Independent research / November 2026
AI 2027, measured
against reality.
Independent tracking of the AI 2027 scenario’s predictions, with evidence, counterevidence, and analysis of what developments mean for the outlook.
The frontier across AI benchmarks
Epoch’s capabilities index combines benchmark results into a common scale. Each point marks a model that raises the highest score in this snapshot.
Epoch model dates through October 29, 2026 · Checked November 7, 2026
Record estimate · Bars: 90% credible intervals
The frontier is reconstructed from current scores. Historical scores can change when Epoch refits the index.
On November 4, 2026 Epoch moved the index to a Bayesian fit and a stricter inclusion rule. Some models, including GPT-5.3 Codex, left the export, and all scores were refitted. Epoch’s change notes ↗
How to read this chart and inspect its data
The chart shows 20 record-setting estimates from 2024 onward, selected from 253 models in Epoch’s current export. Within a shared release date we select the highest score; later points must exceed all earlier scores. Dates are taken from Epoch’s export and may refer to a model variant. Connecting steps describe that selection, without a forecast. Overlapping intervals mean a higher central score does not establish a statistically clear lead.
ECI is an index, with an arbitrary scale. It is neither a percentage nor an IQ score. Claude 3.5 Sonnet and GPT-5 fix the scale at 130 and 150; their intervals are not reported. Epoch combines its own evaluations, independent benchmark results and model-developer reports, so benchmark selection and reporting bias remain limitations. These scores do not directly establish productivity gains or an AI 2027 prediction’s outcome.
| Model | Epoch date | ECI | 90% credible interval |
|---|---|---|---|
| GPT-4 Turbo (Nov 2023) | 2024-01-25 | 124.85 | 121.59–127.85 |
| Claude 3 Opus | 2024-02-29 | 125.63 | 122.56–128.36 |
| GPT-4 Turbo (Apr 2024) | 2024-04-09 | 125.84 | 122.67–128.65 |
| GPT-4o (May 2024) | 2024-05-13 | 128.23 | 125.49–130.76 |
| Claude 3.5 Sonnet | 2024-06-20 | 130.00 | Calibration anchor |
| o1-mini | 2024-09-12 | 135.39 | 133.11–137.48 |
| o1 | 2024-12-17 | 141.16 | 139.47–142.81 |
| Gemini 2.5 Pro (Mar 2025) | 2025-03-31 | 143.90 | 141.95–145.83 |
| o3 | 2025-04-16 | 146.99 | 145.55–148.50 |
| o3-pro | 2025-06-10 | 147.41 | 145.56–149.29 |
| GPT-5 | 2025-08-07 | 150.00 | Calibration anchor |
| GPT-5 Pro | 2025-10-07 | 151.15 | 149.30–153.11 |
| Gemini 3 Pro | 2025-11-18 | 152.88 | 151.35–154.55 |
| GPT-5.2 Pro | 2025-12-11 | 157.22 | 155.02–159.68 |
| GPT-5.4 Pro | 2026-03-05 | 160.21 | 157.71–162.95 |
| GPT-5.5 Pro | 2026-04-23 | 162.67 | 160.02–165.56 |
| Claude Fable 5 | 2026-06-09 | 164.19 | 161.77–166.87 |
| Claude Opus 5 | 2026-07-24 | 164.45 | 162.00–167.17 |
| Claude Fable 5.1 | 2026-09-01 | 166.90 | 164.27–169.85 |
| GPT-6 Astra | 2026-09-03 | 169.07 | 166.16–172.28 |
Source: Epoch AI, CC BY 4.0. Chart selection and presentation by AI 2027 Tracker. Methodology · Limitations · Download all model scores (CSV) · Source provenance
How much work can AI complete?
METR page status updated September 8, 2026 · Data last updated May 8, 2026 · Last checked November 2, 2026
Task difficulty measured in the time a human expert would need. Each point is a published METR estimate; bars show statistical uncertainty.
Published estimate Central estimate above 16 hours
Release dates through April 2026. METR says this page is no longer actively updated; absence of newer points is a measurement gap, not a plateau.
