Writing
Frontier models by scale, 2021–2026
In November 2021 the largest language-model training run on record used about 2×10²⁴ floating-point operations (FLOP). In September 2026 GPT-6 Astra was estimated at 9.4×10²⁶, about 460 times more. This post splits five years of language models into bands of training compute and follows each band through time. Scaling laws describe one band against another at a fixed date. Here the question is what happens to each band as the dates move.
Setup
All data comes from Epoch AI’s public datasets, downloaded on 15 November 2026. The AI Models dataset gives each model’s release date and an estimate of its training compute. We kept language and multimodal models released from 15 November 2021 to 29 October 2026.
Capability is the Epoch Capabilities Index (ECI). Epoch fits one item-response model to 57 benchmarks, so models tested on different benchmarks land on one scale. The scale is anchored at 130 for Claude 3.5 Sonnet and 150 for GPT-5. 256 models have an ECI score; the first was released in February 2023. 104 of them also have a compute estimate, and those 104 carry every band result below. For 2021 and 2022 we use MMLU, a 57-subject multiple-choice exam, because the ECI does not reach back that far.
We cut training compute into four bands, each ten times the one below:
| Band | Training compute | Models with ECI | Typical members |
|---|---|---|---|
| Small | below 10²³ FLOP | 16 | LLaMA-13B, Phi-2, phi-3-mini |
| Mid | 10²³ to 10²⁴ | 29 | Llama 2-70B, Mixtral 8x7B, Phi-4, Qwen3-30B-A3B |
| Large | 10²⁴ to 10²⁵ | 42 | Llama 3.1-70B, DeepSeek-V3, Kimi K2, DeepSeek V4 |
| Frontier | 10²⁵ and above | 17 | GPT-4, Claude 3.5 Sonnet, Grok 4, GPT-5, GPT-6 Astra |
We measure four things. The first is the growth of the largest training runs. The second is each band’s slope: ECI points gained per year, from a least-squares line through the band’s models. The third is catch-up time: the months between the first model in a higher band and the first in a lower band to reach the same score. The fourth is a joint fit of ECI on compute and date, which prices one year of progress in compute.
The largest training run grew 460-fold, about 3× a year
Figure 1 shows the 363 language and multimodal models in the dataset trained with at least 10²³ FLOP. At the start the record was Google’s FLAN 137B, from September 2021, at 2.0×10²⁴. It was then passed by GPT-3.5 and Minerva (2.6 and 2.7×10²⁴, 2022), GPT-4 (2.1×10²⁵, March 2023), Gemini 1.0 Ultra (5.0×10²⁵, December 2023), Grok 3, GPT-4.5 and Grok 4 (3.5 to 5.0×10²⁶, 2025), and GPT-6 Astra (9.4×10²⁶, September 2026).
A line through the five largest models of each year rises 3.0× a year. The record moved in steps, not smoothly. It stood still for 14 months after Gemini 1.0 Ultra and again for 14 months after Grok 4.
| Year | 10²³–10²⁴ | 10²⁴–10²⁵ | 10²⁵–10²⁶ | 10²⁶ and above |
|---|---|---|---|---|
| 2021 | 2 | 2 | 0 | 0 |
| 2022 | 20 | 5 | 0 | 0 |
| 2023 | 51 | 13 | 4 | 0 |
| 2024 | 73 | 41 | 6 | 0 |
| 2025 | 40 | 69 | 6 | 3 |
| 2026 | 1 | 18 | 8 | 1 |
The busiest band moved up one decade in 2025. Most of the 10²⁴–10²⁵ models that year are open-weight mixtures of experts, such as DeepSeek-V3 (3.3×10²⁴) and Qwen3-235B-A22B (4.8×10²⁴). The 2026 counts are low because most closed labs no longer publish enough for Epoch to estimate compute: only 15 of the 87 models scored in 2026 have an estimate.
