AI R&D progress multiplier reaches 2x

Last updated
While the latest Agent-1 could double the pace of OpenBrain’s algorithmic progress, Agent-2 can now triple it, and will improve further with time.

January 2027: Agent-2 Never Finishes Learning · Archived PDF, p. 9

METR’s November 5 assessment adds independent acceleration evidence, not an exact twofold algorithmic/no-AI multiplier. Emerging 50%; the late-2026 target remains open and the earlier 1.5× stage is separate.

At a glance

  • Assessment: Emerging
  • Confidence in assessment: 50%
  • Predicted timing: Late 2026
  • Primary source: AI 2027 archived scenario: January 2027: Agent-2 Never Finishes Learning, p. 9

What AI 2027 Predicted

Tracker summary: The scenario depicts a progression of AI R&D multipliers: 1.5x in early 2026 (AI assists coding), 2x by late 2026 (Agent-1 “could double the pace”), 3x in January 2027 (Agent-2 triples it), and 4x by March 2027 (superhuman coding labor force). The 2x milestone marks a critical transition: the point where AI doesn’t just help with implementation but meaningfully accelerates the overall research cycle. In Appendix B, the authors clarify that the multiplier measures algorithmic progress specifically: “OpenBrain makes as much AI research progress in 1 week with AI as they would in 1.5 weeks without AI usage” (for 1.5x). At 2x, one week of AI-assisted research produces two weeks’ worth of algorithmic progress. Archived PDF, p. 9. Archived PDF, p. 44.

The scenario states this in the January 2027 section about the latest Agent-1; the late-2026 date is the tracker’s inference. See archived PDF, p. 9 for the wording quoted above.

The leap from 1.5x to 2x represents the transition from AI as a useful coding assistant to AI as a genuine research accelerator.

Scenario source: archived version and page-number reference: January 2027: Agent-2 Never Finishes Learning, p. 9.

How We Track This

  • Lab statements about AI accelerating their own research (Anthropic, OpenAI, Google DeepMind)
  • Published estimates of AI’s contribution to AI R&D productivity
  • METR’s research on forecasting impacts of AI R&D acceleration
  • Coding productivity studies (GitHub Copilot, internal tool assessments) as a lower bound
  • Epoch AI analyses of algorithmic efficiency trends
  • Stanford HAI AI Index annual report

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 (50%) remains the assessment. The late-2026 target remains open; the earlier 1.5× stage has its own elapsed-window assessment.

Direct evidence for a 2x multiplier is limited, but directional signals are strong:

Lab statements and industry signals:

Quantitative estimates:

  • Coding-productivity studies concern particular tasks and populations, not the overall research-progress multiplier.

  • METR’s original experienced-developer trial found a slowdown; later work has different selection and measurement limitations. Neither directly estimates frontier-lab research progress. Experiment update, February 24, 2026.

  • Anthropic Institute’s internal analysis (retrieved November 7, 2026), covering Q2 2026 code activity and a March 2026 staff poll, reported that the typical Anthropic engineer merged 8x as much code per day in Q2 2026 as in 2024, while cautioning that lines of code overstates true productivity. It also reported a March 2026 poll of 130 research-team employees where the median respondent estimated about 4x output with Mythos Preview versus no AI models on comparable projects.

  • Tracker assessment of the sources cited here: No qualifying external measurement of the twofold algorithmic-progress multiplier is established by these studies.

  • Earlier first-party assessment, qualified by METR’s later review above: Anthropic’s August 2026 Risk Report said internal AI made research and engineering significantly faster, but not yet by a factor of 2. Anthropic says the estimate is uncertain and measurement is difficult. Its RSP doubling criterion concerns the overall rate of progress relative to the pre-AI-assisted rate (August Risk Report, p. 94); under the scenario’s footnote 31, roughly 2x overall progress corresponds to about 4x algorithmic progress, so it is not the same quantity as this dossier’s 2x algorithmic target (March 2027: Algorithmic Breakthroughs, archived PDF, p. 12). This is first-party context, not a controlled estimate of lab-wide algorithmic progress or a direct below-threshold result for this dossier.

Indirect evidence from capability growth:

  • Tracker interpretation: Increasing measured task capability is precursor evidence, not a direct laboratory research multiplier.
  • Tracker interpretation of the Anthropic sources above: Anthropic’s internal examples are unusually close to the prediction mechanism because they cover production engineering, research-team output, and experiment execution inside a frontier lab. These Anthropic examples are still partly self-reported and come from one lab; METR’s later independent assessment above strengthens, but does not complete, the measurement.

Assessment: Evidence supports meaningful AI assistance but not a precise current multiplier or confirmation of the twofold target. Late 2026 remains the timing reference.

Source:

OpenAI’s September 6 research account adds observed agent use and increased experiment activity within a frontier lab. These input and activity measures complement the existing Anthropic evidence. They do not identify a twofold increase in accepted algorithmic progress against a comparable baseline.

Counterevidence & Limitations

  • The R&D multiplier concept is inherently difficult to measure: there is no standard methodology and labs have incentives both to overstate (for hype) and understate (for competitive secrecy) AI’s contribution
  • Coding throughput is one component of research; experiment selection, execution resources, interpretation and integration can limit overall progress.
  • Algorithmic progress itself is hard to measure: Epoch AI’s work on algorithmic efficiency improvements provides a baseline but doesn’t cleanly separate AI-assisted from human-only contributions
  • The distinction between “AI helps us write code faster” (well-documented) and “AI accelerates our rate of scientific discovery” (much harder to establish) is critical and often conflated
  • Most productivity studies measure individual task completion, not the compound effect on an entire research pipeline
  • Anthropic’s 8x code-throughput metric is not the same as 8x productivity. The company explicitly warned that lines of code measures quantity rather than quality and likely overstates the true speedup.

