- 6 mo46%Gemini 3.1 Pro (high)+8.4 pts
- 1 yr32%GPT-5 Pro+23 pts
- 2 yrtest didn’t exist yet
- 3 yrtest didn’t exist yet
Latest release
Claude Fable 5.2
What AI can actually do, measured and sourced. Updated daily.
AI capabilities, costs and footprint, tracked over time and checked for new results every day. Every number links to its source: papers, labs, independent testers like METR and Epoch AI, and the posts where results first appeared.
Chart: 33 models by release date and Epoch Capabilities Index score. The best score rose from 129 (GPT-4o, May 2024) to 174 (Claude Fable 5.2, Nov 2026).
OpenAIAnthropicGoogleOther labsEach dot is a model on its release date. The line is the best score so far. Source: Epoch AI.
Latest changes
· 2 new models · 2 leaderboards with a new leader · 3 new entries · 5 updated entries · 3 charts with new data · 6 new charts
- New
Anthropic released it on 9 November 2026. Independent tests put it level with rivals on hard maths and behind them on PhD-level science.
- New
OpenAI says an unreleased model built working attacks for 948 Linux flaws in one run, about $1,279 each, and proposed a fix for each.
- Model
Anthropic · Public. Same model as Mythos 5.2, with cyber and biology safeguards. $8 in / $40 out per million tokens, less than Fable 5.1. Leads Epoch's capability index (174.2) and ARC-AGI-3 (81.2%).
- Model
Anthropic · Limited preview. Same model as Fable 5.2 with fewer cyber and biology safeguards. Only for vetted cyber-defence and life-science organisations.
- New leader
Epoch Capabilities Index: 174.2 points (before: GPT-6 Astra) · ARC-AGI-3: 81.2% (before: GPT-6 Astra (Max))
- New
Anthropic says Claude helped redesign a cancer therapy and drew up 35,000 antibody designs in a month. Almost none is tested in people yet.
- Updated
Added what the agents did to the German wiki itself (attempted exploits, overwritten home page, the moderator's clean-up) and retitled the entry.
- Updated
Now uses felonybench.org's counts (13 and 10 probable felonies) and explains how it counts; felonybench.com's incident tally (24) is given alongside.
Show 12 more
- Updated
Reframed around METR's November audit: on AI speeding up AI research, reality has caught up with the scenario's March 2027. Added Daniel Kokotajlo's and Ryan Greenblatt's reactions.
- Updated
Claude Fable 5.2 raised the best ARC-AGI-3 standard-harness score from 62.7% to 81.2%.
- Updated
Added Claude Fable 5.2, the new best score with the standard harness (81.2%).
- New data
Noted that METR's November audit puts AI speeding up AI research at about the scenario's March 2027 level, ahead of the August grade.
- New data
Added Claude Fable 5.2: 81.2% with the standard harness, up from GPT-6 Astra's 62.7%.
- New data
Added Claude Fable 5.2: 95.1% on Tiers 1–3 (a new high, within the margin of error of the previous 93.7%) and 100% on Tier 4.
- New chart
Artificial Analysis combines about ten tests of reasoning, coding, knowledge work and agent tasks into one score out of 100. It is one of the most quoted single numbers for how capable a model is.
- New chart
What the most capable model cost on the day it came out. A million tokens is roughly 750,000 words of input and output.
- New chart
What it costs to use an AI at least as capable as Claude Opus 4.6, judged by Epoch AI's index of more than 50 tests. A million tokens is roughly 750,000 words of input and output.
- New chart
What it costs to use an AI at least as capable as Claude Opus 4.8, judged by Epoch AI's index of more than 50 tests. A million tokens is roughly 750,000 words of input and output.
- New chart
What it costs to use an AI at least as capable as GPT-5.6 Sol, judged by Epoch AI's index of more than 50 tests. A million tokens is roughly 750,000 words of input and output.