In 2021 and 2022, the same compute already bought more
Before the ECI begins, MMLU shows the first sign of the pattern. Gopher scored 60.0% in December 2021. Chinchilla, released by DeepMind in March 2022, used about the same compute (5.8×10²³ FLOP) on a 70-billion-parameter model with four times as much data. It scored 67.5%. PaLM 540B used 2.5×10²⁴ FLOP, four times more than Chinchilla, and scored 69.3%. Then GPT-4, at 2.1×10²⁵, scored 86.4%.
The 10²³–10²⁴ band kept climbing on MMLU without leaving its compute range:
| Date | Model | Training compute (FLOP) | MMLU |
|---|---|---|---|
| Dec 2021 | Gopher (280B) | 6.3×10²³ | 60.0% |
| Mar 2022 | Chinchilla (70B) | 5.8×10²³ | 67.5% |
| Jul 2023 | Llama 2-70B | 8.1×10²³ | 69.9% |
| Nov 2023 | Yi-34B | 6.1×10²³ | 76.3% |
| Dec 2024 | Phi-4 | 9.3×10²³ | 84.8% |
MMLU stops being useful once the best models pass 85%, because the remaining questions include mislabelled ones. From 2023 on we use the ECI.
Every band gains 12 to 20 ECI points a year
Figure 2 plots the 104 models with both an ECI score and a compute estimate. Each band forms its own rising cloud. The four fitted lines are close to parallel. The dashed line is the best score of any model to date, including closed models with no compute estimate.
| Band | Models | ECI points per year | Period | Best model (ECI) |
|---|---|---|---|---|
| below 10²³ FLOP | 16 | 12.6 ± 6.3 | 2023-02 to 2024-11 | phi-3-mini 3.8B (110.6) |
| 10²³–10²⁴ | 29 | 19.9 ± 2.9 | 2023-02 to 2025-08 | Qwen3-30B-A3B-Thinking (Jul 2025) (138.3) |
| 10²⁴–10²⁵ | 42 | 15.2 ± 0.8 | 2023-07 to 2026-09 | DeepSeek V4.1 Flash (156.7) |
| 10²⁵ and above | 17 | 11.6 ± 1.2 | 2023-03 to 2026-10 | GPT-6 Astra (169.0) |
The two well-sampled bands, 10²⁴–10²⁵ and 10²⁵ and above, climb at 15.2 and 11.6 points a year. The 10²⁵ line is pulled down by Grok 3 and GPT-4.5, two very large runs that scored below smaller models of the same months. If we fit only the models that set a new best within their band, the slopes are 20.4, 15.0, 14.2 and 13.1 points a year, from the smallest band to the largest. The smaller bands rise at least as fast as the largest one.
The small band (below 10²³ FLOP) ends in November 2024. After that, no model in the dataset that small has both a compute estimate and an ECI score, so its slope rests on 16 models and has the widest error.
A band ten times smaller reaches the same score 3 to 13 months later
Parallel lines mean a fixed delay. The table gives the first model in each band to reach each score. Read across a row to see the delay.
| ECI | 10²⁵ and above | 10²⁴–10²⁵ | 10²³–10²⁴ | Delay, top two bands |
|---|---|---|---|---|
| 120 | GPT-4, Mar 2023 | Llama 3-70B, Apr 2024 | Phi-4, Dec 2024 | 13 months |
| 130 | Claude 3.5 Sonnet, Jun 2024 | DeepSeek-V3, Dec 2024 | Qwen3-30B-A3B, Apr 2025 | 6 months |
| 140 | Claude 3.7 Sonnet, Feb 2025 | DeepSeek-R1 (May 2025) | not yet | 3 months |
| 150 | GPT-5, Aug 2025 | DeepSeek-V4-Pro, Apr 2026 | not yet | 9 months |
| 160 | GPT-6 Astra, Sep 2026 | not yet (best 156.7) | not yet | 2+ months so far |
The delay between the top two bands fell from 13 months at ECI 120 to 3 months at ECI 140, then rose to 9 months at 150. Two bands apart, the delay is longer. GPT-4 reached 120 with 2.1×10²⁵ FLOP in March 2023. Phi-4 reached 129 with 9.3×10²³ FLOP, 23 times less, 21 months later.