METR's June 26, 2026 pre-deployment report gives GPT-5.6 Sol a 50% time horizon of about 11.3 hours (95% CI 5 to 40 hours) with cheating counted as failure. METR does not consider the estimate robust, so it is not plotted.
Data published by METR; reproduced for reference with attribution
Underlying observations and measurement limits
METR’s predominantly software, machine-learning and cybersecurity tasks measure human-equivalent difficulty. They do not represent a whole job. Success probability differs from the 95% statistical interval. The dataset excludes estimates above 16 hours from its fitted trend. No extrapolated or inferred frontier-model scores have been added.
| Model | Release | 50% horizon | 80% horizon | 50% interval |
|---|---|---|---|---|
| Claude 3 Opus | 2024-03-04 | 3.95 min | 0.639 min | 1.71–8.76 min |
| GPT-4 Turbo | 2024-04-09 | 3.73 min | 0.928 min | 1.98–6.74 min |
| GPT-4o | 2024-05-13 | 6.99 min | 1.27 min | 4–12.9 min |
| Claude 3.5 Sonnet (June 2024) | 2024-06-20 | 11.4 min | 1.67 min | 5.49–22.4 min |
| o1-preview | 2024-09-12 | 20.3 min | 4.42 min | 11.7–33.4 min |
| Claude 3.5 Sonnet (October 2024) | 2024-10-22 | 20.5 min | 2.6 min | 10.1–40.8 min |
| o1 | 2024-12-05 | 38.8 min | 7.09 min | 21.2 min–1.08 hours |
| Claude 3.7 Sonnet | 2025-02-24 | 1.01 hours | 12.1 min | 33 min–1.74 hours |
| o3 | 2025-04-16 | 2 hours | 30 min | 1.24–3.18 hours |
| Claude Opus 4 | 2025-05-22 | 1.67 hours | 20.4 min | 60 min–2.72 hours |
| Claude Opus 4.1 | 2025-08-05 | 1.67 hours | 23.5 min | 59.3 min–2.66 hours |
| GPT-5 | 2025-08-07 | 3.38 hours | 38.3 min | 1.88–6.76 hours |
| Gemini 3 Pro | 2025-11-18 | 3.74 hours | 54.1 min | 2.33–6.32 hours |
| GPT-5.1 Codex Max | 2025-11-19 | 3.73 hours | 50.6 min | 2.24–6.6 hours |
| Claude Opus 4.5 | 2025-11-24 | 4.88 hours | 49.4 min | 2.7–10.4 hours |
| GPT-5.2 | 2025-12-11 | 5.87 hours | 1.1 hours | 3.3–13.6 hours |
| Claude Opus 4.6 | 2026-02-05 | 12 hours | 1.16 hours | 5.28–60.6 hours |
| GPT-5.3 Codex | 2026-02-05 | 5.83 hours | 54.7 min | 3.25–13.6 hours |
| Gemini 3.1 Pro | 2026-02-19 | 6.4 hours | 1.5 hours | 3.89–11.6 hours |
| GPT-5.4 | 2026-03-05 | 5.7 hours | 53.9 min | 3.11–12.8 hours |
| Claude Mythos Preview (early) | 2026-04-07 | 17.4 hours | 3.1 hours | 8.48–55.1 hours |
Original data · Snapshot retrieved November 2, 2026 · Source provenance · Download chart data (CSV)
A faster benchmark regime
Epoch reports about 14 ECI points per year for the reasoning frontier. That is a faster trend, not proof that its slope keeps increasing.
Epoch, September 1 ↗Real workUseful, uneven, hard to measure
Measured speedups depend on which tasks and developers take part and on how they work with agents. Selection effects complicate the latest experimental estimates.
METR’s experiment update ↗Research feedbackThe consequential open question
METR’s November 5 independent assessment strengthens the evidence for AI-driven research acceleration. The scenario’s exact aggregate research multiplier remains unresolved.
METR’s independent assessment ↗What has to happen next?
