FrontierMath

**

FrontierMath
Overview
Full name	FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI
Abbreviation	FrontierMath
Description	A benchmark of research-level mathematics problems designed to evaluate advanced mathematical reasoning in AI systems
Release date	2024-11-08
Latest version	v2 (2026-06-12, 338 corrected problems); Open Problems pilot (2026-01-27)
Benchmark updated	2026-06-12 (v2: errors corrected in 42% of original problems)
Authors	Elliot Glazer, Ege Erdil, Tamay Besiroglu, Diego Chicharro, Evan Chen, Alex Gunning, Caroline Falkman Olsson, Jean-Stanislas Denain, Anson Ho, Emily de Oliveira Santos, Olli Jarviniemi, Matthew Barnett, Robert Sandler, Jaime Sevilla, Qiuyu Ren, Elizabeth Pratt, Lionel Levine, Grant Barkley, Natalie Stewart, Bogdan Grechuk, Tetiana Grechuk, Shreepranav Varma Enugandla
Organization	Epoch AI
Technical Details
Type	Mathematical Reasoning, Research Mathematics
Modality	Text, Code
Task format	Open-ended problem solving with code execution
Number of tasks	v1: 350 (Tiers 1-3: 300; Tier 4: 50); v2: 338 (Tiers 1-3: 295; Tier 4: 43) + 14 Open Problems
Total examples	338 problems (v2) plus the Open Problems pilot
Evaluation metric	Accuracy, Automated verification
Domains	Number theory, Combinatorics, Group theory, Algebraic geometry, Real analysis, Category theory, Probability theory, Algebraic topology, and 20+ additional fields
Languages	English
Performance
Human performance	~90% (expert mathematicians with days of effort)
Baseline	<2% (most models at launch, November 2024)
SOTA score (v1)	52.4% (Tiers 1-3), 39.6% (Tier 4)
SOTA model (v1)	GPT-5.5 Pro
SOTA date	2026-04-23
Saturated	No
Resources
Website	Official website
Paper	Paper
Dataset	Download
License	Proprietary (partial public release)

FrontierMath is an advanced mathematical reasoning benchmark created by Epoch AI in collaboration with over 60 expert mathematicians, including Fields Medalists Terence Tao, Timothy Gowers, and Richard Borcherds. First published on November 8, 2024, FrontierMath consists of hundreds of original, research-level mathematics problems designed to test the outer limits of artificial intelligence systems' mathematical capabilities. At launch, every frontier AI model scored below 2% on the benchmark. By April 2026, the best-performing model, OpenAI's GPT-5.5 Pro, solved 52.4% of Tier 1-3 problems and 39.6% of Tier 4 problems, marking more than a 25-fold improvement in under two years^[1][15].

The project also includes FrontierMath: Open Problems, a pilot collection of 14 genuinely unsolved mathematical problems whose solutions, if found, would advance the state of human mathematical knowledge^[2]. On June 12, 2026, Epoch AI released FrontierMath v2 after an AI-assisted audit found small but critical errors in 42% of the original problems; the v2 release corrected 135 problems and removed 12, leaving a 338-problem set (295 in Tiers 1-3 and 43 in Tier 4)^[16][18][19].

What problem does FrontierMath solve?

By 2024, the most widely used mathematical benchmarks for AI had become saturated. Models routinely scored above 95% on GSM8K (grade-school math), above 90% on the MATH dataset (competition-level problems), and 70-90% on AIME-style questions^[3]. These high scores made it difficult to distinguish between models or to measure genuine progress in mathematical reasoning.

Epoch AI, a nonprofit research organization focused on tracking AI progress, set out to build a benchmark that would remain challenging for years. The core idea was straightforward: recruit active research mathematicians to write problems drawn from their own fields, problems that require hours or days of expert effort and whose answers can be checked automatically by a computer program.