- New chart
What it costs to get answers as good as o3's (April 2025) on hard science questions. Unlike price per token, this counts every token a model actually uses, including the hidden 'thinking' tokens of reasoning models.
The short version
The best AI keeps getting smarter. Any given level of ability keeps getting cheaper. Tap a number for its chart and sources.
More capable261× longer4 min → 17 hThe longest software task an AI agent can finish on its own, half the time.GPT-4 (Mar 2023) → Claude Mythos Preview (limited access) (Apr 2026), when METR stopped measuring: the newest models finish nearly all of its tasks.Smarter+54 pts42% → 96%Score on PhD-level science questions. PhD experts in the field score about 70%.GPT-4 Turbo → Gemini 4 Argon (high).Cheaper for the same ability316× cheaperWhat it costs to use an AI as capable as the original GPT-4, per million tokens.$15 (Nov 2023) → $0.048 (Nov 2025, 1 yr ago).Improving faster×2 every 4 monthsHow often the length of tasks AI can do doubles.Since 2023; METR first reported about 7 months for 2019–2024.More used1.2 billionweekly ChatGPT users, as OpenAI counts them (Sept 2026).
The best AI gets smarter122 → 174Capability score of the best model.GPT-4 (Mar 2023) → Claude Fable 5.2 (Nov 2026), on an index of more than 50 tests.The best AI stays expensive$37.50 → $16Price of the best model, per million tokens.GPT-4 (Mar 2023) → Claude Fable 5.2 (Nov 2026). In between it ranged from $3.44 to $67.50, with no steady fall.Yesterday’s best gets cheap50× cheaperWhat matching Claude Opus 4.8 costs, 4 months after its launch.$10 → $0.20 per million tokens (Claude Haiku 5.5). On average a fixed level gets about 13× cheaper a year.
The best AI keeps getting smarter. Each level of ability keeps getting cheaper.
Top: how capable the best model is. Bottom: what one fixed level of capability costs, not the best model. Dashed lines show where each trend leads by the end of 2031 if it slows, holds its pace or speeds up.
AI capability (Epoch Capabilities Index)
One score built from more than 50 tests, so models stay comparable even as single tests max out. GPT-4 (2023) scored 122, GPT-5 scored 150 and Claude Fable 5.2 scores 174. At 134 a model gets about halfway on PhD-level science questions; at 151, halfway on FrontierMath's hard maths.
100150200250300350202420252026202720282029203020312032Today
Price of one fixed level of capability (o3's, April 2025)
Not the price of the best model: the cost of matching one fixed level. Here, answering a million PhD-level science questions (GPQA Diamond) at o3's score (82%), using the cheapest model that reaches it and counting reasoning tokens. Lighter line: GPT-4's level (36%), from 2023.
$0.000001$0.00001$0.0001$0.001$0.01$0.10$1$10$100$1,000$10,000$100,000$1,000,000202420252026202720282029203020312032Today
- Measured
- Extrapolation (slower, same pace, faster)
- Range between slower and faster
Even the slower lines add up by 2031
If capability grows at its pace from before reasoning models and prices fall at the slow end of Epoch AI’s range, by the end of 2031:
Capability index174 → about 215 points+41 points, about as much as the jump from GPT-4 (Mar 2023) to GPT-5.5 Pro (Apr 2026).
Price at GPT-6 Astra’s level$16 → about $0.0056Per million tokens, about 2,900× cheaper than in Nov 2026, when Claude Fable 5.2 is the cheapest model at that level. Assumes a fall of 4.7× a year. Other models →
Index values above about 180 are harder than every test the index contains, so a score of 215 has no concrete meaning yet; read it as “well beyond today”.
Dashed lines extend past rates. They are not forecasts or probability bounds: trends can speed up, slow down or stop.
No line lets AI speed up its own progress, sometimes called recursive self-improvement. The faster line adds the 1.9× speed-up METR measured at Anthropic once, as a fixed boost. If that speed-up keeps growing as AI gets better, progress could outrun all three lines. How METR measured the speed-up →
Data: Epoch AI (capability index; cost per question incl. reasoning tokens).How the extrapolations work →
Progress over time
The best result available today, compared with 6 months and 1, 2 and 3 years ago.
