AI benchmarks

Every frontier AI model on every major benchmark

Frontier AI benchmark scores as of July 6, 2026: on ARC-AGI-2, GPT-5.5 leads at 85%; on GPQA, GPT-5.5 leads at 93.6%; on SWE Pro, Claude Fable 5 leads at 80.3%. Every score below cites the lab’s announcement post or an independent re-runner.

Last verified: July 6, 2026.

ARC-AGI-2 leader
85%
GPT-5.5
GPQA leader
93.6%
GPT-5.5
SWE Pro leader
80.3%
Claude Fable 5
Coverage
8 × 7
frontier models × benchmarks

How to read this page

● 87.2 — lab-claimed score. Sourced from the lab’s own announcement post or model card. Click the number for the citation. Closed-weights labs report what they choose to report; treat as the upper bound.

□ 86.4 — independent re-run. Sourced from Epoch AI, Artificial Analysis, Aider, or the benchmark’s own public leaderboard. Click the number for the evaluator’s page.

⚠ diverge — lab and independent scores differ by more than 5 percentage points. Often signals a methodology gap (extended thinking enabled vs. not, tools on vs. off, different subset, leaked-test contamination).

— — the lab didn’t publish this score and no independent re-run has landed yet. Honest gap, not zero.

Headline matrix

Each row is a current frontier flagship from one lab; each column is a major benchmark, ordered from most-discriminating to most-saturated. Click any score for its primary-source citation; click any column header to jump to that benchmark’s section below.

Anthropic · June 9, 2026
HLE
●59□53.3⚠ diverge
SWE Pro
●80.3
AIME
—
ARC-AGI-2
—
GPQA
●92.6□92.6
MMMU
—
SWE Verified
●95.5
DeepSeek · April 24, 2026
HLE
●23.4□35.9⚠ diverge
SWE Pro
●54.1
AIME
●91.8
ARC-AGI-2
—
GPQA
●90.1□88.8
MMMU
●76.4
SWE Verified
●80.6
Google · May 19, 2026
HLE
●40.2□41
SWE Pro
●55.1
AIME
—
ARC-AGI-2
●72.1□72.1
GPQA
□92.2
MMMU
—
SWE Verified
—
GPT-5.5Closed
OpenAI · April 23, 2026
HLE
●52.2□44.3⚠ diverge
SWE Pro
●58.6
AIME
—
ARC-AGI-2
●85□85
GPQA
●93.6□93.5
MMMU
—
SWE Verified
—
Grok 4.3Closed
xAI · April 17, 2026
HLE
□35
SWE Pro
—
AIME
—
ARC-AGI-2
—
GPQA
□90.1
MMMU
—
SWE Verified
—
Mistral AI · April 28, 2026
HLE
□12.8
SWE Pro
—
AIME
—
ARC-AGI-2
—
GPQA
□74.8
MMMU
—
SWE Verified
●77.6
Meta · April 8, 2026
HLE
●58□39.9⚠ diverge
SWE Pro
□55
AIME
●87.3
ARC-AGI-2
—
GPQA
●89.5□88.4
MMMU
●81.4
SWE Verified
●77.4
Alibaba · May 31, 2026
HLE
□33.4
SWE Pro
—
AIME
—
ARC-AGI-2
—
GPQA
—
MMMU
—
SWE Verified
—

By benchmark

Ordered by how much each benchmark currently discriminates between frontier labs. Discriminating benchmarks separate models by capability; approaching-saturation benchmarks separate by points within a tight band.

Humanity's Last Exam

KnowledgeDiscriminating

Crowdsourced 3,000-question expert-level exam across 100+ subjects, designed to be the last academic benchmark needed before frontier models match expert humans. Reported with and without tools.

Author
Center for AI Safety · Scale AI
lastexam.ai
Human baseline
Expert humans in their own domains score ~88%; broad humans far below

Saturation: Frontier scores still span a wide band — this benchmark separates the labs.

