AI R&D progress multiplier reaches 1.5×
Tracker summary: AI systems contribute enough to AI research that the effective R&D multiplier reaches 1.5× — AI makes AI research 50% faster.
METR’s November 5 assessment strengthens evidence of AI-driven research acceleration in October. Its 1.9× historical capability-growth ratio does not establish the scenario’s 1.5× algorithmic/no-AI multiplier or its early-2026 timing.
At a glance
- Assessment: Emerging
- Confidence in assessment: 75%
- Outcome: unresolved
- Timing: Original window passed; outcome unresolved
- Evidence: proxy
- Predicted timing: Early 2026
- Primary source: AI 2027 archived scenario: Early 2026: Coding Automation, p. 6; Appendix B - The AI R&D progress multiplier: what do we mean by 50% faster algorithmic progress?, p. 45
What AI 2027 Predicted
Tracker summary: In early 2026, AI assistance makes algorithmic progress about 1.5 times faster, including time spent running experiments. Faster coding alone does not establish this all-inclusive research multiplier. AI 2027
Scenario source: archived version and page-number reference: Early 2026: Coding Automation, p. 6; Appendix B - The AI R&D progress multiplier: what do we mean by 50% faster algorithmic progress?, p. 45.
How We Track This
Compare useful research progress per unit of elapsed time with a credible no-AI baseline. Include experiment execution, failed directions and human review. Keep task speed, self-reported output and measured algorithmic progress distinct.
Current Evidence
November 5, 2026: independent assessment of October conditions. METR’s high-access assessment estimates Anthropic’s capability-index growth at 1.9× its pre-October-2025 rate (95% confidence interval 1.5 to 2.8×). METR considers AI responsible for a majority of the current acceleration likely, with substantial uncertainty. This is stronger evidence than lab self-report alone, but a historical capability-growth ratio is not the scenario’s algorithmic-progress multiplier against a no-AI counterfactual. It neither identifies an exact first crossing nor directly confirms this target.
METR’s separate November 5 review of July 15 conditions regards crossing Anthropic’s twofold early-warning threshold as plausible, not established; most relevant researchers leaned toward it not having been crossed then. That warning threshold, the historical trend ratio and the scenario multiplier are different quantities.
Emerging (75%) remains the assessment. The early-2026 window has passed; later independent evidence strengthens the mechanism without resolving the dated 1.5× algorithmic test.
Anthropic reports that a March poll of 130 research-team employees found median self-reported output gains of four times. The company cautions that actual uplift was likely lower. This is meaningful evidence of internal usefulness, without being a controlled estimate of algorithmic progress. Anthropic’s research account
METR’s May 11 report of a February-April survey of 349 technical workers, including 71 researchers, reported median perceived value gains of roughly 1.4-2 times, compared with three times for speed. The results overlap the scenario’s scale, but selection and self-report limit causal interpretation. METR productivity survey
Later milestone: tenfold algorithmic progress
The scenario’s June 2027 tenfold algorithmic-progress milestone extends this research family. It does not replace the early-2026 1.5× test. No qualifying tenfold result has been verified. The operational test below is the tracker’s specification, not a verbatim scenario threshold definition.
Extend the existing R&D family with a 10× milestone. Measure accepted, reproducible algorithmic progress per unit calendar time over at least three comparable complete projects against a documented no-AI counterfactual. Hold experimental compute and quality criteria comparable, include failed experiments and human interventions, and require independently reviewed methodology. Report the multiplier and uncertainty, separating algorithmic progress from total capability progress.
An estimate at or above 10× is a milestone observation only when uncertainty and the counterfactual are credible. Coding throughput, token output, claimed developer hours saved or lab self-report alone are partial evidence. This is one family extension, not an independent counted dossier.
OpenAI’s September 6 research account reports 3.1 agent-workdays per human workday by mid-August and record experiments per active experimenter. It notes increased compute and uncertain translation from activity to progress. This strengthens the existing mechanism evidence without identifying the all-inclusive multiplier or resolving its early-2026 timing.
