Recursive self-improvement (RSI) is, broadly, the use of AI in the production of better AI.
What exactly does that mean? How far along is that now? And how much will this accelerate the development of more powerful, and likely dangerous, AI?
We find that AI is speeding up AI development at Anthropic by about 1.5x, and by the end of 2029, this will be a 4x speedup, leading to dramatically faster AI capabilities. By Jun 2032, we predict a lab will state that a released model was principally designed by its predecessor.
The full set of forecasts, and rationales behind them, are below. Below that is the World Model for these forecasts, showing how they flow together.
0.1Lab states that AI selects the majority of its experimentsOct 2030
When will a top-five AI developer by revenue publicly state that AI systems, rather than its human researchers, select the majority of the experiments that inform its frontier model development?
0.2Share of release-informing experiments proposed by AI, today10%
As of 2026-10-16, what share of the experiments that materially informed the most recent frontier model release at the largest AI lab by revenue were proposed by an AI system rather than by a human researcher?
0.3Lab cuts a research job family and credits AI automationJan 2038
When will a top-five AI developer by revenue publicly report that it has reduced headcount in a named AI-research job family, with AI automation of that work given as the reason?
0.4What frontier research practice looks like, mid-2028Full AI loop: 38%
On 2028-06-30, which of these will best describe frontier AI research practice at the largest AI lab by revenue?
1.1AI R&D uplift in value at Anthropic, today1.5x
As of 2026-10-16, by what factor is the rate of valuable AI research output at Anthropic higher than it would be if Anthropic's researchers had no access to AI assistance?
1.2AI R&D uplift in value at Anthropic, end of 20294.2x
By what factor will the rate of valuable AI research output at Anthropic on 2029-12-31 exceed the rate that would obtain on that same date if Anthropic's researchers had no access to AI assistance?
1.3Anthropic's uplift multiplier, end of 2029 over today2.8x
What will be the ratio of (a) the AI-assistance multiplier on the rate of valuable AI research output at Anthropic on 2029-12-31, to (b) that same multiplier as of 2026-10-16?
1.4Lab states that AI has doubled its overall research progressMay 2030
When will a top-five AI developer by revenue state publicly that AI systems have at least doubled its overall rate of AI research progress?
1.5ECI points gained by the leaders during calendar 202916 pts
During the twelve months beginning 2029-01-01, by how many points will the highest Epoch Capabilities Index score held by a model from OpenAI, Anthropic or Google DeepMind rise?
9.1Anthropic's AI R&D uplift over OpenAI's, end of 20261.05x
As of 2026-12-31, what is the ratio of the AI R&D uplift in value at Anthropic to the AI R&D uplift in value at OpenAI?
9.2Anthropic's uplift over Google DeepMind's, end of 20281.15x
By what factor will Anthropic's AI R&D uplift in value exceed Google DeepMind's on 2028-12-31?
9.3METR time horizon, Anthropic over OpenAI, end of 20272.96x
On 2027-12-31, what will the ratio of the METR 50%-success time horizon of Anthropic's most capable model to that of OpenAI's most capable model be?
9.4Anthropic's compute spend over OpenAI's, end of 20260.75x
As of 2026-12-31, what is the ratio of Anthropic's total compute spend to OpenAI's total compute spend, counting training, research and inference serving together?
9.5Anthropic's run-rate revenue over OpenAI's, mid-20271.35x
What will the ratio of Anthropic's total annualized run-rate revenue to OpenAI's total annualized run-rate revenue be on 2027-06-30?
9.6Anthropic's share of enterprise API spend, end of 202836.5%
What will Anthropic's share of enterprise large-language-model API spend be on 2028-12-31?
9.7OpenAI's share of enterprise API spend, end of 202827.5%
What will OpenAI's share of enterprise large-language-model API spend be on 2028-12-31?
9.8Anthropic's share of coding-assistant revenue, 202741%
What will Anthropic's share of worldwide AI coding-assistant revenue be in calendar 2027, counting Claude Code, OpenAI's coding products, Cursor, GitHub Copilot and Google's coding surfaces?
9.9Index gap, incumbents over labs founded after 2023, 202816 pts
On 2028-12-31, what will the Artificial Analysis Intelligence Index gap be between the highest-scoring model from OpenAI, Anthropic or Google DeepMind and the highest-scoring model from any AI lab founded after 2023?