Elliot Glazer, the project's lead mathematician, holds a Ph.D. in mathematics from Harvard, where he studied set theory under Hugh Woodin. He was joined by Tamay Besiroglu, Epoch AI's associate director, and Ege Erdil as the three core contributors. The broader team eventually grew to include over 60 mathematicians from institutions such as MIT, Harvard, Princeton, Stanford, Cambridge, Oxford, the University of Leicester, King's College London, Cornell, UC Berkeley, and Bristol University, among others. Fourteen IMO gold medalists and three Fields Medal recipients participated in problem creation or review^[4].

How is FrontierMath structured?

FrontierMath has expanded since its initial release into three distinct components, each targeting a different level of mathematical difficulty.

Tiers 1-3 (base set)

The original base set contains 300 problems spanning difficulty from advanced undergraduate to early postdoctoral level. This set forms the core benchmark used in most published evaluations. Problems are classified using the Mathematics Subject Classification (MSC2020) system and cover virtually every major branch of modern mathematics^[4]. After the June 2026 v2 correction, the base set comprises 295 problems (123 corrected and 5 removed)^[18][19].

Tier 4 (expansion set)

Released on July 1, 2025, Tier 4 adds 50 exceptionally difficult research-level problems to the benchmark. These problems were largely designed or refined during a symposium attended by leading mathematicians, where each problem was tested and approved by a panel of experts. Of the 50 Tier 4 problems, 2 are public and 48 are private. Even the strongest AI systems as of mid-2025, including OpenAI's o4-mini, Anthropic's Claude Opus 4, and Google's Gemini 2.5 Pro, achieved only single-digit success rates on Tier 4^[5]. By April 2026, GPT-5.5 Pro had pushed the Tier 4 frontier to 39.6%^[15]. The v2 update reduced Tier 4 to 43 problems (12 corrected and 7 removed)^[18][19].

Open Problems

On January 27, 2026, Epoch AI launched FrontierMath: Open Problems, a pilot benchmark of 14 genuinely unsolved mathematical research problems. Unlike the main benchmark, where problems have known solutions that an expert created, these are problems that professional mathematicians have attempted and failed to solve. The pilot set tilts toward combinatorics and number theory, where the most problems amenable to automatic verification were found^[2].

Each open problem includes a difficulty estimate from its contributor. Estimated solving times range from one to four weeks at the low end to three to ten years at the high end. The number of serious human attempts per problem ranges from two or three mathematicians to over fifty. Significance ratings span from "moderately interesting results" to "major breakthroughs"^[2].

Two problems added to the benchmark in February 2026 illustrate the scope: finding a Hadamard matrix of order 668 (the smallest order for which none is known) and proving that certain "small" Diophantine equations have infinitely many solutions^[6].

How are FrontierMath problems designed?

Core requirements

Every FrontierMath problem must satisfy four requirements before it enters the benchmark^[4]:

Scroll sideways for more →

Difficulty rating system

Each problem is rated along three dimensions by its creator and at least one peer reviewer^[4]:

Scroll sideways for more →

The authors note that these ratings provide rough guidance rather than definitive claims, since problems can become easier once a specific technique is known, and multiple solution paths of varying difficulty may exist^[4].

Mathematical domain coverage

The benchmark spans most major branches of modern mathematics. The distribution of problems by MSC2020 primary classification is as follows^[4]:

Scroll sideways for more →

Notably, 13% of problems combine number theory and combinatorics, 9% combine combinatorics and group theory, and 8% combine number theory and group theory. Over 200 distinct solution techniques are represented across the benchmark, and even the most common techniques (generating functions, recurrences, special functions) each appear in fewer than 5% of problems^[4].

How are problems created and vetted?