What Would Change Our Assessment

  • Strengthen the assessment: Comparable whole-project measures connect AI use to accepted research progress, including quality, failures and human input.
  • Confirm: Evidence establishes at least twofold progress against a documented no-AI counterfactual, with uncertainty.
  • Reassess timing after late 2026: Record whether the target was met within the original window; unsupported consensus ranges cannot resolve it.

Update History

DateUpdate
2026-11-09Assessment reviewed. Status: Emerging; confidence: 50%. METR’s November 5 assessment adds independent evidence beyond the earlier lab report. Historical capability growth, the lab’s warning threshold and the scenario’s twofold algorithmic multiplier remain distinct. July crossing is plausible, not established; late 2026 remains open. Corrected the August risk-report link and page locator.
2026-11-07Assessment reviewed. Status: Emerging; confidence: 50%. Anthropic’s August risk report estimates substantial internal acceleration below its own 2x progress criterion, while its August essay reports task-level gains. Neither directly measures the scenario’s algorithmic-progress multiplier, supporting Emerging rather than confirmation of that 2x target. Correction: the exact scenario says Agent-1 “could double” progress; late 2026 is the tracker’s timing inference. Anthropic’s overall-progress criterion is not the scenario’s algorithmic multiplier; corrected historical horizon and fit-window figures remain at their entries.
2026-09-07Added OpenAI process measurements as leading indicators alongside existing lab evidence. The twofold research-progress threshold remains unresolved. Status and confidence unchanged.
2026-09-06Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations.
2026-08-17Anthropic said internal AI significantly accelerates R&D but has not yet produced a 2x overall multiplier. This is direct first-party evidence below the late-2026 threshold. Status and confidence remain unchanged because Anthropic also says the estimate is uncertain and difficult to measure.
2026-06-22Anthropic reported internal evidence of large AI-assisted productivity gains in research and engineering, including 8x code throughput and a median 4x self-reported output estimate among research-team staff using Mythos Preview. This strengthens the emerging case for a 2x R&D multiplier, with caveats around self-reporting and code-volume metrics. Confidence adjusted 0.45 -> 0.50.
2026-03OpenAI GPT-5.4 Codex surpasses 2M weekly active users (5x growth since January). AI-2027 report’s 1.5x estimate for early 2026 broadly seen as plausible. The 2x prediction for late 2026 remains within its timeline window but unconfirmed. Status: Emerging. Confidence 0.45.
2026-02METR walks back July 2025 “19% slower” finding — developers now refuse to work without AI, biasing the original experiment. METR states they “believe it is likely that developers are more sped up from AI tools now.” Claude Opus 4.6 time horizon reaches 14.5 hours. 2026-11-07 · Correction: METR’s TH1.1 data gives about 11h59m at 50% for Opus 4.6 (raw results, retrieved November 7, 2026). This is human-task difficulty at a specified success rate, not continuous runtime.
2026-01METR Time Horizon 1.1 (Jan 29): post-2023 doubling time estimated at 130.8 days (4.3 months). METR’s simpler AI timelines model estimates AI R&D automation fraction at 0.25-0.5, implying 1.33x-2.0x uplift. First quantitative third-party estimate placing 2x within the plausible range. 2026-11-07 · Correction: The current TH1.1 raw results, retrieved November 7, 2026, give 128.744 days (about 129 days) for the 2023-onward fit, matching the longer-window figure in the METR doubling dossier. That fit is distinct from its 2024-onward 105-day fit; neither number is a directly measured lab R&D multiplier.
2025-11Claude Code reaches $1B ARR in 6 months — fastest software product to this milestone, validating massive developer adoption of AI coding tools. SWE-bench scores cross 80%. Epoch AI: algorithmic efficiency improving ~3x/year, inference costs falling 5-10x/year.
2025-10Sam Altman announces goal of “automated AI research intern by September 2026” on hundreds of thousands of GPUs. Most explicit public timeline from a leading lab for automating AI R&D. Fortune reports Anthropic engineers use Claude for ~60% of tasks but “fully delegating” remains limited.
2025-09McKinsey R&D Leaders Forum: AI could enhance R&D throughput 75-150% in some industries. Claude Sonnet 4.5 achieves 30+ hours autonomous operation (vs. 7 hours for Opus 4), a prerequisite for extended research tasks.
2025-08Anthropic publishes internal study: 67% increase in merged PRs per engineer per day after Claude Code adoption. Task complexity increased from 3.2 to 3.8. Engineers can only “fully delegate” 0-20% of work. Data suggests multipliers of 1.3-1.7x for coding-adjacent work, below 2x for overall research. METR publishes “Forecasting the Impacts of AI R&D Acceleration” pilot study (Aug 20).
2025-07METR publishes controlled developer productivity study (July 10): experienced developers were 19% slower with AI tools. Injects caution into multiplier estimates — if experts aren’t faster on familiar codebases, path to 2x is harder than lab rhetoric suggests.
2025-06Sam Altman publishes “The Gentle Singularity” (June 10), stating AI will enable “systems that can figure out novel insights” by 2026. Scientists reportedly “two or three times more productive” with AI (self-reported, uncontrolled). Leading labs frame AI-accelerated research as core strategic priority.
2025-05AI-for-AI-research era begins. Anthropic releases Claude Code (May 22); OpenAI launches Codex agent (May 16). Google DeepMind unveils AlphaEvolve (May 14) — a Gemini-powered agent that discovers novel algorithms, including a 4x4 matrix multiplication method beating a 56-year record and a scheduling heuristic recovering 0.7% of Google’s global compute. Concrete example of the R&D feedback loop, but overall multiplier remains well below 2x.