Anthropic's models now improve about twice as fast as before: 1.9× on benchmarks since Oct 2025 (95% CI 1.5–2.8×), per independent evaluator METR, based on Anthropic's internal data. METR guesses that most of this acceleration, perhaps around 80%, comes from Anthropic's own use of AI. METR acceleration
Cognizant Digital Studio Amsterdam · 19 November 2026
Sovereign AI who holds the keys
A journey from the AI frontier to the infrastructure behind it. Europe notices a dependency, then builds. Explore the numbers and their sources.
Explore the storyEurope notices a dependency, then builds.
Aragorn Meulendijks · ItsMrMetaverse01 Frontier update
The frontier, this week
A snapshot of the models, the benchmarks and the pace of progress.
Epoch on Fable 5.2: “above what the pre-Astra trend for reasoning models would have predicted” Epoch AI on X
Axis starts at 150. Both new models are tied for #1 of 241 tracked models (as of 18 Nov 2026). Epoch ECI
New highs at release, our reading of Epoch data; selected milestones. Dashed line: our own straight-line fit to Epoch's frontier records (+16 points per year, in line with the pace Epoch reports for reasoning models). By our own estimate from Epoch data, the two new models sit about 2–3 points above that line. This is our arithmetic, not an Epoch figure. Scores carry about ±4 points of uncertainty and can be revised. Epoch ECIOur calculation
The benchmarks
Three models now score 100%: GPT-6.1 Sol (Sept) and, new in November, Claude Fable 5.2 and GPT-6.2 Astra. GPT-5 (released Aug 2025) scores about 22% in Epoch's current Tier 4 v2 evaluation. Epoch benchmarks
ARC Prize standard harness, each model at its best setting. On the provider-adapter harness most frontier models score 97–100%. ARC Prize
The top models are within about 2 points of each other: this test is nearly maxed out. ARC Prize
The top three are statistically tied, close to the benchmark's ceiling. Epoch benchmarks
Error margins are large (about ±5–9 points). FrontierSWE
Best agent when the index launched (Oct 2025): 2.5%. GPT-5: 1.7%. Remote Labor Index
Margins are about ±4–5 points, so neighbouring models are statistically close. Mercor
METR: an early version of Claude Mythos Preview handles tasks that take a skilled human at least 16 hours, at 50% success; METR warns results above 16 hours are unreliable (raw point estimate 17.4 h, 95% CI 8.5–55 h). Opus 4.6 has wide uncertainty (5 to 60 h). Latest METR data: May 2026; METR stopped updating the page on 8 Sep 2026. No METR horizon exists for Fable 5.2 or GPT-6.2 Astra. METR's Frontier Risk Report (May 2026, data from Feb–Mar 2026): the most capable model labs shared with METR had a 50% point estimate of 16–20 hours, past the roughly 16 hours its task suite can reliably measure. METR horizonsMETR risk
AI is now speeding up AI
Ajeya Cotra estimates Anthropic's researchers may now get roughly a 10× boost from AI. She compares it to living ten days for every day outside, and calls the estimate very uncertain.
Ajeya Cotra: “it looks like the 'dramatic acceleration' threshold has already been crossed” (6 Nov 2026). Ajeya Cotra
the year a superhuman AI researcher will likely arrive, per the AI Futures Project's forecast (10 Nov 2026), assuming no deliberate slowdown. AI Futures Project
the cost of reaching a fixed level of AI performance has fallen about 13× a year since 2023 (about 47% per quarter, average across five benchmarks).
Same score on a PhD-level science exam (GPQA Diamond): from about 30 cents a question to about four hundredths of a cent. About 725× cheaper in under 18 months (Epoch AI estimate). Epoch price
02 Sovereign AI
Europe notices a dependency
History shows how access to infrastructure can become a strategic question.
SIGNAL · 02
The network is subject to law
March 2012: under EU sanctions, SWIFT cut off EU-sanctioned Iranian banks. March 2022: EU sanctions cut seven major Russian banks, including VTB, off from SWIFT. SWIFT sanctionsEU Council
03 Sovereign AI
Compute and models
The frontier is global. Its inputs and outputs are not evenly distributed.