Creation pipeline

The process for creating and reviewing FrontierMath problems involves multiple stages^[4]:

Scroll sideways for more →

Anti-contamination measures

Because the value of the benchmark depends on problems being unknown to AI training pipelines, Epoch AI employs several security measures^[4]:

• All problems are original and previously unpublished
• Communication with contributors uses encrypted channels
• Problem files are shared via password-protected archives
• A core mathematician team manually checks problems against mathematics websites, repositories, and academic publications
• Plagiarism detection tools (Quetext and Copyscape) scan the full dataset
• The majority of the benchmark remains private, with only a handful of sample problems released publicly

Guessproof verification

Each problem undergoes a guessproofness check to confirm that the answer space is large enough (typically exceeding 10^6 possibilities) and that no obvious patterns would allow a model to stumble on the correct answer. Problems typically require large, non-obvious numerical answers or complex mathematical objects as solutions. The target is a less than 1% success rate for random or heuristic guessing^[4].

How are models evaluated?

Interactive environment

Models are evaluated in an interactive Python environment. The evaluation framework gives each model access to the following capabilities^[4]:

Scroll sideways for more →

For Tier 4 evaluations, models receive a 1,000,000-token hard limit with a 660,000-token warning threshold. The model submits a Python function that returns its answer after reasoning and code execution^[5].

Answer verification

When a model submits its answer, verification proceeds automatically^[4]:

Scroll sideways for more →

The model's code must include a specific marker comment (# This is the final answer), save the result using Python's pickle module, and be fully self-contained^[4].

How have AI models performed on FrontierMath?

Timeline of AI performance on FrontierMath (Tiers 1-3)

The following table shows how model performance has evolved since the benchmark's release. These figures are scored against the v1 dataset; v2 scores released after June 12, 2026 are not directly comparable^{[1][7][8][9][10][15]}:

Scroll sideways for more →

The top three Tier 1-3 models (GPT-5.5 Pro, GPT-5.5, GPT-5.4 Pro) cluster within 2.4 percentage points, prompting commentators to describe the benchmark as approaching saturation among frontier models even as more than 45% of v1 problems remain unsolved^[9].

Tier 4 performance

Tier 4 scores are reported separately due to the significantly higher difficulty^[5][15]:

Scroll sideways for more →

Initial model behavior patterns

In the original November 2024 evaluation, the paper's authors documented several behavioral patterns across the six tested models^[4]:

• o1-preview averaged 1.29 responses per problem, while Grok 2 Beta averaged 3.81 responses per problem
• o1-preview and Gemini 1.5 Pro tended to submit answers before seeing experimental results, even when the evaluation framework encouraged iterative testing
• Claude 3.5 Sonnet, GPT-4o, and Grok 2 Beta exceeded the 10,000-token limit in over 45% of attempts
• Gemini 1.5 Pro hit the token limit in only 16.8% of attempts, using roughly 6,000 tokens on average compared to 12,000-17,000 for other models
• Across five runs per model per problem, only four problems total were solved by at least one model; o1-preview was the only model to solve any problem on all five runs

What do expert mathematicians say about FrontierMath?

Four prominent mathematicians were interviewed for the FrontierMath paper: Terence Tao (2006 Fields Medalist), Timothy Gowers (1998 Fields Medalist), Richard Borcherds (1998 Fields Medalist), and Evan Chen (IMO coach and benchmark co-author). Their comments offer a window into how professional mathematicians view the benchmark's difficulty and significance^[4].

Terence Tao

Tao contributed several problems to the benchmark and reviewed others. He described the problems as "extremely challenging" and predicted the benchmark would "resist AIs for several years at least." On the scarcity of relevant training data, Tao observed that for many FrontierMath problems, the relevant material is "almost nonexistent... you're talking like a dozen papers with relevant things"^[4].

Tao suggested that human experts working alongside AI systems could tackle FrontierMath problems within about three years, noting that guiding current AI to correct solutions takes "about five times as much effort" as solving the problems directly. He expected this ratio to improve and eventually drop below 1 for certain problems within a few years, given sufficient tooling and capability improvements^[4].

On practical considerations, Tao remarked that if AI tools require "three days of compute off of all of Google to solve each problem... that's less of a useful tool"^[4].