- 6 mo0.5%Best frontier AI at launch+81 pts
- 1 yrtest didn’t exist yet
- 2 yrtest didn’t exist yet
- 3 yrtest didn’t exist yet
- 6 mosame record as now
- 1 yrsame record as now
- 2 yr$0.13Gemini 1.5 Flash2.8× cheaper
- 3 yr$15GPT-4 Turbo316× cheaper
- 6 mo0%GPT-5.5+4.4 pts
- 1 yrtest didn’t exist yet
- 2 yrtest didn’t exist yet
- 3 yrtest didn’t exist yet
- 6 mo$0.17Gemini 3 Flash Preview (high)15× cheaper
- 1 yr$8.42Grok 4 refinement system714× cheaper
- 2 yrno result yet
- 3 yrno result yet
- 6 mo95%GPT-5.4 Pro+1.8 pts
- 1 yr87%Grok 4+9.4 pts
- 2 yr62%o1-mini (high)+34 pts
- 3 yr42%GPT-4 Turbo+54 pts
Browse by topic
Mathematics
Competition math, open research problems and formal proofs.
21 entries1 new · 1 updated
Computer science
Software engineering, competitive programming and new algorithms.
18 entries1 new · 1 updated
Biology & Medicine
Protein structure, drug discovery, lab science and biosecurity evaluations.
16 entries1 new
Benchmarks
Standardised tests of AI ability, expert baselines, and how fast tests get saturated.
35 entries2 new · 6 updated
Agents & METR
How long a task AI can complete on its own, measured in human working time.
27 entries1 new · 7 updated
Price per IQ
What a fixed level of capability costs, and how fast that price falls.
8 entries1 updated
Energy & water
Electricity, water and emissions per prompt and in aggregate, compared with everything else.
7 entries
Myths
The most repeated claims about AI, stated fairly and checked against the data: water and power, “it’s dumb”, “it’s autocomplete”, “it can’t discover anything”, and hype in the other direction.
13 entries2 updated
Adoption
Who uses AI, how much, and for what.
16 entries1 new
Robotics
Two questions: can robots do physical work on their own yet, and how fast can we build them and at what cost?
10 entries
Fun & strange
Meme benchmarks, a leaderboard of real incidents with AI agents, and strange experiments, from vending machines to Pokémon.
13 entries2 updated
Latest
Newly added and recently updated.
Big news
Claude Fable 5.2 ranks first overall in independent tests (174 vs 169), but not on science
Anthropic released it on 9 November 2026. Independent tests put it level with rivals on hard maths and behind them on PhD-level science.
174.2Epoch AI's combined test score, a new high (GPT-6 Astra: 169.2)
Milestone7 sources
AI is now speeding up AI research about as much as “AI 2027” expected by March 2027
METR found AI roughly doubling Anthropic's pace, a level the scenario reaches in March 2027. Its lead author says the team is ‘spooked’.
~2×AI-driven speed-up of Anthropic's research, Oct 2026 (METR)
Finding13 sources
Big news
OpenAI's AI agents built 948 attacks that let an ordinary user take over a Linux machine
OpenAI says an unreleased model built working attacks for 948 Linux flaws in one run, about $1,279 each, and proposed a fix for each.
948working Linux attacks from one automated run (OpenAI's own count)
Milestone2 sources
Notable
Anthropic's first drug results with Claude: six partners, one 25-hour review cut to 3 hours
Anthropic says Claude helped redesign a cancer therapy and drew up 35,000 antibody designs in a month. Almost none is tested in people yet.
6drug developers in Anthropic's first results, nearly all preclinical
Finding1 source
Same AI, same puzzle games: 62.7% with the standard test setup, 99.9% with its maker's own
How a model is run in a test can matter as much as the model. GPT-6 Astra scored 62.7% or 99.9% depending on the software wrapped around it.