Model
Score
●59□53.3⚠ diverge
Notes
Humanity's Last Exam 59.0 (no tools), per Fable 5's launch comparison table (2026-06-09, via Vellum). Anthropic leads the no-tools HLE field at launch; a separate with-tools figure was not broken out in the launch table (Opus 4.8's with-tools 57.9 is the prior Anthropic reference point).
Model
Score
●58□39.9⚠ diverge
Notes
Per launch blog text: "Contemplating mode provides significant capability improvements in challenging tasks, achieving 58% in Humanity's Last Exam." Standard (non-Contemplating) mode launch chart shows lower; the 58% figure is the headline number.
Model
GPT-5.5
OpenAI
Score
●52.2□44.3⚠ diverge
Notes
With tools enabled (52.2); 41.4 without tools, per the GPT-5.5 announcement.
Model
Score
●40.2□41
Notes
Humanity's Last Exam, full set (text + multimodal). The model card does not split tools vs. no-tools for HLE on 3.5 Flash.
Model
Score
□35
Notes
xAI did not publish a formal Grok 4.3 announcement post or model card. Prior-cycle value (55.3) cited a now-404 URL (x.ai/news/grok-4-3) and is removed pending a primary source.
Model
Score
□33.4
Notes
Not reported in the Qwen3.7-Plus launch post; the launch focused on multimodal vision and agentic capabilities, not academic frontier reasoning.
Model
Score
●23.4□35.9⚠ diverge
Notes
Reported with Thinking mode enabled.
Model
Score
□12.8
Notes
Not reported in the Medium 3.5 release post.

SWE-Bench Pro

CodingDiscriminating

Contamination-resistant, multi-language successor to SWE-bench Verified. Real GitHub issues from production codebases that the model must patch end-to-end. The headline software-engineering benchmark on every frontier release in 2026. With Claude Fable 5's export-control suspension lifted (redeployed globally 2026-07-01), its 80.3 lab-claimed top re-enters the production roster over a 54–59 cluster — a ~26pp band — so the benchmark reads as discriminating again at the production frontier.

Author
Scale AI
labs.scale.com/leaderboard/swe_bench_pro_public
Human baseline
Human SWE pass-rate on a comparable subset estimated ~75% by the benchmark authors

Saturation: Frontier scores still span a wide band — this benchmark separates the labs.

Model
Score
●80.3
Notes
SWE-Bench Pro pass-rate (Anthropic agentic-coding scaffold), Fable 5's headline number per the launch table (2026-06-09, via Vellum): Fable 5 80.3, Mythos Preview 77.8, Opus 4.8 69.2, GPT-5.5 58.6, Gemini 3.1 Pro 54.2 — the top score of any model tested. Vendor-scaffold number; not directly comparable to Scale's standardized SEAL leaderboard.
Model
GPT-5.5
OpenAI
Score
●58.6
Notes
SWE-Bench Pro (Public) per the GPT-5.5 announcement; OpenAI flags evidence of memorization on this public eval.
Model
Score
●55.1
Notes
SWE-Bench Pro (Public), single attempt, per the Gemini 3.5 Flash model card.
Model
Score
□55
Notes
Meta did not report SWE-Bench Pro for the Muse Spark launch.
Model
Score
●54.1
Notes
Model
Score
Not reported
Notes
No published Grok 4.3 SWE-Bench Pro number.
Model
Score
Not reported
Notes
Mistral did NOT publish SWE-Bench Pro in the Medium 3.5 launch post — only SWE-Bench Verified (77.6, text-confirmed). Prior-cycle 50.8 cited the launch post URL but is not text-verifiable on the official post; removed pending primary-source confirmation.
Model
Score
Not reported
Notes
Qwen3.7-Plus launch did not publish SWE-Bench Pro. The text-only sibling Qwen3.7-Max (now retired as Qwen flagship per /ai/models/) had reported 60.6 on SWE-Bench Pro on 2026-05-20.

AIME 2025

MathApproaching saturation

American Invitational Mathematics Examination, 2025 edition. 15 integer-answer problems; widely used as the frontier math benchmark because the problems are public after release but the answer space rewards reasoning over recall.