Timothy Gowers

Gowers reported that "all of the problems I looked at were not really in my area and all looked like things I had no idea how to solve." He emphasized that the problems "appear to be at a different level of difficulty from IMO problems," requiring familiarity with "the tricks of the trade of some particular branch of maths," a kind of domain knowledge that is hard to acquire without substantial, specialized training data^[4].

Gowers also offered a practical vision for AI in mathematics, suggesting that AI systems could help with "slightly boring bits of doing research where you, for example, make some conjecture that would be useful, but you're not quite sure if it's true... it could be a very, very nice time saving device"^[4].

Richard Borcherds

Borcherds was described in the paper as "the most bullish" among the interviewees about AI's potential in mathematics. He did note, however, that the benchmark problems "aren't quite the same as coming up with original proofs," drawing a distinction between solving a problem with a known answer and generating new mathematical knowledge^[4].

Evan Chen

Evan Chen, a well-known mathematics educator and IMO coach who also co-authored the FrontierMath paper, published a separate blog post analyzing the benchmark's design philosophy. He noted that FrontierMath inverts two of the three desirable properties of traditional competition problems (like those at the IMO or Putnam exam). While FrontierMath retains the requirement for creative insight, it deliberately abandons the simplicity requirement and assumes the solver has "access to a Python console and a lot of reference text." Chen praised the authors for being "pretty ruthless about rejecting problems for which they felt it was possible to guess the answer" through engineer's induction^[11].

Chen identified a key advantage of FrontierMath's design: its ability to use "easily verifiable solutions" through code implementation, similar to the International Olympiad in Informatics or Project Euler. This contrasts with pencil-and-paper competitions where human coordinators must evaluate proofs^[11].

What was the o3 score controversy?

OpenAI's initial claim

On December 20, 2024, OpenAI announced its o3 reasoning model and reported a 25.2% score on FrontierMath, a dramatic leap from the previous best of under 2%. This result was highlighted during the o3 launch event as evidence of a breakthrough in mathematical reasoning^[7].

Epoch AI's independent evaluation

On April 18, 2025, Epoch AI published its own independent evaluation of the publicly released o3 model, reporting a score of approximately 10%, significantly below OpenAI's claim. Epoch AI identified several factors that could explain the discrepancy^[8]:

Scroll sideways for more →

Epoch AI noted: "The difference between our results and OpenAI's might be due to OpenAI evaluating with a more powerful internal scaffold, using more test-time computing, or because those results were run on a different subset of FrontierMath"^[8].

Funding disclosure controversy

The o3 announcement also triggered scrutiny of the financial relationship between OpenAI and Epoch AI. On the same day o3 was announced (December 20, 2024), Epoch AI disclosed that OpenAI had funded the creation of FrontierMath. Several problems quickly emerged^[12][13]:

• OpenAI had visibility into many of the problems and solutions in the benchmark before the public announcement
• The more than 60 contributing mathematicians were not informed of OpenAI's involvement or exclusive early access
• Six mathematicians who contributed significantly to the benchmark confirmed to a Stanford PhD student that they were unaware OpenAI would have exclusive access
• Epoch AI's associate director acknowledged being "restricted from disclosing the partnership until around the time o3 launched" and stated that "in hindsight we should have negotiated harder for the ability to be transparent to the benchmark contributors as soon as possible"
• OpenAI and Epoch AI had a "verbal agreement" that OpenAI would not use FrontierMath's problem set to train its AI models

The controversy drew criticism from multiple outlets. Fortune described it as "manipulative and disgraceful." TechCrunch reported that the benchmarking organization was "criticized for waiting to disclose funding from OpenAI." The incident raised broader questions about independence in AI benchmarking and the risks of conflicts of interest when AI companies fund the benchmarks used to evaluate their own models^[12][13].

Epoch AI is primarily funded by Open Philanthropy, and the OpenAI funding for FrontierMath was a separate, project-specific arrangement^[12].

Why did Epoch AI correct 42% of FrontierMath problems?

