Projecting fully autonomous AI research at Anthropic

Research note · 6 November 2026 · Data (CSV)

Extending Anthropic's own R&D automation data with METR's independent measurement of AI R&D acceleration puts the median date for fully autonomous AI research at Anthropic in December 2027. That's when AI works with no human in the loop on at least 95% of the company's model R&D work, which is Automation Level 5 on Epoch AI's scale.

There's a 50% chance it falls between July 2027 and September 2028, and an 80% chance between May 2027 and February 2030. The projection assumes development isn't slowed on purpose and that no mandatory human sign-off is required.

The data

Until recently there was no public time series for how much of a frontier lab's research AI does. Anthropic's note Measurements for understanding the pace of AI development inside frontier labs changed that. It released an R&D Automation Index (version 2026.07) with monthly readings from August 2025 through August 2026, scoring every category of model R&D work at the company on Epoch AI's automation scale. The scale goes from AL0, with no AI involvement, through AL3 ("AI collaborates") and AL4 ("AI leads") to AL5, where AI runs on its own with no human in the loop.

On 5 November 2026 METR published an independent assessment of AI R&D acceleration at Anthropic, done with insider access. METR reckons that since October 2025 Anthropic's benchmark scores have grown at 1.9× their earlier rate (95% CI 1.5–2.8). Its best guess is that roughly 80% of the current acceleration comes from AI, and it found no big bottleneck for the next 12 months.

Each automation level has moved through Anthropic's R&D task basket as its own S-shaped wave. The table shows the shares read from the published chart at pixel precision. That chart also has 90% measurement intervals for the AL4 series, and the fit uses them.

In twelve months the share at AL3 or above went from 3.8% to 95.6%, and in about seven the share at AL4 or above went from zero to 26.3%. AL5 is still at zero for every measured category of work. The projection below tries to pin down when the AL5 wave, where people leave the loop, will be complete.

Method

Each automation level is treated as a logistic wave across the task basket. Waves AL1 to AL4 are fit to the 13 monthly observations in logit space, with measurement error from the published intervals. Each wave gets its own process-noise term, set so that chi-squared per degree of freedom is about 1 with a floor of 0.2 logit units. Months near 0% or 100% are censored, and the parameter uncertainty is widened for residual autocorrelation.

There's nothing to fit for AL5 yet. Its wave is projected from how the spacing and steepness of the earlier waves have changed, and it has to stay at zero through August 2026.

The data can't settle how the waves are driven or whether AL4's recent slowdown will last, so the projection averages four variants that differ on both.

In the feedback variants the waves move in effective R&D time. Automation cuts the human time each task needs, that speeds up R&D, and faster R&D brings the next wave forward. The pace function is tuned to METR's evidence by importance sampling, drawing on all 16 AECI benchmark observations before and after the October 2025 trend break, METR's judgment that the current pace is about 2× (1.3–3×), its estimate that about 80% of the acceleration comes from AI, and the growth of staff-reported uplift. A compute ceiling of 6–40×, from published estimates of compute bottlenecks in AI R&D, caps the feedback. In the calendar variants the waves just keep the rhythm they have shown so far. The feedback variants get two thirds of the weight (see "Evidence behind the weights").

AL4 has slowed since May 2026. From June to August the AL4 wave rose at about 60% of the AL3 wave's average pace in logit units. AL3 paused in a similar way before it jumped in March 2026, when Anthropic got a much stronger model (Mythos Preview). Variants that fit the whole AL4 series and variants that fit only its recent months get equal weight.

Results

The four variants put the median for full automation in July 2027 (feedback, full AL4 series, weight one third), December 2027 (calendar, full series, one sixth), April 2028 (feedback, recent AL4 months, one third) and September 2028 (calendar, recent months, one sixth). Most of the width comes from this spread between variants. Parameter uncertainty inside any one variant is smaller, so both open questions stay in the average.

In the feedback variants the resulting pace of R&D is 2.5× the 2024–25 rate in October 2026. That's above METR's central judgment of about 2× and inside its 1.3–3× range. That number is a snapshot, while the 1.9× AECI estimate averages the trailing year. The pace reaches about 5× by the median full-automation date, and after that the compute ceiling starts to bite.

Full automation here means AL5 on at least 95% of the work because the index can't tell 95% from 100%. The earlier waves top out between 96% and 100% on the published chart, and the measurement intervals cover several points. Anthropic also says its model judge agrees exactly with human raters only 59% of the time. With AL5 on 95% of the basket and AL4 on the rest, people would still do about 0.4% of the original human work.

Validation and sensitivity

I back-tested the model on held-out months. Fitted only on data through April 2026, when AL4 stood at 2.6%, the feedback variants predicted a median of 24.5% for August (80% band 7.6–74.5%). The actual value was 26.3%. The calendar variants predicted 15.7% (5.8–45.5%). Fitted through January 2026, the feedback and calendar variants both put all four held-out AL4 values inside their 80% bands, though at that horizon the bands are very wide.