Author
Mathematical Association of America
maa.org/math-competitions/aime
Human baseline
Strong high-school competitors solve ~50%; AIME qualifiers (top USAMO contenders) ~80%

Saturation: Frontier scores cluster near the top — this benchmark separates labs by points, not by capability.

Model
Score
●91.8
Notes
DeepSeek-V3.2-Speciale (V3.2's high-compute variant) claimed IMO/IOI gold medals at this scale; V4-Pro inherits the reasoning gains.
Model
Score
●87.3
Notes
Model
Score
Not reported
Notes
Anthropic's Fable 5 launch table does not report AIME 2025; the math-and-reasoning slot is filled by HLE (no tools) and FrontierCode instead.
Model
Score
Not reported
Notes
Gemini 3.5 Flash model card does not report AIME 2025; reports Humanity's Last Exam, ARC-AGI-2, and agentic benchmarks instead.
Model
GPT-5.5
OpenAI
Score
Not reported
Notes
OpenAI's GPT-5.5 announcement post (openai.com/index/introducing-gpt-5-5) reports FrontierMath instead of AIME 2025. The GPT-5.5 system card (deploymentsafety.openai.com/gpt-5-5) was checked on 2026-06-13 and is a safety document — it carries no AIME capability score. Prior-cycle value (96.4) was not text-verifiable on any OpenAI primary source and is permanently removed; this cell stays null unless OpenAI republishes AIME 2025 for GPT-5.5.
Model
Score
Not reported
Notes
xAI did not publish a formal Grok 4.3 announcement post or model card. Prior-cycle value (95.2 Heavy / ~88 single-instance) cited a now-404 URL (x.ai/news/grok-4-3) and is removed pending a primary source.
Model
Score
Not reported
Notes
Mistral did NOT publish AIME 2025 in the Medium 3.5 launch post. The launch post's body text names only SWE-Bench Verified and τ³-Telecom; multiple third-party reviewers explicitly note that Mistral skipped AIME / GPQA / MMLU / HumanEval / MATH for this release. Prior-cycle 88.1 cited the launch post URL but is not text-verifiable; removed pending primary-source confirmation.
Model
Score
Not reported
Notes
Qwen3.7-Plus launch did not publish AIME 2025. The text-only sibling Qwen3.7-Max (now retired as Qwen flagship) had reported HMMT 2026 Feb (97.1) rather than AIME 2025; neither was carried forward on the multimodal Plus model.

ARC-AGI-2

ReasoningApproaching saturation

Successor to the original ARC-AGI prize. Visual-pattern abstraction puzzles designed to resist memorization. Frontier flagships now cluster in the 70–85% band — still informative but tightening as the top of the field converges.

Author
François Chollet · ARC Prize Foundation
arcprize.org/arc-agi/2
Human baseline
Human-untrained ~60%, expert ~95% on the public set; the hard private set is by design lower

Saturation: Frontier scores cluster near the top — this benchmark separates labs by points, not by capability.