On May 11, 2026, Epoch AI announced that it was "conducting an AI-assisted review of FrontierMath: Tiers 1-4" and that the review had "flagged fatal errors in about a third of problems," adding "we believe most are valid flags." The organization said it would release updated scores on a corrected dataset once a thorough human review was complete^[16][17].

The disclosure significantly raised the estimated error rate of the benchmark. The paper's original second-round review of 25 problems had flagged roughly 1 in 25 problems (about 4%) as incorrect; the new AI-assisted pass flagged closer to one-third, an order-of-magnitude increase^[4][16]. OpenAI researcher Noam Brown publicly credited GPT-5.5 with producing the first flags, an inversion of the usual relationship in which the benchmark evaluates the model rather than the model evaluating the benchmark^[16][17].

FrontierMath v2 (June 12, 2026)

Epoch AI completed the human review and released FrontierMath v2 on June 12, 2026. The final audit found small but critical errors in 42% of the original problems, a far higher rate than the roughly 5% (about 1 in 20) suggested by the earlier human quality reviews^[18][19]. According to Epoch AI's account, the project began in April 2026 when OpenAI shared that it had found more errors than expected during an internal review; Epoch AI then ran an independent audit, using frontier models such as GPT-5.5 and Claude Opus 4.7 to flag candidate errors before engaging mathematicians to adjudicate the flags. Almost all of the flagged issues were determined to be real and severe enough to render the affected problems unsolvable as stated^[18][19].

The v2 release corrected 135 problems and removed 12, leaving 338 problems in total^[18][19]:

Scroll sideways for more →

Epoch AI framed the correction as routine benchmark maintenance rather than a benchmark failure, characterizing the original error rate as "comparable to error rates in other major ML benchmarks like ImageNet"^[18]. The organization cautioned that scores measured on v1 are not directly comparable to scores measured on the corrected v2 set, so leaderboard positions established before June 12, 2026 should be treated with care until models are re-run on v2^[18][19].

The review carried several implications:

• Previously reported scores on Tiers 1-3 and Tier 4 may shift once errored problems are removed or corrected, and Epoch AI cautioned against treating pre-v2 leaderboard positions as final^[16][18]
• A frontier model contributing materially to the audit of its own evaluation set complicates the long-standing principle that benchmarks should be created and graded independently from the systems they measure
• Commentators noted that GPT-5.5's audit performance was itself a capability demonstration: identifying valid mathematical errors at scale in research-level problems requires the same skills the benchmark is designed to test^[17]

Epoch AI maintained that final corrections were made by human mathematicians, not by AI alone, and emphasized that the models were used as filters to surface candidate errors rather than as the arbiters of validity^[16][18].

FrontierMath: Open Problems and the first AI solution

The Ramsey hypergraph breakthrough

On March 24, 2026, Epoch AI confirmed that GPT-5.4 Pro had produced a verified solution to a genuinely open mathematical problem on FrontierMath: a Ramsey-style problem on hypergraphs that had remained unsolved since it was posed by mathematicians Will Brian and Paul Larson in a 2019 paper. The solution was first elicited by researchers Kevin Barreto and Liam Price using GPT-5.4 Pro. Problem contributor Will Brian confirmed the solution's correctness, and a write-up is being prepared for publication^[14].

This marked the first time an AI model produced a novel solution to an open problem on the FrontierMath benchmark. After the initial solve, several other frontier models also solved the same problem: Claude Opus 4.6 (max), Gemini 3.1 Pro, and GPT-5.4 (xhigh). The fact that multiple models could solve it suggests the problem sat at the boundary of current frontier model capabilities^[14].

Broader context

The Ramsey hypergraph result is part of a wider trend. Since Christmas 2025, 15 open mathematical problems have moved from unsolved to solved, with 11 of them (73%) credited to AI involvement. However, Epoch AI also noted that when GPT-5.4 Pro was evaluated on the full set of FrontierMath Open Problems, it "did not solve any problems" other than the Ramsey one, and its novel observations on one other problem were "of a form that the author had anticipated and characterized as relatively uninteresting"^[14].

