Foxxe Labs Research

AI Model Performance Timeline

Two charts, two stories. The first tracks MMLU — the knowledge benchmark that defined 2022–2025 — across every major OpenAI and Anthropic release. The headline: Claude entered 2023 thirteen points behind GPT-4 and had fully caught up by mid-2024. Neither lab reports MMLU any more, so the 2026 tail is shaded as estimate. The second chart covers where the frontier is actually contested. Both labs shipped in the first week of September — Claude Fable 5.1 on the 1st, GPT-6 Astra on the 3rd — and neither published a full sweep, so the bars name the model that actually produced each figure rather than implying one pairing. The headline is ARC-AGI-3: the widest gap on this page has reversed, with Astra claiming 99.9% on a stateful harness against Opus 5's 30.2%. Read the notes before the gaps; several of these numbers are measured under settings that are not directly comparable. Hover any data point for model details and source notes.

Last updated: 11 September 2026 · Todd McCaffrey, Foxxe Labs

Part 1 — Knowledge benchmark (MMLU), 2022–2026

MMLU was the dominant benchmark from 2022 to 2025, tracking broad knowledge across 57 academic disciplines. Neither lab reports it in launch materials any more: both sit above 96% and the spread between them is inside the noise. The historical arc is still the clearest picture of how fast the gap closed — Claude entered 2023 thirteen points behind GPT-4 and had caught up by mid-2024.

+27.8pp

OpenAI MMLU gain

GPT-3.5 → GPT-6 Astra (3.8 yrs)

+24.6pp

Claude MMLU gain

Claude 1 → Claude Fable 5.1

8 releases

Anthropic 2026 cadence

Opus 4.6 → Fable 5.1, Feb–Sep 2026

ARC-AGI-3

Widest live gap

99.9% vs 30.2% — reversed in September

OpenAI / ChatGPTAnthropic / ClaudeShaded region = estimated, no longer reported

MMLU BENCHMARK (%) — HIGHER IS BETTER

OpenAI / ChatGPT

Sep '26

GPT-6 AstraLATEST

99.9% ARC-AGI-3 on a stateful harness; 72.6% OSWorld 2.0; 97.6% FrontierMath T4

97.8% est

Jul '26

GPT-5.6 Sol

Sol / Terra / Luna tiers; 96.2% SWE-bench Verified; $5/$30 per MTok

97.4% est

Apr '26

GPT-5.5

Codename 'Spud'; first fully-retrained base since GPT-4.5

96.5% est

Mar '26

GPT-5.4

Native computer use; 1M context; merges the Codex line

97.2% est

Feb '26

GPT-5.2

Improved polish; spreadsheet & finance tasks

96.8% est

Nov '25

GPT-5.1

Adaptive reasoning; 8 personality presets

96.4%

May '25

GPT-5

Unified reasoning + conversational AI

96%

Apr '25

o3 / o4-mini

Top AIME 2024/25 benchmark scores

96.7%

Jan '25

o3-mini

Compact, fast reasoning

93%

Dec '24

Full o1 + $200/mo Pro tier

92.3%

Sep '24

o1-preview

Chain-of-thought reasoning model

90.8%

May '24

GPT-4o

Omni-modal; free-tier access

88.7%

Nov '23

GPT-4 Turbo

128K context; updated knowledge cutoff

87.4%

Mar '23

GPT-4

Major leap; launched in Bing and ChatGPT

86.4%

Nov '22

GPT-3.5

ChatGPT launch — 1M users in 5 days

70%

Anthropic / Claude

Sep '26

Claude Fable 5.1LATEST

81.2% SWE-bench Pro; cache reads cut 4x to $0.25/MTok; Mythos 5.1 is the same model, restricted

97.6% est

Jul '26

Claude Opus 5

96.0% SWE-bench Verified; 1M context; Opus pricing held at $5/$25

97.5% est

Jun '26

Claude Sonnet 5

GA 30 Jun; 85.2% SWE-bench Verified at $2/$10 per MTok

96.4% est

Jun '26

Claude Fable 5

95.0% SWE-bench Verified, 80.0% Pro; export-control pause 12 Jun – 1 Jul

97.2% est

May '26

Claude Opus 4.8

88.6% SWE-bench Verified; 41-day release cadence

96.9% est

Apr '26

Claude Opus 4.7

87.6% SWE-bench Verified; xhigh effort tier

96.7% est

Mar '26

Claude Sonnet 4.6

Cost-efficient tier; 12 days after Opus 4.6

96.2% est

Feb '26

Claude Opus 4.6

1M context beta; adaptive thinking; agent teams

96.5% est

Nov '25

Claude 4.5 Opus

First model >80% SWE-bench (80.9%)

95.8%

Sep '25

Claude 4.5 Sonnet

77.2% SWE-bench; 30+ hour task focus

95%

May '25

Claude 4 Opus

ASL-3 safety classification; Claude Code

94.5%

Feb '25

Claude 3.7 Sonnet

First hybrid reasoning model

91%

Oct '24

Claude 3.5 Sonnet v2

Computer use capability; upgraded Haiku

89.5%

Jun '24

Claude 3.5 Sonnet

Beats Opus at Sonnet price; Artifacts

88.7%

Mar '24

Claude 3 Opus

Multimodal; noted self-awareness in tests

86.8%

Nov '23

Claude 2.1

200K context; reduced hallucination rate

80%

Jul '23

Claude 2

General public access; 100K context

78.5%

Mar '23

Claude 1

First public Claude via limited API

73%

Part 2 — Where the frontier is contested (September 2026)

MMLU measures what a model knows. These measure what it can do: ship a repository-level change, drive a terminal, operate a desktop, and reason about problems it has never seen. Both current flagships — OpenAI's GPT-5.6 Sol (9 July) and Anthropic's Claude Opus 5 (24 July) — cleared 96% on SWE-bench Verified, so the live comparison has moved to SWE-bench Pro, OSWorld 2.0, Frontier-Bench and ARC-AGI-3. Hover any bar for the source note.

