gapnium

Tier-2 coefficient · weight 0.20

gapnium.autonomy

How long and how independently AI agents can work.

Now0.3190
12 months+0.0970
Since 2020+0.279

Hover the chart to read values · click a marker to open its articles

What it measures

The length, reliability and independence of agentic work: the duration of tasks an agent can complete unsupervised (METR time horizons), real-world software engineering (SWE-bench), computer use (OSWorld), and long-running deployments. A doubling of reliable task length is treated as a large step.

Evidence the sorting agent looks for

  • METR 50% / 80% time horizons
  • SWE-bench, Terminal-Bench, OSWorld
  • Agent products shipped to users
  • Hours-to-days unsupervised runs

What 1.0 would mean

Agents can be handed open-ended, months-long projects and complete them with expert-level reliability and no human in the loop.

Feedback on this coefficient

Events that moved gapnium.autonomy

  1. Claude Opus 5.5: frontier performance at lower cost

    Opus 5.5 reports 66.4% on Terminal-Bench 4.0, 81.8% on OSWorld 2.1 and matches Fable 5.1 on most work for ~40% less than Opus 5.

    Autonomy +0.016 Automation +0.010 Compute +0.006
    Anthropic · Introducing Claude Opus 5.5

  2. Claude Opus 5 sets agentic and knowledge-work records

    Anthropic's Opus 5 comes close to Claude Fable 5 at half the price and scores roughly three times the next-best model on ARC-AGI 3.

    Cognition +0.022 Autonomy +0.016 Automation +0.012
    Anthropic · Introducing Claude Opus 5

  3. GPT-5.5 released

    OpenAI's GPT-5.5 reports 82.7% on Terminal-Bench 2.0, a large step for terminal-based agentic work.

    Autonomy +0.014 Cognition +0.008
    Wikipedia · GPT-5.5 — Wikipedia

  4. Claude Opus 4.6 brings 1M context to Opus

    Opus 4.6 posts state-of-the-art results on Terminal-Bench 2.0, Humanity's Last Exam and BrowseComp, and leads GDPval-AA.

    Autonomy +0.018 Automation +0.010 Cognition +0.008
    Anthropic · Introducing Claude Opus 4.6

  5. METR Time Horizon 1.1 and its limits

    METR expands its task suite to 228 tasks with more than twice as many 8h+ tasks, and cautions that a 50% horizon is not a delegation guarantee.

    Autonomy +0.006 Automation −0.003
    METR · Time horizon limitations

  6. Claude Opus 4.5 outscores human engineering candidates

    Anthropic says Opus 4.5 scored higher than any human candidate on its take-home engineering exam, with 80.9% on SWE-bench Verified.

    Autonomy +0.015 Recursion +0.012
    Anthropic · Introducing Claude Opus 4.5

  7. First reported AI-orchestrated cyber-espionage campaign

    Anthropic reports disrupting a state-sponsored campaign in which an AI agent performed 80–90% of the intrusion work autonomously.

    Autonomy +0.012
    Anthropic · Disrupting the first reported AI-orchestrated cyber espionage campaign

  8. Claude Sonnet 4.5 sustains 30-hour tasks

    Anthropic reports Claude Sonnet 4.5 maintaining focus for more than 30 hours on complex multi-step tasks, and 61% on OSWorld.

    Autonomy +0.020 Recursion +0.006
    Anthropic · Introducing Claude Sonnet 4.5

  9. ChatGPT agent unifies browsing, research and actions

    OpenAI merges Operator and Deep Research into one agent with its own virtual computer.

    Autonomy +0.010
    OpenAI · Introducing ChatGPT agent

  10. Claude 4 works autonomously for hours

    Claude Opus 4 leads SWE-bench (72.5%) and is reported to run a multi-hour open-source refactor independently.

    Autonomy +0.018 Cognition +0.006
    Anthropic · Introducing Claude 4

  11. o3 and o4-mini use tools while reasoning

    OpenAI's reasoning models can agentically search, run code and analyse images inside their chain of thought.

    Cognition +0.010 Autonomy +0.008
    OpenAI · Introducing OpenAI o3 and o4-mini

  12. METR: agent task length doubling every ~7 months

    METR finds the length of tasks frontier agents complete with 50% reliability has doubled roughly every 7 months for six years.

    Autonomy +0.020
    METR · Measuring AI Ability to Complete Long Tasks

  13. Claude 3.7 Sonnet and Claude Code

    Anthropic releases a hybrid reasoning model and Claude Code, an agentic command-line coding tool.

    Autonomy +0.018 Recursion +0.012
    Anthropic · Claude 3.7 Sonnet and Claude Code

  14. Deep Research agents produce analyst-grade reports

    OpenAI's Deep Research autonomously browses for up to 30 minutes and writes cited, multi-page reports.

    Autonomy +0.012 Automation +0.008
    OpenAI · Introducing deep research

  15. OpenAI Operator browses the web for users

    A browser-using agent that can fill forms, order groceries and book travel is released to users.

    Autonomy +0.012
    OpenAI · Introducing Operator

  16. RE-Bench: agents vs. human ML researchers

    METR finds AI agents beat human experts on ML research-engineering tasks at 2 hours, but humans win decisively given 8+ hours.

    Recursion +0.010 Autonomy −0.004
    arXiv · RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts

  17. Claude gains computer use

    Anthropic releases a model that operates a computer by looking at the screen, moving a cursor and typing.

    Autonomy +0.020 Automation +0.005
    Anthropic · Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku

  18. Claude 3.5 Sonnet raises the coding bar

    Claude 3.5 Sonnet outperforms Claude 3 Opus at a fifth of the cost and solves 64% of problems in an internal agentic coding eval.

    Autonomy +0.020 Cognition +0.010 Recursion +0.006
    Anthropic · Introducing Claude 3.5 Sonnet

  19. Devin is billed as an autonomous AI software engineer

    Cognition unveils Devin, an agent that plans and executes multi-step software tasks, resolving ~14% of SWE-bench issues.

    Autonomy +0.018
    Cognition · Introducing Devin, the first AI software engineer

  20. GPT-4 passes professional exams

    GPT-4 scores around the top 10% on a simulated bar exam and accepts image inputs.

    Cognition +0.050 Automation +0.015 Autonomy +0.010
    OpenAI · GPT-4