Sample problems

While most problems remain private to prevent contamination, the original paper includes five public sample problems at varying difficulty levels^[4]:

Scroll sideways for more →

These samples illustrate several features of the benchmark: answers are large, non-obvious integers (making them guessproof); problems span multiple mathematical fields; and even the "easiest" problem requires two hours of creative work from an expert.

How does FrontierMath compare with other benchmarks?

Difficulty scaling

Scroll sideways for more →

What sets FrontierMath apart

Scroll sideways for more →

Notable contributors

Fields Medalists

Scroll sideways for more →

Key team members

Scroll sideways for more →

Institutional participation

Over 60 mathematicians from leading institutions contributed, including researchers from MIT, Harvard, Princeton, Stanford, Cambridge, Oxford, Cornell, UC Berkeley, King's College London, the University of Leicester, the University of Siegen, ICMC USP (Brazil), and Bristol University, among others.

Implementation details

Evaluation setup

Models interact with a Python environment where they can write and execute code, test hypotheses, and submit answers. A simplified conceptual overview of the evaluation framework:

# Conceptual evaluation framework (simplified)
class FrontierMathEvaluator:
    def evaluate_model(self, model, problem):
        environment = PythonEnvironment()
        max_attempts = 10
        for attempt in range(max_attempts):
            code = model.generate_code(problem, environment.state)
            result = environment.execute(code)
            if model.verify_answer(result, problem):
                return self.check_solution(result, problem.answer)
        return False

Access tiers

Scroll sideways for more →

Funding and development

Funding sources

FrontierMath's development has been supported by:

• Open Philanthropy (primary funder of Epoch AI as an organization)
• OpenAI (project-specific funding, disclosed December 2024)^[12]
• Additional academic and industry partners

Ongoing development

Scroll sideways for more →

Impact and significance

Influence on AI research

FrontierMath has had a measurable effect on the AI research community since its release:

• It demonstrated that benchmark saturation on easier datasets (GSM8K, MATH) did not indicate genuine mathematical reasoning capability
• The benchmark's design principles, particularly its emphasis on originality, automated verification, and guessproofness, have influenced the design of subsequent benchmarks
• The o3 score controversy prompted broader discussion about transparency in AI benchmarking and the risks of vendor-funded evaluation
• The Open Problems component established a new category of AI evaluation: testing whether models can advance the frontier of human knowledge, not merely match it
• The May-June 2026 AI-assisted review and v2 correction illustrated a new dynamic: frontier models becoming capable enough to audit the benchmarks designed to evaluate them^[16][17][18]

Progress tracking

The trajectory from under 2% (November 2024) to 52.4% (April 2026) on Tiers 1-3 is one of the fastest rates of improvement on any major AI benchmark. Yet the benchmark remains far from saturated. The leading Tier 4 score is below 40%, most models still score in single digits on Tier 4, and virtually all Open Problems remain unsolved. The June 2026 v2 correction, which revised or removed 42% of the original problems, also means pre-v2 figures should be read as approximate until models are re-scored on the corrected 338-problem set^{[1][5][14][15][16][18]}.

Limitations and criticisms

Known limitations

Scroll sideways for more →

Criticisms

Several criticisms have been raised since the benchmark's launch:

1. The funding transparency failure undermined trust in the benchmark's independence, even though the problems themselves were created by independent mathematicians^[12][13]
2. The discrepancy between OpenAI's reported 25.2% and Epoch AI's measured 10% for o3 highlighted the difficulty of comparing results when testing conditions differ^[8]
3. Limited access to the full problem set makes independent replication difficult, though this restriction exists to prevent data contamination
4. Some mathematicians have questioned whether problems with automatically verifiable answers represent the full range of mathematical reasoning, since much of research mathematics involves constructing proofs rather than computing specific values^[4]
5. The June 2026 disclosure that 42% of the original Tier 1-4 problems contained errors raised fresh questions about how confidently pre-v2 year-over-year scores can be compared^[16][17][18]

