Two businessmen enthusiastically analyzing market charts on computer monitors in a modern office
TrackersBased on company claims

AI research acceleration tracker: what OpenAI and Anthropic have disclosed, dated

By Himanshu Sakre·

Photo: AlphaTradeZone / Pexels

A living log of the public measures of how much AI is speeding up AI research at the frontier labs, with each figure dated and attributed, as of November 21, 2026. Every number here is the lab's own unless marked.

This page tracks the public evidence on a single question: how much is AI speeding up the work of building AI? It lists what each frontier lab has actually published, with dates, and what is still missing. It is updated through our corrections process when a lab releases new figures. Last updated: November 21, 2026.

The reason to track it is simple. Arguments about recursive self-improvement are only as good as the measurements behind them, and until this autumn there were few. Our explainer on recursive self-improvement defines the strict and loose readings. This page records the data.

OpenAI: usage, success rates and a target date

September 6, 2026. In "Research acceleration: The view inside OpenAI," the company said that, by its own measurements, it had reached the goal announced the previous fall of having an automated research intern by September. It defined that as a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. It added that it was making strong progress toward an automated AI researcher by March 2028.

November 18, 2026. In "Research acceleration: The jagged frontier," OpenAI added new measures. Unaided agent success on tasks estimated at two to four human hours rose from 23% in March to 46% in October, and for four-to-eight-hour tasks from 13% to 35%. The median researcher's agent use reached $2,400 a day at API list prices in October, and the 90th percentile $23,000. Agent workdays per researcher per day rose from 0.5 to 5.7. Safety and Alignment documented workweeks rose 77% from January to September.

The same report records the limits. From August 1 to October 18, unaided success was 79% for monitoring training runs, 35% for launching them and 25% for inference and serving code. For GPT-6 Astra, researchers said people originated the ideas they named. For GPT-6.2 Astra, pretraining researchers described working with Codex. Read our full report analysis for the caveats.

Anthropic: a threshold, not a series

November 2, 2026. Anthropic's amended registration statement does not publish a time series like OpenAI's. It says instead that "our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet" its Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI. It adds that it has not designated any current model as exceeding that threshold, that it continues to assess the underlying data, and that it will release further reports.

“we may imminently meet our Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.””

Anthropic, PBC, Form S-1/A, filed November 2, 2026

November 5, 2026. METR published an independent assessment related to Anthropic's AI research and development acceleration, which we covered in our METR analysis.

Cryptocurrency analysis on a digital screen with candlestick charts showing market trends and volatility
Analysts reviewing performance charts: the kind of data labs are now starting to publish. Photo: Rafael Minguet Delgado / Pexels

How the measures compare

The two disclosures do not line up. OpenAI reports inputs and process: spend, agent time, success rates by task length and survey responses, plus an internal capability index. Anthropic reports a policy threshold, tied to its own safety framework, with the evidence partly withheld. A reader cannot convert one into the other.

OpenAI itself calls for the missing piece: "we believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI)." It notes that other organizations have begun to publish related work, and hopes the report will help sustain shared measurement norms.

What would change this page

  • A third lab, such as Google DeepMind or Meta, publishing comparable usage and success-rate data.
  • An independent evaluator confirming OpenAI's task-duration estimates.
  • Anthropic stating that it has met the threshold, or releasing the underlying series.
  • A regulator or legislator writing the reporting requirement into law.
  • A change in how any lab defines its tasks, which would break comparisons over time.

Definitions used on this page

Unaided success rate. OpenAI's estimated share of tasks an agent completes without human intervention, bucketed by how long the task would take a person. The task categories follow the O*NET AI R&D taxonomy developed by Epoch AI. OpenAI notes that researchers may change which tasks they delegate as tools improve, so a rising rate can reflect both capability and choice.

Agent workday. Every eight hours of agent use counted as one workday. OpenAI says it generally covers interactive use only and may undercount automation.

OpenAI ECI. An OpenAI-specific internal version of the Epoch Capabilities Index, introduced in the November 18 report to track model capability gains.

Dramatically accelerated progress attributable to AI. The wording of Anthropic's Responsible Scaling Policy threshold, quoted in its S-1/A. Anthropic says it has not designated any current model as exceeding it.

The survey evidence

OpenAI also surveyed 101 research staff from late August to early September 2026. Asked what they would trade to restore frontier coding agents after a hypothetical three-month loss of access, the median answers were 40% of team headcount and 30% of current team compute, rising to 50% of compute if asked in January 2027. Staff estimated a median of three months to match one month of their previous output without agents. OpenAI flags the limits of self-reports and the ad hoc sampling.

How to read it

Treat each row as a claim with a date. Self-reports are a start, since they were rare a year ago, but they come from organizations that benefit from the story. Success rates also depend on which tasks researchers choose to delegate, which OpenAI notes can change as tools improve. Our position: the right test is replication by someone outside the lab, and the right policy ask is the one OpenAI already made, mandatory and standardized reporting. Until that exists, this ledger is the best public record, and we will keep it current.

Frequently asked questions

How much is AI speeding up AI research?

No audited figure exists. OpenAI reports that agents completed an estimated 46% of two-to-four-hour research tasks without human help in October 2026, up from 23% in March, and calls uplift jagged. Anthropic says it may imminently meet its acceleration threshold.

Has OpenAI achieved an automated AI researcher?

Not by its own account. On September 6, 2026, OpenAI said it had reached an automated research intern and was making strong progress toward an automated AI researcher by March 2028. Its November 18 report says humans remain central.

What does Anthropic say about AI accelerating its research?

Its S-1/A, filed November 2, 2026, says its internal estimates indicate it may imminently meet its Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI. It has not designated any current model as exceeding that threshold.

What are indicators of recursive self-improvement?

They are measurable signs that AI is accelerating the development of future AI, such as agent usage, task success rates, code output and capability indexes. OpenAI argues labs should be required to track and publicly report them.

Is any of this independently verified?

Not in the sources we reviewed. The OpenAI figures are internal measurements and surveys. METR has published an assessment concerning Anthropic, which we cover separately, but neither lab's full series has been audited in public.

How often is this tracker updated?

We update it when a lab publishes new figures, through our corrections process, and each entry carries its date. This version reflects disclosures through November 21, 2026.

Sources

What each one is, and whose it is.

  1. 1

    Research acceleration: The jagged frontier, OpenAI (November 18, 2026)

    Vendor announcement
  2. Vendor announcement
  3. 3

    Anthropic, PBC Form S-1/A, Amendment No. 1, U.S. Securities and Exchange Commission (November 2, 2026)

    FilingThe vendor’s own