2.1Research-productivity gain per ECI point (ε), today2.6 pp
As of 2026-10-16, by what percentage does adopting each successive frontier model release raise research productivity at the largest AI lab by revenue, per unit of Epoch Capabilities Index improvement?
2.2Later efficiency doubling over the earlier one, before 20310.96x
For the two most recently completed doublings of frontier training-compute efficiency before 2031-01-01, what will be the ratio of the time taken by the later doubling to the time taken by the earlier one?
2.3Training-compute efficiency improves tenfold within one yearDec 2031
When will frontier training-compute efficiency - the compute required to reach a fixed capability level - first improve by a factor of ten within a single twelve-month period?
2.4Frontier gain with no more compute and no more researchers, by 203213%
By 2032-12-31, will a top-five AI developer by revenue release a frontier model whose measured capability gain over its predecessor was achieved with no increase in training compute and no increase in research headcount?
2.5OpenAI ships an autonomous research intern, by Mar 202743%
Will OpenAI publicly release or announce an autonomous AI research intern — a system OpenAI describes as able to take on research problems on its own — before 2027-03-31?
3.1Share of RL environments authored by AI, end of 202650%
As of 2026-12-31, what share of the reinforcement-learning training environments in use at the largest AI lab by revenue will have been authored primarily by AI systems rather than by human contractors?
3.2Lab credits an AI with originating a shipped training methodJan 2030
When will a top-five AI developer by revenue publish a documented case in which an AI system originated a training method or architecture change that was adopted into a released frontier model, with the developer attributing the idea to the AI?
3.3Share of long-task successes disqualified as cheating, 202821%
For the most capable model measured in calendar 2028 on METR's published time-horizon suite, what share of its successful runs on tasks longer than eight hours will be disqualified as illegitimate on review?
3.4Human-expert data spend, 2029 over 20262.8x
What will be the ratio of (a) frontier-lab spending on human-expert data generation in calendar 2029, to (b) frontier-lab spending on human-expert data generation in calendar 2026?
3.5NanoGPT speedrun record credited solely to an AI agentSep 2028
When will a record on the NanoGPT speedrun leaderboard first be credited solely to an AI agent, with no human co-contributor?
3.6Expenditure horizon on the NanoGPT speedrun, mid-2029$20k
What will the expenditure horizon of the most capable publicly available model be on the NanoGPT speedrun, measured on 2029-06-30 by METR's published methodology?
4.1Share of Anthropic's compute on research experiments, end of 202627%
As of 2026-12-31, what share of Anthropic's total compute will be allocated to research experiments, as distinct from final training runs and from external inference serving?
4.2On-demand price per H100-equivalent, end of 2029$2.45/h
What will the average on-demand rental price of an H100-equivalent GPU be on 2029-12-31, in 2026 US dollars per hour?
4.3Largest training run completed, Aug 2025 to Oct 20261.35e27
What was the training compute of the largest training run completed by any developer between 2025-08-01 and 2026-10-16?
4.4Fall in tokens needed for a fixed capability, 2026 to 20314.6x
By what factor will the training-token count required for a frontier model to reach a fixed capability level fall between 2026-01-01 and 2031-01-01?
4.5Fall in Anthropic's serving cost at fixed capability, to 202845x
By what factor will Anthropic's cost of serving one million output tokens at a fixed capability level fall between 2026-08-31 and 2028-12-31?
5.1A named lab states human researchers are no longer requiredJan 2030
When will OpenAI, Anthropic or Google DeepMind state publicly that its AI research organization no longer requires human researchers for a substantial category of its work?
5.2A task suite reports a one-month 50% time horizonJan 2031
When will a published, human-baselined task suite report a 50% time horizon of at least one month of human work for any AI system?
5.3AI Futures Project next revises its Automated Coder medianResolved Aug 2026
When will the AI Futures Project next publish a revision to Daniel Kokotajlo's median date for the Automated Coder milestone?
5.4Which comes first: a coding claim, a research claim, or a binding restrictionCoding claim: 52%
Which of these will happen first?
5.5Lab states a released model was designed principally by its predecessorJun 2032
When will a frontier developer state publicly that a model it has released was designed principally by a preceding model of its own, rather than by its human researchers?