Model
GPT-5.5
OpenAI
Score
●85□85
Notes
ARC-AGI-2 (Verified) reported in the GPT-5.5 announcement table (Abstract reasoning section) at xhigh reasoning. OpenAI uses 'Verified' to describe their own test protocol; this is not the same as the ARC Prize Foundation's verified public leaderboard.
Model
Score
●72.1□72.1
Notes
Per the Gemini 3.5 Flash model card (Evaluation > Results, Reasoning section). The launch blog headlined Terminal-Bench 2.1 / GDPval-AA / MCP Atlas instead; ARC-AGI-2 is from the model card.
Model
Score
Not reported
Notes
Anthropic's Fable 5 launch comparison table (2026-06-09) does not include ARC-AGI-2; the published rows are SWE-Bench Pro / Verified, FrontierCode (incl. Diamond split), Terminal-Bench 2.1, HLE (no tools), GPQA Diamond, GDP.pdf vision, and tau-squared-Bench. Fable 5 is not yet a distinct row on the ARC Prize v2 public leaderboard, so no independent ARC-AGI-2 re-run is recorded.
Model
Score
Not reported
Notes
DeepSeek does not report ARC-AGI-2 on their announcement post (text-confirmed: the V4 launch post is general-claims only). Prior-cycle ARC Prize independent value (18.6) is no longer present on the ARC Prize public leaderboard as of 2026-06-03 and has been nulled pending re-submission.
Model
Score
Not reported
Notes
xAI has not published a Grok 4.3 model card; the rollout was a silent model-selector update with no formal announcement. ARC-AGI-2 not reported.
Model
Score
Not reported
Notes
Mistral did not report ARC-AGI-2 on the Medium 3.5 launch post.
Model
Score
Not reported
Notes
Meta's Muse Spark launch blog did not include an ARC-AGI-2 score; the model is closed-weights and access is gated, limiting independent re-runs.
Model
Score
Not reported
Notes
Qwen3.7-Plus is the multimodal vision+language flagship that replaced Qwen3.7-Max as the current Qwen flagship on 2026-05-31. The launch did not publish ARC-AGI-2; the only published benchmark at launch was Vision Arena rank (#16). Independent ARC-AGI-2 re-runs not yet present on the ARC Prize public leaderboard.

GPQA Diamond

KnowledgeApproaching saturation

Graduate-level physics, chemistry, and biology multiple-choice questions written by domain experts and validated to be “google-proof”. The Diamond subset (~198 questions) is the hardest tier.

Author
Rein et al. · NYU · Cohere
arxiv.org/abs/2311.12022
Human baseline
PhD-level experts in matched domains score ~65%; non-experts with web access ~34%

Saturation: Frontier scores cluster near the top — this benchmark separates labs by points, not by capability.

Model
GPT-5.5
OpenAI
Score
●93.6□93.5
Notes
Model
Score
●92.6□92.6
Notes
GPQA Diamond 92.6 per Fable 5's launch materials (2026-06-09, via Vellum's launch-day breakdown). Anthropic characterizes GPQA as state-of-the-art / effectively beaten at the frontier and leads with SWE-Bench Pro and FrontierCode rather than GPQA.
Model
Score
□92.2
Notes
Gemini 3.5 Flash model card does not report GPQA Diamond; emphasizes agentic-coding (Terminal-Bench 2.1, MCP Atlas, OSWorld-Verified) and multimodal (CharXiv, MMMU-Pro) instead.
Model
Score
●90.1□88.8
Notes
Lab-claimed score per DeepSeek's V4 launch chart (the body text only carries general capability claims, no specific GPQA number) and corroborated by multiple third-party reviewers reporting 90.1 from the official April 24, 2026 announcement. Prior cycle had 86.7 which appears to have been a chart-OCR misread. The DeepSeek V4 tech report is the more rigorous source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf.
Model
Score
□90.1
Notes
Carried forward 88.5 from prior cycle would cite a 404 URL (x.ai/news/grok-4-3); xAI never published a formal Grok 4.3 announcement post or model card. Per the data-freshness rule, lab_score nulled until a primary xAI source is found.
Model
Score
●89.5□88.4
Notes
Lab-claimed score per the launch chart (the blog body text does not name a GPQA number; chart-OCR retained from prior cycle).
Model
Score
□74.8
Notes
Mistral did NOT publish GPQA Diamond in the Medium 3.5 launch post. The launch post's body text only names SWE-Bench Verified (77.6) and τ³-Telecom (91.4) as the headline numbers; prior-cycle 82.4 cited the launch chart but is not text-verifiable on the official post and is contradicted by multiple third-party reviewers explicitly noting that Mistral skipped MMLU / GPQA / AIME / HumanEval / MATH for this release. Removed pending primary-source confirmation.
Model
Score
Not reported
Notes
Qwen3.7-Plus launch did not publish GPQA Diamond. The Plus model is the multimodal half of the 3.7 generation, with the launch positioning on vision (Vision Arena #16) rather than reasoning-knowledge benchmarks. Independent Artificial Analysis GPQA Diamond re-run for qwen3.7-plus pending.

