna at 78.6%. On the harder Tier 4 bracket, GPT-6 Astra tops the board at 97.6%, with GPT-5.6 Sol at 83.0% and GPT-5.6 Terra at 68.3%.
Those scores are climbing fast enough that Epoch AI launched an entirely new, harder benchmark on September 3, 2026: FrontierMath Erdős, built from 68 genuinely unsolved problems posed by mathematician Paul Erdős, with models required to write solutions directly in Lean. No model had solved any of the 68 problems before this benchmark existed. GPT-6 Astra has since solved two of the 68, a 3% clearance rate that sounds modest until you remember these are problems with no known solution at all, not just problems a model hasn’t memorized.
| Benchmark | Model | Score | Source |
|---|---|---|---|
| FrontierMath v2, Tier 4 | GPT-6 Astra | 97.6% | BenchLM / Epoch AI |
| FrontierMath v2, Tier 4 | GPT-5.6 Sol | 83.0% | BenchLM / Epoch AI |
| FrontierMath v2, Tier 4 | GPT-5.6 Terra | 68.3% | BenchLM / Epoch AI |
| FrontierMath v2, Tiers 1-3 | GPT-5.6 Sol | 89.0% | BenchLM / Epoch AI |
| FrontierMath v2, Tiers 1-3 | GPT-5.6 Terra | 84.9% | BenchLM / Epoch AI |
| FrontierMath v2, Tiers 1-3 | GPT-5.6 Luna | 78.6% | BenchLM / Epoch AI |
| FrontierMath Erdős (68 unsolved problems) | GPT-6 Astra | 2/68 (3%) | Epoch AI |
| 2024 IMO (competition math) | AlphaProof + AlphaGeometry 2 | 4/6 (silver-medal standard) | Kingy AI |
Anthropic’s models are notably absent from the public FrontierMath v2 and Erdős leaderboards, which track released, publicly available systems. That gap underscores a key distinction in this week’s news: the model behind the Riemann result is explicitly unreleased, meaning it cannot yet be independently benchmarked by Epoch AI or anyone else outside Anthropic. For a fuller picture of how released frontier models stack up against each other on reasoning and coding tasks more broadly, see this site’s comparison of the LMArena and Artificial Analysis leaderboards.
In Anthropic’s Own Words
Anthropic’s public statements about the Riemann work are unusually direct about both the achievement and its limits. Announcing the result, the company wrote: “We asked an unreleased research version of Claude to take a stab at the Riemann hypothesis. It didn’t solve it, but it did make strides on a related problem: it increased the lower bound for the fraction of zeros of the Riemann zeta function that satisfy the hypothesis from 41.6% to 67.2%,” according to a post from the company’s official account.
In the accompanying research writeup, Anthropic described the origin of the challenge in almost casual terms: “Recently, a member of staff at Anthropic gave Claude an unreasonable challenge. It was about one of the most famous unsolved problems in mathematics: Take a real stab at the Riemann hypothesis,” the company wrote on its research page. It followed that with a clear-eyed acknowledgment of the outcome: “Claude did take a real stab, but as you might have expected if you’re familiar with the difficulty of the task (the Riemann hypothesis dates back to 1859 and has a million-dollar bounty), it didn’t succeed.”
On the specific technical gain, Anthropic’s research post stated plainly: “An unreleased research version of Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis,” and credited the broader field for the foundation that made the jump possible: “Drawing on extensive prior research by mathematicians over the past decades, it has increased this bound from 41.6% to 67.2%.”
Market Stakes: Why Math Credibility Matters Right Now
The timing of these results is not happening in a vacuum. Anthropic has been building toward a public offering, and the company has simultaneously been vocal about the risks its own technology poses, a tension explored in this site’s earlier coverage of Anthropic warning of existential AI risk inside its own IPO filing. A steady drumbeat of verifiable, peer-checked mathematics results functions as a counterweight to that risk narrative: it gives prospective investors concrete, falsifiable evidence of capability, rather than asking them to take safety or capability claims purely on faith.
For OpenAI, the 722-paper release carries a different kind of market logic. The company is not currently running a comparable public offering process, but it is competing directly with Anthropic for enterprise customers, researcher talent and the broader narrative of which lab is closer to genuinely novel scientific discovery rather than incremental benchmark gains. Both companies have strong incentives to be the one associated with the next verified mathematical breakthrough, which helps explain why both chose the same two-week window in early October to make their biggest math claims of the year.