62.7% → 99.9%GPT-6 Astra on ARC-AGI-3: standard harness vs OpenAI's own adapter
Finding5 sources
AI went from about 4% to 95% on puzzles people can solve, and under 1% to 81% on puzzle games
Puzzles ordinary people can solve stumped AI in early 2025. By September 2026 the best model solved 95%, checked by the ARC Prize team.
95.0%GPT-6 Astra on ARC-AGI-2, verified by ARC Prize (human panel: 100%)
Milestone10 sources
Key charts
New tests get beaten within one to two years
Best AI score on each benchmark, starting when the benchmark was published.
What this measures: Each line is a test built to be hard for AI. Watch how quickly each one climbs from near zero toward the top once it exists.
- GPQA Diamond
- FrontierMath Tiers 1–3
- Humanity's Last Exam
- ARC-AGI-2
- ARC-AGI-3
0%50%100%202420252026
GPQA and FrontierMath points are Epoch AI's independent runs (FrontierMath on the corrected 2026 set, re-run on older models); HLE is Scale AI's no-tools leaderboard; ARC points are ARC Prize-verified with the standard setup. Fast saturation partly reflects benchmark flaws and contamination, not only skill.
Show data table and sources
| Series | Date | Label | Value |
|---|---|---|---|
| GPQA Diamond | 20 Nov 2023 | GPT-4 Turbo | 42% |
| GPQA Diamond | 20 Jun 2024 | Claude 3.5 Sonnet | 54% |
| GPQA Diamond | 12 Sept 2024 | o1-mini | 62% |
| GPQA Diamond | 17 Dec 2024 | o1 | 77% |
| GPQA Diamond | 25 Mar 2025 | Gemini 2.5 Pro | 84% |
| GPQA Diamond | 9 Jul 2025 | Grok 4 | 87% |
| GPQA Diamond | 18 Nov 2025 | Gemini 3 Pro | 93% |
| GPQA Diamond | 5 Mar 2026 | GPT-5.4 Pro | 95% |
| GPQA Diamond | 30 Sept 2026 | Gemini 4 Argon | 96% |
| FrontierMath Tiers 1–3 | 8 Nov 2024 | Launch: under 2% | 2% |
| FrontierMath Tiers 1–3 | 17 Dec 2024 | o1 | 15% |
| FrontierMath Tiers 1–3 | 16 Apr 2025 | o4-mini | 36% |
| FrontierMath Tiers 1–3 | 7 Aug 2025 | GPT-5 | 55% |
| FrontierMath Tiers 1–3 | 11 Dec 2025 | GPT-5.2 Pro | 74% |
| FrontierMath Tiers 1–3 | 23 Apr 2026 | GPT-5.5 Pro | 88% |
| FrontierMath Tiers 1–3 | 3 Sept 2026 | GPT-6 Astra | 94% |
| Humanity's Last Exam | 24 Jan 2025 | o1 | 8% |
| Humanity's Last Exam | 16 Apr 2025 | o3 | 20% |
| Humanity's Last Exam | 7 Aug 2025 | GPT-5 | 25% |
| Humanity's Last Exam | 6 Oct 2025 | GPT-5 Pro | 32% |
| Humanity's Last Exam | 18 Nov 2025 | Gemini 3 Pro | 38% |
| Humanity's Last Exam | 19 Feb 2026 | Gemini 3.1 Pro | 46% |
| Humanity's Last Exam | 2 Sept 2026 | GPT-6 Astra | 55% |
| ARC-AGI-2 | 24 Mar 2025 | o3-preview (est.) | 4% |
| ARC-AGI-2 | 9 Jul 2025 | Grok 4 | 16% |
| ARC-AGI-2 | 24 Nov 2025 | Claude Opus 4.5 | 38% |
| ARC-AGI-2 | 11 Dec 2025 | GPT-5.2 Pro | 54% |
| ARC-AGI-2 | 12 Feb 2026 | Gemini 3 Deep Think | 85% |
| ARC-AGI-2 | 9 Jul 2026 | GPT-5.6 Sol | 93% |
| ARC-AGI-2 | 2 Sept 2026 | GPT-6 Astra | 95% |
| ARC-AGI-3 | 25 Mar 2026 | At launch | 0.5% |
| ARC-AGI-3 | 1 Jun 2026 | Claude Opus 4.8 | 1.5% |
| ARC-AGI-3 | 9 Jul 2026 | GPT-5.6 Sol | 7.8% |
| ARC-AGI-3 | 24 Jul 2026 | Claude Opus 5 | 30% |
| ARC-AGI-3 | 2 Sept 2026 | GPT-6 Astra | 63% |
AI 2027 scenario vs reality
How fast each measure moved compared with the AI 2027 scenario, as graded by its authors. 1.0 = exactly on the scenario's pace.