Benchmarks moved under the models this year. OSWorld went to the harder 2.0 set, Terminal-Bench to 2.1, SWE-bench to Pro, and GDPval-AA to v2. Scores below are not comparable to the 2025 numbers that carried the same benchmark names.

OpenAIAnthropicSCORE (%) — HIGHER IS BETTER

GDPval-AA v2

Economically valuable knowledge work across 44 professions, scored as an Elo rating

GPT-5.6 Sol1,736 Elo

Claude Opus 51,861 Elo

Opus 5 leads by 125 Elo — roughly a 67% expected win rate head-to-head. Elo is an interval scale and is not plotted alongside the percentage benchmarks. Bars start at 1,650 Elo, not zero.

SWE-bench Verified — saturated

Reported for reference, not plotted: the spread no longer separates the frontier.

GPT-5.6 Sol96.2%

Claude Opus 596.0%

Claude Fable 595.0%

Claude Sonnet 585.2%

Four points separate the two flagships. Verified is finished as a differentiator; the contested version is Pro. Neither September release adds a row — Anthropic published no Verified score for Fable 5.1, and the widely-quoted 95.0% belongs to Fable 5.

SWE-bench Pro

Repository-level software engineering — harder and less contaminated than Verified

GPT-5.6 Sol64.6%

Claude Fable 5.181.2% ▲

Fable 5.1 (Sep 2026) takes the Anthropic figure to 81.2, up from 79.2 for Opus 5. OpenAI published no SWE-bench Pro score for GPT-6 Astra, so the OpenAI bar is still GPT-5.6 Sol and the gap shown is not a like-for-like September comparison.

Terminal-Bench

Agentic coding in real terminal environments — planning, iteration, tool coordination

GPT-5.6 Sol89.5% ▲

Claude Opus 589.1%

Effectively a tie (0.4 points, Sol at xhigh vs Opus 5 at max effort). GPT-5.6 Sol Ultra reaches 91.9%. Not updated for September: Fable 5.1 reports against Terminal-Bench 4.0 (55.8%) and Terminal-Bench-Science (52.6%), which are different, harder sets and cannot be put on this bar.

OSWorld 2.0

Autonomous desktop navigation via screenshots and keyboard/mouse

GPT-6 Astra72.6%

Claude Fable 5.177.9% ▲

Both September flagships, but read the settings before reading the gap: Fable 5.1 scores 77.9 on the partial setting and only 41.7 strict, and Anthropic ran with production safeguards on, taking zeros where a classifier intervened. OpenAI reports 72.6 for Astra against 65.7 for Sol. OSWorld moved to the harder 2.0 set in 2026, so none of this is comparable to the ~80% figures on OSWorld-Verified.

Frontier-Bench

Long-horizon agentic coding on unsolved problems

GPT-5.6 Sol34.4%

Claude Opus 543.3% ▲

Both labs are under 50%. The newest benchmark on this page and the least saturated.

ARC-AGI-3

Novel reasoning on problems absent from any training set

GPT-6 Astra99.9% ▲

Claude Opus 530.2%

Astra saturates ARC-AGI-3 at 99.9% and turns the widest gap on this page completely around — but the headline number depends on a stateful, expensive harness, and stateless API calls score far lower. Anthropic has published no ARC-AGI-3 figure for Fable 5.1, so the Anthropic bar is still Opus 5. Treat this pairing as the least settled on the page.

Benchmark notes: MMLU is official where published. Entries marked est (February 2026 onward, shaded on the chart) are estimates positioned from the ArtificialAnalysis Intelligence Index — neither lab reports MMLU in launch materials any more. Capability scores are from each model's launch post and the comparison round-ups that followed, and reflect each lab's highest published effort setting. They are not all from the same release: where a lab published no figure for a benchmark with its September model, the bar names the last model that did, and that benchmark's note says so. Anthropic's September figures were measured with production safeguards on, taking zeros where a classifier intervened, which makes them conservative. GDPval-AA v2 is an Elo rating and is shown on its own scale rather than on the percentage axis. Benchmark versions changed during 2026 — OSWorld 2.0, Terminal-Bench 2.1, SWE-bench Pro, GDPval-AA v2 — so these are not comparable with earlier figures published under the same names. Last updated 11 September 2026.

Live tracking: llm-stats.com · Artificial Analysis · ARC Prize leaderboard

Interactive · full screen

The Expanding Universe of Large Language Models

An observatory globe that grows the whole model landscape from the 2017 Transformer to today — a curated 135 models plotted on the city of their lab, sized by parameters, raised by capability, and coloured by maker, with a playable timeline you can scrub to any month. Follows the site theme.

Open the globe →