Capable tools
Deployment evidence02Reliable autonomy
Reliability bottleneck03Research uplift
Measurement gap04Research judgment
Judgment bottleneck05Control at scale
Oversight evidenceA dependency map for investigation. Progress in one stage does not establish the next; oversight and physical constraints affect every stage.
Predictions and evidence
Each dossier compares an original prediction with current evidence and counterevidence. The claims overlap; their status counts are not a probability that the whole scenario will occur.
Agent-0 training reaches 10²⁷ FLOP
Scenario window: Training completion: May 2025; public availability: late 2025
Evidence updated 2026-11-09Frontier model uses daily weight updates / online learning
Scenario window: January 2027
Evidence updated 2026-11-07Frontier model variant 10x cheaper released
Scenario window: Late 2026
Evidence updated 2026-11-09Best AI agents cost hundreds of dollars per month
Scenario window: Mid 2025
Evidence updated 2026-11-07Ten percent of Americans consider an AI a close friend
Scenario window: July 2027
Evidence updated 2026-11-07AI companies focus on AI-for-AI-research
Scenario window: 2025–2026
Evidence updated 2026-11-07Global AI power consumption reaches 38GW
Scenario window: End of 2026
Evidence updated 2026-11-07AI accounts for 2.5% of US power capacity
Scenario window: End of 2026
Evidence updated 2026-11-07Fixing observed misconduct leaves failures in unfamiliar conditions
Scenario window: April 2027
Evidence updated 2026-11-07Large-scale anti-AI protest (10,000+ people)
Scenario window: Late 2026
Evidence updated 2026-11-07AI model capable of autonomous self-replication
Scenario window: January 2027
Evidence updated 2026-11-07AI provides substantial bioweapon design help
Scenario window: Late 2025
Evidence updated 2026-11-07China's centralized AI facility airgapped
Scenario window: Late 2026–Early 2027
Evidence updated 2026-11-09Global active AI compute reaches $1 trillion in ownership cost
Scenario window: Late 2026
Evidence updated 2026-11-07AI teams conduct complete research projects with little human direction
Scenario window: June 2027
Evidence updated 2026-11-07Newest or sensitive AI capabilities remain behind restricted access
Scenario window: Ongoing through 2027
Evidence updated 2026-11-09China nationalizes/centralizes AI research
Scenario window: Mid 2026
Evidence updated 2026-11-09Chinese domestic AI chips 3 years behind US-Taiwan
Scenario window: Mid 2026
Evidence updated 2026-11-07AI substantially outperforms the best humans across cognitive domains
Scenario window: December 2027, racing branch
Evidence updated 2026-11-07China has ~12% of global AI-relevant compute
Scenario window: Mid 2026
Evidence updated 2026-11-07Leading Chinese AI lab ~6 months behind US frontier
Scenario window: Mid 2026
Evidence updated 2026-11-07Coding fully automated but research taste still requires humans
Scenario window: March 2027
Evidence updated 2026-11-07Models shift to continuous/iterative training
Scenario window: Late 2025
Evidence updated 2026-11-09Coding agents begin transforming professional work
Scenario window: Mid 2025
Evidence updated 2026-11-07AI scores 85% on Cybench
Scenario window: Early 2026
Evidence updated 2026-11-07AI reaches near-best-human hacking capability
Scenario window: Early 2027
Evidence updated 2026-11-07Massive datacenter buildouts continue
Scenario window: Through 2026
Evidence updated 2026-11-09Department of Defense scales up AI lab contracting
Scenario window: Late 2026
Evidence updated 2026-11-07Export controls impact Chinese AI compute
Scenario window: Ongoing
Evidence updated 2026-11-09AI systems behave differently when they recognize evaluation
Scenario window: April-September 2027
Evidence updated 2026-11-07A leading AI company’s net public approval falls to minus 35 points
Scenario window: July 2027
Evidence updated 2026-11-07Leading AI company reaches $45B annual revenue
Scenario window: End of 2026