The mid band (10²³–10²⁴) has no model past 138.4. That is Qwen3-30B-A3B-Thinking, from July 2025, the band’s last model with a compute estimate. Labs still release models this size, but without the figures Epoch needs, so the row ends there.
One year of progress is worth about 17 times the compute
We fit one line through all 104 models: ECI against log₁₀ of training compute and release date together. The fit explains 89% of the variance (R² = 0.888).
| Term | Estimate |
|---|---|
| ECI points per tenfold increase in training compute | 12.2 |
| ECI points per year at fixed training compute | 15.2 |
| Compute worth one year of progress | 17.5× (95% bootstrap interval 7.8× to 59×) |
Dividing the two coefficients gives the compute a year of progress is worth: 15.2 ÷ 12.2 = 1.24 decades, or 17.5×. The largest training runs grew 3.0× a year over the same period. By this measure, what a fixed compute budget can do grew faster than the budgets themselves.
The 17.5× is not pre-training efficiency alone. Epoch’s compute estimates count pre-training and leave out most post-training. Reinforcement learning on reasoning, which arrived with OpenAI’s o1 in September 2024 and DeepSeek-R1 in January 2025, raised scores at almost unchanged estimated compute. DeepSeek-V3 and DeepSeek-R1 share the estimate of 3.3×10²⁴ FLOP; their ECI scores are 131.7 and 138.6.
A lab's smallest model trails the best model by about seven months
Since 2025, compute estimates are missing for most closed models, so the bands thin out. Labs still publish a size ordering through their product tiers. Anthropic ships Haiku, Sonnet and Opus. OpenAI ships nano, mini and full models, and since July 2026 Luna, Terra and Sol. Google ships Flash-Lite, Flash and Pro. We took the 22 small-tier and 22 middle-tier models from those three labs. For each one we counted the months since any model first reached its ECI score.
| Year | Small tier: models | Median months behind | Middle tier: models | Median months behind |
|---|---|---|---|---|
| 2023 | 3 | 5 | 0 | — |
| 2024 | 5 | 8 | 5 | 4 |
| 2025 | 8 | 7 | 8 | 2 |
| 2026 | 6 | 8 | 9 | 5 |
| All | 22 | 7 | 22 | 3.5 |
The small tier has stayed about seven months behind for three years. The delay has not shrunk. At 14 ECI points a year, seven months is about eight points. Claude Haiku 5.5, released in October 2026, scored 160.4: the level GPT-5.5 Pro first reached in April 2026, five and a half months earlier. GPT-5.4 Nano, March 2026, scored 146.4, the level of o3 eleven months before. Gemini 3.5 Flash-Lite, July 2026, is the farthest behind at 15 months.
The middle tier sits closer, at a median of 3.5 months. Claude 3.5 Sonnet was the best model of any lab when it was released in June 2024. One small-tier model was too: o1-mini, a reasoning model, in September 2024.
The top changes hands every six weeks, and the field sits closer to it
Figure 4 looks at all 256 scored models, with or without a compute estimate. Each week, the field is every model released in the year before. We rescale each model in the field so that the week’s best model scores 1 and its weakest scores 0, then smooth each curve over about two weeks. A curve starts at its model’s release and ends a year later. The 20 models that were best in the field for at least three weeks are drawn in full and labelled, and so is Claude Fable 5.2, best in the latest week. The dashed line is the median model of the field.
GPT-4 held the top for 45 weeks, from March 2023 to January 2024. No model has held it as long since. From 2025 on, the top changed hands 14 times, and the median hold was six and a half weeks. The longest was 12 weeks, by GPT-5.2 Pro from December 2025. OpenAI models were on top for 153 of the 193 weeks, Anthropic models for 31 and Google models for 7.
The field closed in on the top through 2025. The scaled median rose from 0.54 in 2023 to 0.76 in 2025. In ECI points, the best model led the median model by 32.5 in 2023, 23.6 in 2024, 15.6 in 2025 and 15.9 in 2026.