What this measures: AI 2027 is a detailed 2025 forecast in which AI that out-codes the best humans arrives inside a lab by March 2027. A bar longer than the dashed line means reality is moving faster than the scenario on that measure.
Global AI data-centre spending~$850B/yr vs ~$510B predicted (Aug 2026)
1.76×
Leading-lab revenueAnthropic $47B annualised (May 2026)
1.43×
AI chips in the world~24.5M H100-equivalents vs ~22M predicted
1.14×
Length of tasks AI can do (METR)Claude Mythos Preview, Apr 2026
1.05×
Using a computer (OSWorld)60.8% vs 65% predicted, Aug 2025
0.84×
Top lab valuation$852B vs $1.3T predicted (Mar 2026)
0.7×
Largest training runauthors' estimate
0.58×
AI speeding up AI researchauthors' Aug 2026 estimate; METR's Nov 2026 audit points higher
0.31×
Coding benchmark (SWE-bench)74.5% vs 85% predicted for mid-2025
0.18×
Public concern about AIGallup, 4-month average to Jul 2026
0.1×
These grades stop in August 2026. Since then, METR's audit (5 November) found AI roughly doubling Anthropic's research pace, a level the scenario reaches around March 2027, so the “AI speeding up AI research” bar is out of date. GPT-6 Astra and Claude Fable 5.2 are not graded yet.
Paces are the AI Futures Project's own grading (Feb and Aug 2026) and are sensitive to how each metric is measured. They are the scenario authors' estimates, not independent ones.
Source: AI Futures ProjectData as of 10 Nov 2026
Show data table and sources
| Item | Value | Note |
|---|---|---|
| Global AI data-centre spending | 1.76× | ~$850B/yr vs ~$510B predicted (Aug 2026) |
| Leading-lab revenue | 1.43× | Anthropic $47B annualised (May 2026) |
| AI chips in the world | 1.14× | ~24.5M H100-equivalents vs ~22M predicted |
| Length of tasks AI can do (METR) | 1.05× | Claude Mythos Preview, Apr 2026 |
| Using a computer (OSWorld) | 0.84× | 60.8% vs 65% predicted, Aug 2025 |
| Top lab valuation | 0.7× | $852B vs $1.3T predicted (Mar 2026) |
| Largest training run | 0.58× | authors' estimate |
| AI speeding up AI research | 0.31× | authors' Aug 2026 estimate; METR's Nov 2026 audit points higher |
| Coding benchmark (SWE-bench) | 0.18× | 74.5% vs 85% predicted for mid-2025 |
| Public concern about AI | 0.1× | Gallup, 4-month average to Jul 2026 |
- AI 2027 grading spreadsheet (Aug 2026) (AI Futures Project)
- AI 2027 grading spreadsheet (Feb 2026) (AI Futures Project)
- Q2.5 2026 Timelines Update: Uplift and Revenue (AI Futures Project)
- Grading AI 2027's 2025 Predictions (AI Futures Project)
Water: one AI prompt vs everyday things
Millilitres of water. Log scale: each gridline is 10× the previous. Blue = AI.