Historical Context: A 167-Year-Old Problem Meets 2026 AI
It is worth stepping back on scale. The Riemann hypothesis has survived serious attack from essentially every major number theorist of the past century and a half. Fermat’s Last Theorem, stated by Pierre de Fermat around 1637, went unproven for 358 years until Andrew Wiles closed it in 1994 using techniques far beyond anything available in Fermat’s lifetime. Watson’s lattice-integral problem sat open for 87 years before this fall’s resolution. Against that backdrop, an AI system raising a partial bound by 25 percentage points, or formalizing an already-proven theorem in machine-checkable form, is not the same thing as solving a new Millennium Prize Problem outright, and neither Anthropic nor OpenAI has claimed otherwise this month.
What has genuinely changed is speed and the division of labor. The Fermat formalization that took Anthropic’s multi-agent system under two weeks would, by any historical comparison, have taken a human team of specialists in formal verification substantially longer to encode into Lean, even working from Wiles’s already-complete proof. That compression of formalization time, rather than the invention of new mathematics from scratch, may end up being the more durable and less controversial contribution these systems are making in 2026.
What Comes Next: Five Predictions
- Expect both OpenAI and Anthropic to keep using unreleased, internal-only model versions for their highest-profile math claims through 2027, reserving public model releases for commercial products rather than research headlines.
- Lean formalization will keep becoming the minimum bar for a math claim to be taken seriously by the professional community, pushing labs that skip it, as OpenAI’s 722-paper batch has been criticized for in places, toward adopting it more consistently.
- Epoch AI and similar benchmark groups will keep raising the difficulty ceiling, following the same pattern as FrontierMath Erdős, as soon as current leaderboards start to saturate above 90%.
- More named mathematicians will publicly weigh in on AI-generated proofs, mirroring the scrutiny already directed at OpenAI’s math release, turning outside verification into a routine step rather than an afterthought.
- Expect at least one more Clay Millennium Prize Problem, most plausibly Navier-Stokes given the rumors already trading on prediction markets, to become the subject of a formal lab announcement before the end of 2026.
Frequently Asked Questions
Did Anthropic’s Claude model solve the Riemann hypothesis?
No. Anthropic has been explicit that its unreleased research model did not solve the Riemann hypothesis. It improved a related lower bound, the guaranteed percentage of zeta function zeros known to satisfy the hypothesis, from 41.6% to 67.2%.
What is the Riemann hypothesis in simple terms?
It is an 1859 conjecture by Bernhard Riemann about where the zeros of the Riemann zeta function sit on the complex plane. If proven true, it would tighten guarantees about how prime numbers are distributed. It is one of the Clay Mathematics Institute’s seven Millennium Prize Problems and carries a $1 million reward.
What does a “lower bound” of 67.2% actually mean?
It means mathematicians can now formally guarantee that at least 67.2% of the zeta function’s zeros lie where the Riemann hypothesis says they should, up from a previous guaranteed figure of 41.6%. It is not a percentage of the hypothesis being “true,” but a measure of how much of the problem space has been formally closed off from potential counterexamples.
Is this the same as OpenAI’s 722 math papers?
No, they are separate claims from separate companies. OpenAI published 722 papers covering 377 problem solutions from its own internal model on October 6, 2026. Anthropic’s Riemann-related result, published October 5, 2026, is a single, narrowly scoped numeric result formalized in Lean and reviewed internally before release.
What is Lean, and why does it matter for AI math claims?
Lean is an open-source proof assistant that checks every logical step of a mathematical argument by computer rather than relying on a human reviewer to catch gaps. Results formalized in Lean carry more weight with the mathematical community than natural-language proof sketches, because the formal checker either accepts or rejects the logic outright.
Has Anthropic made other big math claims recently?
Yes. On September 23, 2026, Anthropic said it produced the first complete, computer-checked formalization of Fermat’s Last Theorem. On October 1, 2026, it published a roundup describing Claude-assisted work across 18 scientific fields, including the resolution of a lattice-integral problem mathematician George Watson began working on in 1939.
Does this result collect the Clay Institute’s $1 million prize?
No. The Clay Mathematics Institute’s Millennium Prize is reserved for a full proof of the Riemann hypothesis itself. Anthropic’s result addresses a related partial bound and does not constitute a proof of the hypothesis, so it is not eligible for the prize.
Related Coverage
- Meta, Microsoft Cut Claude Use, Lean on In-House AI Tools [2026]
- Claude Opus 5.5 API Setup: 13 Steps, $4/$20 per 1M [2026]
- Claude Sonnet 5.5 API Setup: 13 Steps, $2/$10 per 1M [2026]
- Clippy Moment: Meta Muse Mascot Hits 5M Downloads [2026]
- SiMa.ai Raises $150M, Hits $1.45B Valuation [2026]