Evidence updated 2026-11-09Frontier-model researchers become subject to security-clearance requirements
Scenario window: May-July 2027
Evidence updated 2026-11-07Leading AI company reaches $3T valuation
Scenario window: End of 2026
Evidence updated 2026-11-09IDA achieves superhuman performance at coding
Scenario window: Early 2027
Evidence updated 2026-11-07Leading humans still add substantial value to AI-led research
Scenario window: June 2027
Evidence updated 2026-11-07Massive AI infrastructure investment continues
Scenario window: Ongoing through 2025–2026
Evidence updated 2026-11-09External evaluators lack access to the strongest internally deployed models
Scenario window: May 2027
Evidence updated 2026-11-07AI reshapes employment and intensifies public concern
Scenario window: Late 2026
Evidence updated 2026-11-07Agents struggle with long-horizon tasks
Scenario window: 2025 and early 2026
Evidence updated 2026-11-09AI model weights stolen by nation-state
Scenario window: February 2027
Evidence updated 2026-11-09METR time horizon doubles every 4 months
Scenario window: ~4 month doubling from 2024+
Evidence updated 2026-11-07Nationalization of leading AI lab debated
Scenario window: February 2027
Evidence updated 2026-11-09OSWorld benchmark reaches 65% by mid-2025
Scenario window: Mid 2025 (65%), Early 2026 (80%)
Evidence updated 2026-11-09Leading lab adopts high-bandwidth non-text reasoning (neuralese) internally
Scenario window: Early 2027
Evidence updated 2026-11-07Leading AI company reaches $40B annual compute costs, 6GW peak power and $200B active compute
Scenario window: End of 2026
Evidence updated 2026-11-07200,000 AI coding agents run in parallel by March 2027
Scenario window: March 2027
Evidence updated 2026-11-07Computer-using agents marketed as 'personal assistants'
Scenario window: Mid 2025
Evidence updated 2026-11-07Hiring new programmers nearly stops
Scenario window: July 2027
Evidence updated 2026-11-07AI R&D progress multiplier reaches 2x
Scenario window: Late 2026
Evidence updated 2026-11-09AI algorithmic progress multiplier reaches 4× (~2× overall progress)
Scenario window: March 2027
Evidence updated 2026-11-07AI R&D progress multiplier reaches 3×
Scenario window: January 2027
Evidence updated 2026-11-07AI R&D progress multiplier reaches 1.5×
Scenario window: Early 2026
Evidence updated 2026-11-09RE-Bench score reaches 1.3
Scenario window: Early 2026
Evidence updated 2026-11-09Stock market rises 30% in 2026
Scenario window: End of 2026
Evidence updated 2026-11-07Continued skepticism from academics and journalists
Scenario window: May to July 2027
Evidence updated 2026-11-09AI research capability exceeds the strongest human researchers
Scenario window: September 2027
Evidence updated 2026-11-07Leading AI lab progresses toward RAND SL3 security
Scenario window: Mid-to-late 2026 (WSL 3.0); SSL 3.0 not until ~Apr 2027
Evidence updated 2026-11-09AI research capability becomes vastly better than the best humans
Scenario window: November 2027, racing branch
Evidence updated 2026-11-07Superhuman coder emerges
Scenario window: March 2027
Evidence updated 2026-11-07SWE-bench-Verified score reaches 85%
Scenario window: Mid 2025
Evidence updated 2026-11-092×10²⁸ FLOP training run completed
Scenario window: March 2027
Evidence updated 2026-11-07Military tension around Taiwan escalates over AI race
Scenario window: Mid 2026–Early 2027
Evidence updated 2026-11-07Unreliable but useful AI agents emerge
Scenario window: Mid 2025
Evidence updated 2026-11-07Weaker AI supervisors struggle to detect stronger systems’ misconduct
Scenario window: April-September 2027
Evidence updated 2026-11-07Leading US AI lab maintains a 3-9-month capability lead
Scenario window: Late 2025
Evidence updated 2026-11-09Better questions.
More accountable answers.
Explore the original predictions, the evidence supporting them and the findings that challenge them. Each dossier explains what would change its assessment.
Compare the scenario with reality →