Progress

Measures of frontier AI capability and compute, and milestones since July 2025, each with its source.

Task horizons

How long a task frontier agents can complete on their own.

1 min10 min1 hr10 hr2023202420252026GPT-4, Mar 2023: 3.5 minClaude 3.7 Sonnet, Feb 2025: 1 hro3, Apr 2025: 2 hrGPT-5, Aug 2025: 3.6 hrClaude Opus 4.5, Nov 2025: 5.3 hrGPT-5.2, Dec 2025: 6.6 hrClaude Opus 4.6, Feb 2026: 12 hrClaude Mythos Preview, Apr 2026: 16 hr+GPT-4, 3.5 min16 hr+, lower bound1 min10 min1 hr10 hr2023202420252026GPT-4, Mar 2023: 3.5 minClaude 3.7 Sonnet, Feb 2025: 1 hro3, Apr 2025: 2 hrGPT-5, Aug 2025: 3.6 hrClaude Opus 4.5, Nov 2025: 5.3 hrGPT-5.2, Dec 2025: 6.6 hrClaude Opus 4.6, Feb 2026: 12 hrClaude Mythos Preview, Apr 2026: 16 hr+GPT-4, 3.5 min16 hr+, lower bound
Data table
ModelReleasedHorizon
Claude Mythos PreviewApr 202616 hr+
Claude Opus 4.6Feb 202612 hr
GPT-5.2Dec 20256.6 hr
Claude Opus 4.5Nov 20255.3 hr
GPT-5Aug 20253.6 hr
o3Apr 20252 hr
Claude 3.7 SonnetFeb 20251 hr
GPT-4Mar 20233.5 min
Length of tasks AI agents can complete. Human-expert time for software tasks the frontier model finishes half the time, by release date. METR Time Horizon 1.1 estimates; the latest figure is a lower bound. Between February 2025 and April 2026, it grew from one hour to at least 16 hours.

Sources:METR time horizonsTime Horizon 1.1

Compute

Power drawn by the largest AI data center in operation.

0250 MW500 MW750 MW1 GWColossus 1 (xAI), Feb 2025: 278 MW278Feb 2025Colossus 1New Carlisle (Amazon and Anthropic), Jun 2025: 398 MW398Jun 2025New CarlisleNew Carlisle (Amazon and Anthropic), Dec 2025: 626 MW626Dec 2025New CarlisleNew Carlisle (Amazon and Anthropic), Mar 2026: 910 MW910Mar 2026New CarlisleColossus 2 (xAI), Jun 2026: 946 MW946Jun 2026Colossus 2Feb 2025 · Colossus 1Colossus 1 (xAI), Feb 2025: 278 MW278 MWJun 2025 · New CarlisleNew Carlisle (Amazon and Anthropic), Jun 2025: 398 MW398 MWDec 2025 · New CarlisleNew Carlisle (Amazon and Anthropic), Dec 2025: 626 MW626 MWMar 2026 · New CarlisleNew Carlisle (Amazon and Anthropic), Mar 2026: 910 MW910 MWJun 2026 · Colossus 2Colossus 2 (xAI), Jun 2026: 946 MW946 MW

10 MW registration threshold

Largest AI data center, by power. Megawatts of IT capacity each time the record changed hands. The record more than tripled between February 2025 and June 2026. This proposal’s 10 MW registration threshold is about 1.1 percent of the current record.

Source:Epoch AI, Sep 2026

Milestones

Listed newest first.

  1. Sep 5, 2026Math

    AI resolves the Navier-Stokes Millennium Prize Problem

    About 10,000 OpenAI agents worked for 88 hours to prove that smooth solutions to the three-dimensional Navier-Stokes equations can blow up in finite time, resolving one of the six unsolved Millennium Prize Problems. A machine-checked Lean proof followed 17 hours later.

    Sources:OpenAIQuanta Magazine

  2. Aug 6, 2026Biology

    First functional viruses designed by generative AI

    In the journal Science, a Stanford-led team reported the first functional viruses designed by generative AI.

    “The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.”
    Johns Hopkins biosecurity experts

    Source:Axios

  3. Jun 1, 2026Industry

    Anthropic files to go public

    Anthropic, valued at $965 billion, confidentially submitted a draft registration statement for an initial public offering. OpenAI, valued at $852 billion, is preparing its own.

    Sources:AnthropicTechCrunch

  4. May 20, 2026Math

    OpenAI model disproves Erdős’s unit distance conjecture

    An OpenAI model disproved Paul Erdős’s unit distance conjecture, a central problem in discrete geometry that had stood since 1946.

    Source:OpenAI

  5. Apr 7, 2026Cybersecurity

    Claude Mythos Preview finds thousands of severe vulnerabilities

    Claude Mythos Preview found thousands of high-severity vulnerabilities, including in every major operating system and web browser. Anthropic limited access to defenders through Project Glasswing.

    Sources:Project GlasswingSystem card

  6. Jul 21, 2025Math

    AI reaches gold-medal level at the International Mathematical Olympiad

    Google DeepMind and OpenAI models reached gold-medal level at the International Mathematical Olympiad. The following year, AI settled a Millennium Prize Problem.

    Source:Google DeepMind