MMMU

MultimodalApproaching saturation

Massive Multi-discipline Multimodal Understanding & Reasoning — 11.5K college-exam-level questions across 30 subjects mixing text with diagrams, charts, and images. The canonical multimodal benchmark.

Author
MMMU Benchmark · University of Waterloo + collaborators
mmmu-benchmark.github.io
Human baseline
College students with web access ~83%; domain experts ~90%

Saturation: Frontier scores cluster near the top — this benchmark separates labs by points, not by capability.

Model
Score
●81.4
Notes
Natively multimodal across text, image, audio, and video tokenizers.
Model
Score
●76.4
Notes
DeepSeek's multimodal pipeline trails the US frontier labs; the V-series is text-and-code-first.
Model
Score
Not reported
Notes
Anthropic's Fable 5 launch table does not report the plain MMMU benchmark; multimodal capability is described via GDP.pdf vision (Fable 5 leads at 29.8) and screenshot-to-code demos rather than an MMMU number.
Model
Score
Not reported
Notes
Gemini 3.5 Flash model card reports MMMU-Pro (83.6%, no tools) and CharXiv Reasoning (84.2%, no tools), but not the plain MMMU benchmark tracked here.
Model
GPT-5.5
OpenAI
Score
Not reported
Notes
OpenAI's GPT-5.5 announcement post reports MMMU-Pro (81.2 no tools / 83.2 with tools), not the plain MMMU tracked here. The GPT-5.5 system card (deploymentsafety.openai.com/gpt-5-5, checked 2026-06-13) carries no plain-MMMU score. Prior-cycle value (85.6) was not text-verifiable on any OpenAI primary source and is permanently removed; this cell stays null unless OpenAI republishes plain MMMU for GPT-5.5.
Model
Score
Not reported
Notes
xAI did not publish a formal Grok 4.3 announcement post or model card. Prior-cycle value (79.8) cited a now-404 URL (x.ai/news/grok-4-3) and is removed pending a primary source.
Model
Score
Not reported
Notes
Mistral did NOT publish MMMU in the Medium 3.5 launch post. Medium 3.5 is described in the launch post as the first Mistral flagship to merge multimodal vision with chat in a single weights set, but the launch post body text does not name an MMMU number. Prior-cycle 75.2 cited the launch post URL but is not text-verifiable; removed pending primary-source confirmation.
Model
Score
Not reported
Notes
Despite being the multimodal flagship, Qwen3.7-Plus launch did not publish MMMU. The launch reported Vision Arena rank (#16 overall, #5 lab) as the headline multimodal signal; MMMU was not surfaced. Independent Artificial Analysis MMMU-Pro coverage for qwen3.7-plus pending.

SWE-Bench Verified

CodingApproaching saturation

The 500-issue human-verified subset of SWE-bench. The canonical ‘can the model do real software work’ benchmark from 2024–2025; partially saturated in 2026 but still cited because most frontier models report it.

Author
OpenAI · Princeton NLP
openai.com/index/introducing-swe-bench-verified
Human baseline
Subset construction targets human-solvable issues; success rate not directly comparable to model pass@1

Saturation: Frontier scores cluster near the top — this benchmark separates labs by points, not by capability.