The method does poorly when a wave changes pace partway through. Fitted through June 2026, partway into AL4's slowdown, both kinds overshoot July and August. For August the feedback variants predicted 73.5% and the calendar variants 60.1%, against 26.3% observed. This is why the recent-months variants keep half the weight.

Setting METR's current-pace judgment to either end of its range (1.3× or 3×), the feedback variants' median moves by about a month either way, because the AECI series and the automation data already pin down the pace. A wider 8–100× prior for the compute ceiling, in place of 6–40×, moves no variant's median by more than a month. With full automation defined at 90% or 99% of the work, the median becomes November 2027 or January 2028.

Evidence behind the weights

Anthropic's recursive self-improvement page (updated 18 September 2026) says Claude's success on the most open-ended engineering tasks rose from about 26% in August 2025 to 76% in May 2026 and 91% in September 2026. The biggest jump came in March 2026, after Mythos Preview arrived internally, which is also when the AL3 wave jumped. By METR's model dating, new internal frontier snapshots arrived about once a month from April to September 2026, roughly three times as often as in 2024.

Anthropic's amended S-1 filing of 2 November 2026 says that Claude wrote over 90% of the contributions to its codebase in September 2026. It adds that its internal estimates and expectations for future models suggest it may soon meet its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI". It hasn't called any current model past that threshold. METR's review of the AI R&D section of Anthropic's August 2026 Risk Report says it can't rule out a big further acceleration over the next six months.

But evidence cuts the other way too. Anthropic's Opus 5.5 system card (September 2026) shows that a one-time capability jump at Mythos Preview fits its internal index better than a lasting change of slope in 99 of 100 resampled fits. METR, which saw newer internal models, now thinks a lasting change is more likely. And public models from July to September 2026 improved only in small steps. Anthropic's August 2026 Risk Report says researchers use only moderate amounts of AI because they don't trust the models on steps requiring calibration, self-monitoring or judgment, and adds that the gap is narrowing. If that trust builds on an organisational timescale that model releases don't change, the waves would keep a calendar rhythm. This is why the calendar variants keep a third of the weight.

The projection assumes no deliberate pacing, and that may be wrong. Anthropic's amended S-1 says it doesn't expect to be able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision, even once it could. It's weighing changes to its Responsible Scaling Policy to keep humans in control. Its Frontier Safety Roadmap aims to publish, by 15 January 2027, a policy framework for targeting the pace of AI development or an explanation of why it decided against one. If measures like these take hold, the dates above show when the work could be automated, and work at AL5 level might still run at AL4 by policy.

Limitations

• The data covers one lab and one index. The task basket was frozen in July 2026, so work that came along later, such as overseeing agents, isn't counted. Anthropic found no rise in new kinds of tasks between January and July 2026 and plans to redo the basket from time to time.
• AL5 on 95% of the July 2026 task categories is necessary for a model to build its successor on its own, and it may not be enough. People could keep research direction and final sign-off even when nearly every task is autonomous.
• The last step may be different in kind. Moving from AL4 to AL5 removes the human check itself. If dropping that check happens on an organisational-trust timescale, the later variants are the better guide. The calendar variants capture part of this.
• Much of the evidence is redacted. METR says Anthropic redacted most of the evidence it found most useful, including METR's own forecasts for future models. This analysis works with the public AECI series that's left.

A few of these predictions can be checked soon. The next Automation Index release should show AL4 or above passing 50% of work between September and November 2026 (80% interval). The median for a first AL5 reading is May 2027. If AL4 or above is still well below 50% in late 2026, every date moves later and the slower variants look better. An AL5 reading before mid-2027 would back the faster ones.

Sources

• Anthropic, Measurements for understanding the pace of AI development inside frontier labs (R&D Automation Index v2026.07, data to August 2026).
• METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 (5 November 2026).
• METR, review of the AI R&D section of Anthropic's August 2026 Risk Report (5 November 2026).
• Anthropic, Recursive self-improvement (updated 18 September 2026).
• Anthropic, Risk Report, August 2026 (redacted) and Claude Opus 5.5 system card (September 2026).
• Anthropic, PBC, Form S-1/A (SEC, 2 November 2026), and Frontier Safety Roadmap (goals as of 19 October 2026).
• Compute-bottleneck estimates from Davidson and Houlden (Forethought, 2025), Whitfill and Wu (2025) and the AI Futures Project model (2026).

This note is an independent analysis by The Singularity Gate. Anthropic and METR didn't review or endorse it, and neither has published a forecast of these dates. The numbers were read from published charts and may differ slightly from the raw data. The projection follows current trends under the conditions stated above and says nothing about any specific future model. The monthly projected series is available as a CSV file.