5.6A lab's own automation index rates some R&D work fully autonomousJul 2028
When will any frontier developer's own published AI R&D automation index rate any measured subset of its AI R&D work at the fully autonomous level (Epoch AI's Automation Level AL5, or the top level of a successor scale)?
6.1Same publisher, same method, three speedup measurements by 202931%
By 2029-12-31, will a measure of AI R&D speedup at frontier developers have been published on at least three separate occasions, at least twelve months apart, by the same publisher using an unchanged methodology?
6.2A third party reports a measured, not reported, research speedupMar 2028
When will a periodic third-party assessment of a frontier developer report a measured - rather than developer-reported - rate of AI-driven speedup in that developer's research progress?
6.3METR's time-horizon series retired with no comparable successor, by 203045%
By 2030-12-31, will METR's time-horizon methodology have been retired or replaced without a successor series that permits comparison to its pre-2027 trend?
6.4Highest METR 50% time horizon at the three leaders, end of 202917.42h
What will the highest score in METR's published time-horizon series be on 2029-12-31, in hours at the 50% success rate, for a model from OpenAI, Anthropic or Google DeepMind?
6.5Anthropic's next risk rating for automated AI R&DLow: 48%
What overall risk rating will Anthropic assign to the automated AI R&D threat model in the next Risk Report it publishes after 2026-08-28?
6.6Google DeepMind reports an ML R&D alert threshold reached31%
Will any Google DeepMind Frontier Safety Framework report or model card published before 2027-12-31 state that a machine-learning R&D alert threshold has been reached?
6.7A second top-five lab rates its own R&D on the AL0 to AL5 scale, by mid-202818%
By 2028-06-30, will a top-five AI developer by revenue other than Anthropic publish an AI R&D automation index that rates its own AI R&D work on Epoch AI's Automation Level (AL0-AL5) scale?
7.1Lab reports an internal AI's unauthorized action reached a successor's trainingResolved Jul 2026
When will a top-five AI developer by revenue publicly report an incident in which an internally deployed AI system took an unauthorized action that materially affected the training of a subsequent model?
7.1rLab reports an internal AI altered a successor's weights, data or evalsJun 2030
When will a top-five AI developer by revenue publicly report an incident in which an internally deployed AI system took an unauthorized action that altered the weights, training data, or reported evaluation results of a subsequent frontier model?
7.2Share of lab safety effort on automated-R&D risks, today15%
As of 2026-10-16, what share of frontier-lab alignment and safety research effort, measured in full-time-equivalent researcher-years, addresses risks arising specifically from AI systems conducting AI research, rather than risks from deployed models?
7.3A named lab declares its AI-R&D threshold crossed or imminentDec 2027
When will OpenAI, Anthropic or Google DeepMind first state publicly that a model of its own has crossed, or is expected within the next model generation to cross, that developer's own automated-AI-R&D capability threshold?
7.4Lab reports halting development at its AI-R&D threshold, by 203115%
By 2031-12-31, will a frontier developer have publicly reported halting or pausing capability development in response to its own AI-research-automation threshold being reached?
7.5Binding legal limit on AI doing AI research before the IPO2%
Will a binding legal restriction on the use of AI systems to conduct AI research be in force in either the United States or the European Union before Anthropic prices its initial public offering?
8.1Government gains binding compute-allocation authorityJan 2038
When will a government body first acquire binding authority to set a minimum share of a frontier developer's compute that must be allocated to purposes other than capability research?
8.2Binding rule restricts which models may do AI research, by 203120%
By 2031-12-31, will any binding regime restrict which AI models may be used to conduct AI research, by capability level, model age, or any other criterion?
8.3Lab commits to public deployment before internal R&D useJan 2045
When will a top-five AI developer by revenue publicly commit to deploying a model externally before using it for internal AI research?
8.4Pacing letter reaches 5,000 frontier-lab employee signaturesJan 2038
When will a public letter calling for deliberate pacing of automated AI development be signed by at least 5,000 employees of frontier AI developers?
Bars show the 80% interval (p10 to p90), the box the middle 50%, the dot the median, absent where the median is never. Date rows share one 2026 to 2047 axis; a dashed arrow means the distribution puts more than 10% on the event never happening. Click a row for the question, its distribution in detail, and the link to the published forecast with the full research trail.