What this measures: How much water one AI text prompt uses, next to everyday activities and products.
One AI text prompt (Gemini median, on-site)data-centre cooling
0.26 mL
Same prompt incl. power-plant water≈ 5.2 prompts · our calculation, US average
1.35 mL
One AI response, full life cycle (Mistral Large 2)≈ 170 prompts · incl. hardware manufacturing
45 mL
2024 “bottle per email” estimate (GPT-4)≈ 2,000 prompts · assumed ~140 Wh per email
519 mL
One toilet flush (US standard)≈ 23K prompts · direct use
6.06 L
One 10-minute shower≈ 360K prompts · direct use
94.6 L
One cup of coffee≈ 540K prompts · supply-chain footprint
140 L
One quarter-pound beef patty≈ 6.7M prompts · supply-chain footprint
1.75 m³
One pair of jeans, cradle to grave≈ 15M prompts · life-cycle footprint
3.78 m³
Scopes differ, so compare like with like. The AI prompt, toilet and shower rows are direct use. Coffee, beef and jeans are supply-chain or life-cycle footprints (coffee and beef include rainwater taken up by crops and pasture); the fair AI comparison for those is the life-cycle row (Mistral, 45 mL). Long-document, reasoning and agentic requests use far more than a median text prompt.
Sources: Google Cloud, Mistral AI, Lawrence Berkeley National Laboratory, Fortune, US EPA WaterSense, University of Twente (Chapagain & Hoekstra), Mekonnen & Hoekstra, Levi Strauss & Co.Data as of 9 Nov 2026
Show data table and sources
| Item | Value | Note |
|---|---|---|
| One AI text prompt (Gemini median, on-site) | 0.26 mL | data-centre cooling |
| Same prompt incl. power-plant water | 1.35 mL | our calculation, US average |
| One AI response, full life cycle (Mistral Large 2) | 45 mL | incl. hardware manufacturing |
| 2024 “bottle per email” estimate (GPT-4) | 519 mL | assumed ~140 Wh per email |
| One toilet flush (US standard) | 6.06 L | direct use |
| One 10-minute shower | 94.6 L | direct use |
| One cup of coffee | 140 L | supply-chain footprint |
| One quarter-pound beef patty | 1.75 m³ | supply-chain footprint |
| One pair of jeans, cradle to grave | 3.78 m³ | life-cycle footprint |
- How much energy does Google's AI use? We did the math (Google Cloud)
- Our contribution to a global environmental standard for AI (Mistral AI)
- 2024 United States Data Center Energy Usage Report (Lawrence Berkeley National Laboratory)
- AI water usage, droughts, ChatGPT, Microsoft (reports the 519 mL estimate) (Fortune)
- Residential toilets (US EPA WaterSense)
- Showerheads (US EPA WaterSense)
- The water footprint of coffee and tea consumption in the Netherlands (University of Twente (Chapagain & Hoekstra))
- The green, blue and grey water footprint of farm animals and animal products (Mekonnen & Hoekstra)
- Levi's 501 life cycle assessment results (Levi Strauss & Co.)
Myth checks
Popular claims about AI, stated fairly and compared with the data.
Myth check: “AI can't create new knowledge, it only remixes its training data”
Verdict: FalseAI disproved an 80-year-old conjecture and designed antibiotics that worked in mice. Real finds come where answers can be checked.
Myth check7 sources
Hype check: “Humanoid robots in demos and viral videos act on their own”
Verdict: MisleadingSome do, but often a person steers them remotely without saying so. About 60% of robots in a 2026 Beijing race were remote-controlled.
Myth check7 sources
Hype check: “AI solved the Navier–Stokes Millennium Prize problem”
Verdict: MisleadingOpenAI's proof needs a designed outside force pushing the fluid, which the prize rules allow. A fluid left alone is still an open question.
Myth check8 sources