In 2026 the scaled median fell to 0.47 while the gap in points held at about 16. The difference is the floor. Min-max scaling depends on the weakest model in the field. Through most of 2025 that was Qwen2.5-Coder-0.5B, at 68.4. It left the field in September 2025 and INTELLECT-1, at 83.5, took its place. When INTELLECT-1 left at the end of November, the floor jumped to 126.1 (Llama 3.3 70B). Each time every curve dropped at once, but no model got worse. Read the curves for order and timing, and the point gaps for distance.
Discussion
Over five years, scale and time traded against each other at a steady rate. Each tenfold step in training compute is worth about 12 ECI points. Each year is worth about 15. The frontier band leads, and the bands below follow on nearly parallel lines. The gap between the top two bands has varied from 3 to 13 months with no clear trend. The gap between a lab’s small tier and the best model has stayed near seven months.
For someone choosing a model, the practical reading is direct. A model one tier down is, on this index, the model one tier up from about half a year earlier. A model trained with a tenth of the compute is a frontier model from three to thirteen months before.
Limits
- Training compute is Epoch AI’s estimate, not a figure labs report. Only 104 of 256 scored models have one, and only 15 of 87 in 2026. Since 2025 the 10²⁴–10²⁵ band is mostly open-weight models from DeepSeek, Moonshot and Alibaba, so it describes those labs more than the field.
- Compute estimates mostly leave out post-training. The joint fit therefore credits reasoning post-training to the passage of time.
- Each ECI score has an uncertainty; the median 90% interval is 2.6 points either way. Delays measured from a single first model to cross a threshold move by months when one score moves by one point.
- The ECI measures benchmarks. It does not measure speed, price, context length or behaviour in long agent runs.
- Tiers were assigned from model names. OpenAI’s o1-mini, o3-mini and o4-mini count as small tier, though each was near the top of the index when released.
More head-to-head evals
| Measure | Particle-1 | DeepSeek V4.1 Flash | GLM-5.3-Flash |
|---|---|---|---|
| Rows right, blind | 91.2% | 89.4% | 90.8% |
| Median seconds per run | 375 | 315 | 144 |
| Measure | GLM-5.3 | DeepSeek V4.1 Flash |
|---|---|---|
| Right, of 75 | 75 | 71 |
| Seconds a question, median | 3.9 | 1.8 |
| Measure | First run | Best since |
|---|---|---|
| Landscape, seconds | 937 | 233 |
| Memo, seconds | 759 | 602 |
| Measure | Qwen3 235B | GLM-5.2 | Gemini 3.8 Flash | Claude Opus 5 |
|---|---|---|---|---|
| Seconds to the plan, median | 27 | 48 | 57 | 94 |
| Plans written, of three runs | 2 | 3 | 3 | 3 |
| Measure | SubQ | jev | Decider | Nimble |
|---|---|---|---|---|
| Found, mean of five columns at section length | 64% | 65% | 66% | 68% |
| Mechanism of action, found at section length | 57% | 51% | 61% | 54% |
| Measure | ask | Apollo | treg |
|---|---|---|---|
| Usable leads | 13 of 40 | 7 of 40 | 6 of 28 |
| Cost per usable lead | $0.037 | $0.27 | $0.23 |
| Measure | ask with jev | ask with Decider |
|---|---|---|
| Questions right, of 92 | 89 | 89 |
| Wrong answers | 1 | 2 |
| Measure | jev | ask |
|---|---|---|
| Our 370 clinical decisions, right | 94.6% | 93.0% |
| PubMedQA, right (best EdgeEvals scored: 62.2%) | 77.0% | 74.4% |
| Measure | ask | Claude Opus 5.5 |
|---|---|---|
| Questions right, of 95 | 59 | 63 |
| Small companies right, of 20 | 13 | 9 |
| Measure | realestate-us | Best of four assistants |
|---|---|---|
| Floor-plan rents right, of 26 | 26 | 23 |
| Amenities claimed the listing doesn't state | 12 | 129 |
| Measure | ask | jev |
|---|---|---|
| Right, of the questions answered | 98% | 79% |
| Wrong answers, of 92 | 2 | 19 |
| Measure | Trained localiser | Best of four frontier models |
|---|---|---|
| Mean overlap with the defect (IoU, 0 to 1) | 0.421 | 0.109 |
Llama 2-70B