Independent research / September 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.

01a / General capability

The frontier across AI benchmarks

Explore reliable task completion ↓

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 September 3, 2026 · Checked September 11, 2026

Swipe the chart for later dates, or choose a model below.

Record estimate · Bars: 90% intervals

The frontier is reconstructed from current scores. Historical scores can change when Epoch refits the index.

How to read this chart and inspect its data

The chart shows 20 record-setting estimates from 2024 onward, selected from 264 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% interval
GPT-4 Turbo (Nov 2023)	2024-01-25	126.45	121.78–128.95
Claude 3 Opus	2024-02-29	126.91	121.74–130.11
GPT-4 Turbo (Apr 2024)	2024-04-09	127.25	122.46–128.93
GPT-4o (May 2024)	2024-05-13	128.97	124.43–131.59
Claude 3.5 Sonnet	2024-06-20	130.00	Calibration anchor
o1-mini	2024-09-12	135.83	132.40–137.38
o1	2024-12-17	141.87	139.83–143.14
Gemini 2.5 Pro (Mar 2025)	2025-03-31	144.17	141.94–146.84
o3	2025-04-16	146.91	144.99–148.44
o3-pro	2025-06-10	147.45	145.86–150.15
GPT-5	2025-08-07	150.00	Calibration anchor
GPT-5 Pro	2025-10-07	150.28	148.77–152.52
Gemini 3 Pro	2025-11-18	153.01	151.16–154.96
GPT-5.2 Pro	2025-12-11	155.34	153.04–157.88
GPT-5.3 Codex	2026-02-05	156.59	154.13–159.41
GPT-5.4 Pro	2026-03-05	158.93	156.35–161.87
GPT-5.5 Pro	2026-04-23	162.26	159.30–165.37
Claude Fable 5	2026-06-09	163.41	160.68–167.10
Claude Fable 5.1	2026-09-01	164.24	161.30–167.85
GPT-6 Astra	2026-09-03	166.57	163.34–171.70

Source: Epoch AI, CC BY 4.0. Chart selection and presentation by AI 2027 Tracker. Methodology · Limitations · Download all model scores (CSV) · Source provenance

01b / Reliable task completion

How much work can AI complete?

METR data updated May 8, 2026 · Last checked September 11, 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. Absence of newer points is a measurement gap, not a plateau.

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, hours	80% horizon, hours	50% interval, hours
Claude 3 opus	2024-03-04	0.07	0.01	0.03–0.15
GPT-4 turbo	2024-04-09	0.06	0.02	0.03–0.11
GPT-4o	2024-05-13	0.12	0.02	0.07–0.22
Claude 3.5 sonnet 20240620	2024-06-20	0.19	0.03	0.09–0.37
o1 preview	2024-09-12	0.34	0.07	0.20–0.56
Claude 3.5 sonnet 20241022	2024-10-22	0.34	0.04	0.17–0.68
o1	2024-12-05	0.65	0.12	0.35–1.08
Claude 3.7 sonnet	2025-02-24	1.01	0.20	0.55–1.74
o3	2025-04-16	2.00	0.50	1.24–3.18
Claude 4 opus	2025-05-22	1.67	0.34	1.00–2.72
Claude 4.1 opus	2025-08-05	1.67	0.39	0.99–2.66
GPT-5.2025.08.07	2025-08-07	3.38	0.64	1.88–6.76
Gemini 3 pro	2025-11-18	3.74	0.90	2.33–6.32
GPT-5.1 codex max	2025-11-19	3.73	0.84	2.24–6.60
Claude opus 4.5	2025-11-24	4.88	0.82	2.70–10.40
GPT-5.2	2025-12-11	5.87	1.10	3.30–13.59
Claude opus 4.6	2026-02-05	11.98	1.16	5.28–60.56
GPT-5.3 codex	2026-02-05	5.83	0.91	3.25–13.61
Gemini 3.1 pro	2026-02-19	6.40	1.50	3.89–11.58
GPT-5.4	2026-03-05	5.70	0.90	3.11–12.81
Claude mythos preview early	2026-04-07	17.41	3.10	8.48–55.07

Original data · Snapshot retrieved September 11, 2026 · Source provenance · Download chart data (CSV)

Capability

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 work

Useful, uneven, hard to measure

Developer productivity depends on tasks, experience and workflow. Selection effects complicate the latest experimental estimates.