Counterevidence & Limitations
Epoch’s October 7 InnovationEval found limited gains relative to a reference human method on one post-training research task despite substantial experimental compute. This constrains extrapolation from engineering throughput to novel output. It measures neither all-inclusive laboratory progress nor the early-2026 multiplier.
Google’s October safety report finds no statistically significant acceleration in its Gemini capability-index trend after an April 1 breakpoint, including when internal frontier models are included. Its internal index combines benchmark results rather than accepted algorithmic progress and does not identify the no-AI counterfactual. This first-party result limits extrapolation from rising agent activity to a measured research multiplier; it does not establish zero AI uplift or resolve the early-2026 timing.
Open-ended shadow evaluations show that successful implementation can fail to produce accepted research results. Conversely, automated alignment research demonstrates useful method discovery where objectives are measurable. Neither study yields a whole-lab multiplier. Shadow evaluations · Automated alignment research
The evidence makes material acceleration plausible. It does not identify the exact contribution to algorithmic progress or retrospectively resolve the early-2026 deadline. Ratios of benchmark scores and code volume are not substitutes for that missing quantity.
What Would Change Our Assessment
- Strengthen: Multi-team studies measure accepted research progress, including experiment latency and unsuccessful work, against a defensible baseline.
- Confirm: Comparable early-2026 evidence establishes approximately 1.5× all-inclusive progress.
- Weaken: Measured gains disappear after quality, experiment and coordination costs are included.
The later research-multiplier milestones remain separate timing tests; evidence for this initial uplift cannot automatically resolve them.
Update History
| Date | Update |
|---|---|
| 2026-11-09 | Assessment reviewed. Status: Emerging; confidence: 75%. METR’s November 5 independent October assessment strengthens the acceleration evidence. Its 1.9× historical capability-growth estimate is not the scenario’s algorithmic/no-AI multiplier; the early-2026 1.5× test remains unresolved. The July warning-threshold retrospective is uncertain, not confirmation. |
| 2026-11-07 | Assessment reviewed. Status: Emerging; confidence: 75%. The early-2026 window has passed and the dated 1.5× all-inclusive algorithmic-progress result remains unresolved. Under the methodology’s D3 post-deadline exception, Emerging is retained because Anthropic’s internal research account, METR’s technical-worker survey and constrained automated alignment research provide positive partial evidence of research assistance. Code volume, self-reported output and narrow method discovery do not identify the registered no-AI counterfactual multiplier. Later evidence supports the mechanism, not retrospective confirmation; the original date and operational test remain unchanged. The scenario statement is labeled as a summary. The later tenfold milestone remains separate from the early-2026 1.5× test; activity and perceived output cannot establish that dated all-inclusive multiplier. |
| 2026-10-19 | Google reported no statistically significant acceleration in its internal capability-index trend. This qualifies throughput-based extrapolations but does not identify the early-2026 algorithmic-progress multiplier. Status and confidence unchanged. |
| 2026-10-12 | Epoch’s single-target method test found limited gains relative to the reference method. This bounds one research-output claim; it does not estimate a whole-lab multiplier or determine the early-2026 timing. Status and confidence unchanged. |
| 2026-09-07 | Added OpenAI measurements of agent use and experiment activity. They strengthen research-assistance evidence; the overall multiplier and original timing remain unresolved. Status and confidence unchanged. |
| 2026-09-07 | Added direct lab and cross-organization productivity evidence supporting material acceleration. The all-inclusive algorithmic-progress multiplier and its early-2026 timing remain unresolved. Status and confidence unchanged. |
| 2026-09-06 | Assessment revised from on-track (0.80) to emerging (0.75). Meaningful AI research assistance is documented. A comparable 1.5-fold whole-research multiplier was not established within the original early-2026 window. |