See also

• MATH benchmark
• GSM8K
• AIME
• Epoch AI
• OpenAI o3
• Automated theorem proving
• Terence Tao
• AI benchmark
• Humanity's Last Exam

References

1. Epoch AI. "FrontierMath: A benchmark for evaluating advanced mathematical reasoning in AI." https://epoch.ai/frontiermath ↩
2. Epoch AI. "Introducing FrontierMath: Open Problems." Epoch AI Substack, January 2026. https://epochai.substack.com/p/introducing-frontiermath-open-problems ↩
3. VentureBeat. "AI's math problem: FrontierMath benchmark shows how far technology still has to go." November 2024. https://venturebeat.com/ai/ais-math-problem-frontiermath-benchmark-shows-how-far-technology-still-has-to-go ↩
4. Glazer, E., Erdil, E., Besiroglu, T., et al. "FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI." arXiv:2411.04872, November 2024. https://arxiv.org/abs/2411.04872 ↩
5. Epoch AI. "FrontierMath Tier 4." https://epoch.ai/benchmarks/frontiermath-tier-4 ↩
6. Epoch AI. "FrontierMath: Open Problems - Hadamard Matrices." https://epoch.ai/frontiermath/open-problems/hadamard ↩
7. OpenAI. "Announcing o3." December 20, 2024. ↩
8. TechCrunch. "OpenAI's o3 AI model scores lower on a benchmark than the company initially implied." April 20, 2025. https://techcrunch.com/2025/04/20/openais-o3-ai-model-scores-lower-on-a-benchmark-than-the-company-initially-implied/ ↩
9. llm-stats.com. "FrontierMath Benchmark Leaderboard." https://llm-stats.com/benchmarks/frontiermath ↩
10. OpenAI. "Advancing science and math with GPT-5.2." https://openai.com/index/gpt-5-2-for-science-and-math/ ↩
11. Chen, E. "FrontierMath." Power Overwhelming (blog), November 10, 2024. https://blog.evanchen.cc/2024/11/10/frontiermath/ ↩
12. TechCrunch. "AI benchmarking organization criticized for waiting to disclose funding from OpenAI." January 19, 2025. https://techcrunch.com/2025/01/19/ai-benchmarking-organization-criticized-for-waiting-to-disclose-funding-from-openai/ ↩
13. Fortune. "'Manipulative and disgraceful': OpenAI's critics seize on math benchmarking scandal." January 2025. https://fortune.com/2025/01/21/eye-on-ai-openai-o3-math-benchmark-frontiermath-epoch-altman-trump-biden/ ↩
14. WinBuzzer. "GPT-5.4 Pro Cracks Open Math Problem." March 24, 2026. https://winbuzzer.com/2026/03/24/gpt-54-pro-solves-open-math-problem-epoch-ai-frontiermath-xcxwbn/ ↩
15. Vellum. "Everything You Need to Know About GPT-5.5." April 2026. https://www.vellum.ai/blog/everything-you-need-to-know-about-gpt-5-5 ↩
16. Epoch AI. "FrontierMath: Tiers 1-4." May 11, 2026 update. https://epoch.ai/frontiermath/tiers-1-4 ↩
17. Startup Fortune. "GPT-5.5 is turning AI benchmarks into an audit problem." May 2026. https://startupfortune.com/gpt-55-is-turning-ai-benchmarks-into-an-audit-problem/ ↩
18. Epoch AI. "FrontierMath Tiers 1-4 (v2)." June 12, 2026. https://epoch.ai/frontiermath/tiers-1-4/about ↩
19. DigitalApplied. "FrontierMath v2: When AI Benchmarks Get Error-Corrected." June 2026. https://www.digitalapplied.com/blog/epoch-frontiermath-v2-error-corrected-ai-benchmark-analysis ↩