SIGNAL · 03
Where new AI compute landed
Of the new AI compute that arrived worldwide from 2022 to June 2026, the US received about 72%, China about 10%, and the EU about 5%. Epoch AI estimates about 27 million H100-equivalents arrived worldwide. Estimates outside the US are less certain. Epoch AI imports
SIGNAL · 04
Notable models by origin
Stanford’s AI Index counts 59 notable AI models from the US, 35 from China and 2 from Europe, including the UK, in 2025. Stanford AI Index
SIGNAL · 05
Open weights follow the frontier
As of May 2026, Epoch AI found that since January 2026 the best open-weight models trailed the closed frontier by about 4 months on average, or 8 ECI points. The earlier average was about 3 months from January 2023 to October 2025. Epoch open weights
04 Sovereign AI
The cloud layer
European providers grew, while the market grew faster.
SIGNAL · 06
A smaller share of a larger market
European providers’ share of Europe’s cloud infrastructure market fell from 29% in 2017 to 15% in 2022 and held around 15% through mid-2025. Amazon, Microsoft and Google held 70% in 2025. Europe includes the UK. Synergy Research
05 Sovereign AI
Finance at the frontier
Regulated institutions face both the opportunity and the operational responsibility.
SIGNAL · 07
Generative AI is already in banking
About 85% of the banks the ECB supervises directly already use generative AI, according to ECB supervisory data from September 2026. In July 2026 the ECB told every bank it directly supervises to submit an action plan on frontier-AI risks by 31 October 2026. The ECB is assessing those plans. On 16 November 2026, Google Threat Intelligence Group reported that attackers use agentic frameworks to automate parts of live intrusions. ECBECB letterGoogle Threat Intelligence
SIGNAL · 08
Critical providers enter the frame
In November 2025 the EU’s financial supervisors named 19 technology providers as critical under DORA. This was the first list. DORA list
SIGNAL · 09
The high-risk timetable has changed
The EU Digital Omnibus on AI has been in force since 27 July 2026. It delays AI Act high-risk obligations to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems. EU regulation
06 Sovereign AI
Then Europe builds
From policy to infrastructure, the response is taking shape.
SIGNAL · 10
The gigafactory ambition
July 2026: the EU opened its AI Gigafactories call. Up to seven sites, up to €10B in public money, expected to unlock at least €20B in private investment. More than €30B in total. European Commission
SIGNAL · 11
A capacity proposal
The EU Cloud and AI Development Act, a Commission proposal of 3 June 2026, aims to triple EU data-centre capacity within 5–7 years. CADA proposal
SIGNAL · 12
A Dutch AI factory
Over €200 million for a Dutch AI factory in Groningen: €130 million from the national government and the region, plus about €70 million from the EU, granted October 2025. Dutch governmentEU funding granted
SIGNAL · 13
A vital piece of the supply chain
ASML is the only company in the world that makes EUV lithography machines, according to its annual report filed in February 2026. Annual report
SIGNAL · 14
Output per person
One AI lab reports over $6.6M revenue per employee for the trailing twelve months to September 2026. For IT services, about $63K is our calculation from published figures, annualising a quarter. Revenue is not profit, and the time bases differ. S-1/AOur calculation
Sources
Trace every number
Explore the source behind each figure and chart.
More numbers from the talk
- 85 / 120GPT-3.5 and o1-preview on Mensa Norway puzzles; not validated IQ scoresTrackingAI
- 174 / 174joint ECI record, rounded · 18 Nov 2026Epoch ECI
- 100%three models on FrontierMath Tier 4 v2, maximum reasoningEpoch raw data
- 81.2%Fable 5.2 on ARC-AGI-3, standard harnessARC Prize
- 1.9×benchmark capability growth at one lab; METR estimateMETR acceleration
- ≈13×annual decline in fixed-performance AI costEpoch price
- 88% / 8%finance professionals: expected impact / very well preparedAICPA & CIMA
- ≈10×research acceleration estimate; highly uncertainAjeya Cotra
- 2027likely superhuman AI researcher, conditional forecastAI Futures Project
- 12.7% / 50.3%maths benchmark forecast / actual score in 2022Steinhardt
Reliability: HIGH = primary source or official statistics; MEDIUM = analyst, research firm or secondary source; LOW = speaker perspective.