Skip to main content
EvidenceAI ModelsVisualization detail

Estimated training FLOP over time.

Training compute growth

Takeaway: A roughly exponential climb off the GPT-3 baseline.

Training compute

Estimated training compute of frontier models by year (Epoch AI).

Takeaway: A roughly exponential climb off the GPT-3 baseline.

Training compute

A roughly exponential climb off the GPT-3 baseline.

Rendering the interactive chart…

Expand data table
Download CSV
Training compute growth: data table
PeriodSeriesValueValue basisSource
2020Flops estimate3.1e+23ActualEpoch AI: compute & training trends
2022Flops estimate2.5e+24ActualEpoch AI: compute & training trends
2023Flops estimate2e+25ActualEpoch AI: compute & training trends
2024Flops estimate5e+25ActualEpoch AI: compute & training trends
2025Flops estimate5e+26ActualEpoch AI: compute & training trends

Training compute growth: full data table

SOURCEEpoch AI: compute & training trends
VALUE BASISActual
LAST VERIFIEDSep 10, 2026

Caveat: Flop figures are third-party estimates (Epoch AI), not disclosed by labs.

How this is calculated

Data table

Download CSV
Training compute growth: training_compute data table (The Frontier AI Model Race (2026))
YearSeriesSourceValue basisFlops estimate
2020training_computeEpoch AI: compute & training trendsActual3.1e+23
2022training_computeEpoch AI: compute & training trendsActual2.5e+24
2023training_computeEpoch AI: compute & training trendsActual2e+25
2024training_computeEpoch AI: compute & training trendsActual5e+25
2025training_computeEpoch AI: compute & training trendsActual5e+26

Method

How this evidence was built

What it measures

Source-backed values are seeded for four of the five charts: the release cadence by lab (2024 → 2026, from each lab’s release notes), the capability-vs-cost scatter (Artificial Analysis Intelligence Index vs output $/M tokens, June 2026 snapshot), training-compute growth by year (Epoch AI, corroborated by Stanford hai), and a safety-gated capabilities matrix.

How it is calculated

  1. This is an interpretation of each model’s published safety policy, not a measured benchmark.
  2. Placeholder: the per-lab benchmark-trajectory chart is left unseeded, a consistent historical per-quarter, per-lab benchmark series was not sourceable without mixing incompatible benchmarks.

Key context

  • Every numeric point carries a source citation and a basis label.
  • Sources are each lab’s own model cards/release notes, Artificial Analysis, Epoch AI, and Stanford hai: cross-checked against public release timelines.
  • Editorial encoding: the safety-gated matrix scores each model × domain as 3 = allowed / 2 = gated / 1 = blocked.

Key numbers

  • Four labs trade the lead quarter to quarter: releases are clustered, not steady (anthropic-news / openai-research / google-deepmind / meta-ai).
  • A leading model tops the Artificial Analysis Intelligence Index at ~62 while costing ~$25/M output tokens, capability up, price-per-capability down (artificial-analysis).
  • Training compute scales off a ~3.1e23-flop gpt-3 (2020) baseline (epoch-ai, corroborated by Stanford hai).
  • The safety matrix shows hard limits blocking high-risk cyber/bio/chem while gating legal/medical use, an editorial, not measured, scale (anthropic-news).

Caveats

Caveat: Flop figures are third-party estimates (Epoch AI), not disclosed by labs.

Caveat: Editorial 1=blocked / 2=gated / 3=allowed scale, a posture summary, not a measured benchmark.

Do not use this to

  • Exact, lab-disclosed training compute (flop figures are third-party estimates).
  • Real-world safety outcomes (the matrix is a posture summary on an editorial scale).
  • Market share or revenue by lab (not in this index).

Scope

June 2026

Sources

Last verified 10 Sept 2026

How relevant was this information?