Model
Score
●95.5
Notes
SWE-Bench Verified 95.5 per Fable 5's launch materials (2026-06-09). The harder Pro variant (80.3) is the launch headline; Verified is reported as a secondary, near-saturated number (six models from four labs cluster near 80% on the vendor-reported tracker).
Model
Score
●80.6
Notes
Lab-claimed score per DeepSeek's V4 launch chart (the announcement body text only carries general capability claims, no specific SWE-Bench Verified number) and corroborated by multiple third-party reviewers reporting 80.6 from the official April 24, 2026 announcement, with CAISI's independent reproduction also matching. Prior cycle had 78.9 which appears to have been a chart-OCR misread (same failure shape as yesterday's GPQA Diamond 86.7→90.1 fix on this row). The DeepSeek V4 tech report is the more rigorous source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf.
Model
Score
●77.6
Notes
Per the launch post body text: "Mistral Medium 3.5 scores 77.6% on SWE-Bench Verified, ahead of Devstral 2 and models like Qwen3.5 397B A17B."
Model
Score
●77.4
Notes
Model
Score
Not reported
Notes
Gemini 3.5 Flash model card does not report SWE-Bench Verified; it reports SWE-Bench Pro instead.
Model
GPT-5.5
OpenAI
Score
Not reported
Notes
OpenAI's GPT-5.5 announcement post pivoted to SWE-Bench Pro (58.6%) and does not surface SWE-Bench Verified. The GPT-5.5 system card (deploymentsafety.openai.com/gpt-5-5, checked 2026-06-13) references SWE-Bench Verified only as a linked definition, with no GPT-5.5 score. Prior-cycle 87.3 was not text-verifiable on any OpenAI primary source and is permanently removed; this cell stays null unless OpenAI republishes SWE-Bench Verified for GPT-5.5. The Gemini 3.5 Flash model card cross-references GPT-5.5 SWE-Bench Pro 58.6, the cell populated above.
Model
Score
Not reported
Notes
xAI did not publish a formal Grok 4.3 announcement post or model card. Prior-cycle value (82.6) cited a now-404 URL (x.ai/news/grok-4-3) and is removed pending a primary source.
Model
Score
Not reported
Notes
Qwen3.7-Plus launch did not publish SWE-Bench Verified.

About this page

Cross-family comparison page in the /ai/ section. The roster is the current frontier flagship from every major lab on this site — Claude, GPT, Gemini, Grok, Llama / Muse, DeepSeek, Mistral, Qwen — matched to /ai/models/ so each row links back to the per-family version page for the full lineage.

Lab-claimed vs. independent. Each cell can carry two values. The lab-claimed score (filled circle) is what the lab published in its announcement post, system card, or model card — the lab chooses the configuration (extended thinking, tool use, eval subset). The independent re-run (open square) is what Epoch AI, Artificial Analysis, Aider, or the benchmark’s own public leaderboard reports under their own protocol. When the two diverge by more than 5 percentage points, the page flags it — the gap is the editorial signal that matters here. Where no independent re-run exists yet, the cell shows the lab number alone; closed-weights labs are harder to re-run, so independent coverage is concentrated on the open-weights side.

Benchmark selection. Seven benchmarks covering reasoning, knowledge, coding, math, and multimodal capability. Picked for citation volume (every frontier launch reports these), discrimination (the score band is wide enough to separate labs), and primary-source availability (the benchmark author publishes a leaderboard or the eval protocol is public). Vendor-only proprietary benchmarks that no other lab reports are excluded. LMArena’s Elo is widely cited but is a different measurement type (human preference voting, not standardized eval); the page omits it but the LMArena leaderboard covers that signal.

Saturation framing. A benchmark is treated as discriminating when frontier scores span at least 20 percentage points, approaching saturation when the band tightens below that, and saturated when all frontier models cluster within a few points of the ceiling. Saturated benchmarks (HumanEval, MMLU, HellaSwag, GSM8K) are intentionally omitted from the v1 matrix; they no longer separate labs by capability. The saturation labels are re-evaluated on every refresh.

Sources. Primary lab announcements: Anthropic at anthropic.com/news, OpenAI at openai.com/index, Google at blog.google/technology/google-deepmind, xAI at x.ai/news, Meta at ai.meta.com/blog, DeepSeek at api-docs.deepseek.com/news, Mistral at mistral.ai/news, Alibaba at qwenlm.github.io/blog. Independent re-runners: Epoch AI, Artificial Analysis, Aider polyglot, ARC Prize Foundation.

Refreshed on every major model launch and at least monthly between launches. The page’s job is to stay current within a release cycle; the worst failure mode is showing a stale lab number after that lab has shipped a newer flagship.

Last verified: July 6, 2026. 8 frontier models · 7 benchmarks · 8 labs.