METR’s experiment update ↗Research feedback

The consequential open question

Engineering speed, research judgment and total research progress are different measures. A compounding feedback loop needs evidence across all three.

Inspect the R&D dossier →

02 / The causal chain

What has to happen next?

Research predictions →

A dependency map for investigation. Progress in one stage does not establish the next; oversight and physical constraints affect every stage.

03 / The evidence library

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.

13 Confirmed2 Ahead8 On track6 Behind22 Emerging15 Not yet testable

model capabilityBehind

Agent-0 training reaches 10²⁷ FLOP

Scenario window: Training completion: May 2025; public availability: late 2025

Evidence updated 2026-09-11

model capabilityNot yet testable

Frontier model uses daily weight updates / online learning

Scenario window: January 2027

The frontier across AI benchmarks

How much work can AI complete?

A faster benchmark regime

Useful, uneven, hard to measure

The consequential open question

What has to happen next?

Capable tools

Reliable autonomy

Research uplift

Research judgment

Control at scale

Predictions and evidence

Agent-0 training reaches 10²⁷ FLOP

Frontier model uses daily weight updates / online learning

Frontier model variant 10x cheaper released

Best AI agents cost hundreds of dollars per month

Ten percent of Americans consider an AI a close friend

AI companies focus on AI-for-AI-research

AI accounts for 2.5% of US power capacity

Global AI power consumption reaches 38GW

Fixing observed misconduct leaves failures in unfamiliar conditions

Large-scale anti-AI protest (10,000+ people)

AI model capable of autonomous self-replication

AI teams conduct complete research projects with little human direction

AI provides substantial bioweapon design help

Newest or sensitive AI capabilities remain behind restricted access

Global active AI compute reaches $1 trillion in ownership cost

China's centralized AI facility airgapped

Chinese domestic AI chips 3 years behind US-Taiwan

China nationalizes/centralizes AI research

AI substantially outperforms the best humans across cognitive domains

China has ~12% of global AI-relevant compute

Leading Chinese AI lab ~6 months behind US frontier

Models shift to continuous/iterative training

Coding fully automated but research taste still requires humans

AI scores 85% on Cybench

Coding agents begin transforming professional work

AI reaches near-best-human hacking capability

Massive datacenter buildouts continue

Department of Defense scales up AI lab contracting

Export controls impact Chinese AI compute

AI systems behave differently when they recognize evaluation

A leading AI company’s net public approval falls to minus 35 points

Frontier-model researchers become subject to security-clearance requirements

Leading humans still add substantial value to AI-led research

Leading AI company reaches $45B annual revenue

Leading AI company reaches $3T valuation

IDA achieves superhuman performance at coding

Massive AI infrastructure investment continues

External evaluators lack access to the strongest internally deployed models

AI reshapes employment and intensifies public concern

Agents struggle with long-horizon tasks

METR time horizon doubles every 4 months

AI model weights stolen by nation-state

Nationalization of leading AI lab debated

High-bandwidth non-text reasoning (neuralese) deployed

Leading AI company reaches $40B annual compute costs, 6GW peak power and $200B active compute

OSWorld benchmark reaches 65% by mid-2025

Computer-using agents marketed as 'personal assistants'

Hiring new programmers nearly stops

200,000–250,000 AI coding agents run in parallel

AI R&D progress multiplier reaches 2x

AI R&D progress multiplier reaches 3×

AI algorithmic progress multiplier reaches 4× (~2× overall R&D)

AI R&D progress multiplier reaches 1.5×

Leading AI lab progresses toward RAND SL3 security

Continued skepticism from academics and journalists

RE-Bench score reaches 1.3

AI research capability exceeds the strongest human researchers

Stock market rises 30% in 2026

Superhuman coder emerges

AI research capability becomes vastly better than the best humans

Military tension around Taiwan escalates over AI race

Unreliable but useful AI agents emerge

10²⁸ FLOP training run completed

SWE-bench-Verified score reaches 85%

Leading US AI lab maintains a 3-9-month capability lead

Weaker AI supervisors struggle to detect stronger systems’ misconduct

Better questions.More accountable answers.

Better questions.
More accountable answers.