| 2026-08-17 | Anthropic said internal AI significantly accelerates research and engineering, while its concrete AI R&D evaluations have saturated and no controlled lab-wide multiplier is available. This supports the prediction’s direction without establishing a 1.5x result. Status and confidence remain unchanged. |
| 2026-08-03 | Independent shadow evaluations found frontier agents could complete research engineering but failed to make substantial progress on two open-ended AI research questions. This cautions against inferring a 1.5x R&D multiplier from coding throughput alone. Status and confidence remain unchanged because the sample is limited. |
| 2026-06-22 | Anthropic reported internal evidence that AI tools are accelerating engineering and research work at the lab, including more than 80% Claude-authored merged code and a median 4x self-reported output estimate among research-team employees using Mythos Preview. This supports the current on-track assessment for a 1.5x R&D multiplier. Confidence adjusted 0.75 -> 0.80. |
| 2026-05-25 | METR’s Frontier Risk Report reported broad autonomous-agent use inside frontier labs and cited measured productivity benefits of roughly 4-20% in its latest RCT of late-2025 public agents, plus larger self-reported survey gains of 1.6x-4x depending on design. This modestly strengthens the case for R&D acceleration but still falls short of a clean public demonstration of a 1.5x frontier-lab R&D multiplier. Confidence adjusted 0.70 → 0.75. |
| 2026-04-13 | No confirmed evidence of 1.5× R&D multiplier at frontier labs. METR simpler timelines model (Feb 2026) discusses measuring “uplift fraction” and acknowledges difficulty of measuring real-world R&D productivity gains (METR). Academic literature (GovAI, March 2026, arXiv:2603.03992) notes “it is unclear how directly such results translate to productivity boosts given real-world integration frictions.” The most advanced model had 80% success rate on tasks taking human expert coders 1h10m, but benchmark-to-productivity translation remains unproven. No lab has publicly claimed 1.5× R&D speedup from AI tools. Prediction remains behind. 2026-11-07 · Correction: the “remains behind” wording was incorrect; the assessment was On Track (70%), with no status change in this entry. |
| 2026-04-02 | AI Futures Project Q1 update: AI company researchers “doubling down” on near-term automated AI R&D timelines in private discussions. Kokotajlo: “Rather than walking back their predictions, they are doubling down, both in public and in private.” The authors consider this some evidence (though far from conclusive) that the multiplier may be closing faster than external measurements suggest. Source: LessWrong |
| 2026-03 | AI R&D automation infrastructure in place at all major labs. Current estimates suggest ~1.1-1.3× multiplier, approaching but not yet reaching the 1.5× target. |
| 2026-02 | METR experiment redesign update (Feb 24): “developers are more sped up from AI tools now — in early 2026 — compared to our estimates from early 2025.” Directionally positive but not yet quantified. |
| 2026-01 | AI Futures clarification post (Jan 27): “AI software R&D uplift: behind pace.” METR RCT and gap between benchmarks and real-world productivity remain primary evidence. Anthropic 2x internal coding uplift (self-reported, non-randomized) is best evidence in favor. |
| 2025-12 | AI Futures Project Dec 2025 model update explicitly cited R&D multiplier as key reason timelines shifted longer: “less pre-SC AI R&D automation effect, due to fixing a bug and revising the interpolation method.” AI labs report meaningful but modest acceleration of internal R&D workflows through AI tools. Estimated multiplier approaching but not yet at 1.5×. |
| 2025-08 | METR’s August 12 research update confirms and elaborates its July finding: a controlled study found experienced developers working with early-2025 AI tools were 19% slower, not faster. METR attributes this to systematic overestimation of AI utility on tasks requiring human judgment. This is direct contrary evidence to the 1.5x multiplier prediction. The prediction’s mechanism (benchmark performance translating to R&D productivity) is not confirmed. |
| 2025-07 | METR publishes RCT (July 10): experienced open-source developers using Cursor Pro with Claude 3.5/3.7 Sonnet took 19% longer to complete tasks than control group. 16 developers, 246 real GitHub issues. Developers perceived 20% speedup while experiencing slowdown. Significant counterevidence against near-term 1.5x R&D multiplier. |