Most Capable Frontier AI Models Infographic

What this graphic shows

Figure 1 tracks the most capable AI model in the world, one record at a time. Capability is measured with the Epoch Capabilities Index (ECI), which Epoch AI builds by combining scores from dozens of benchmarks into one scale. The scale is anchored so that Claude 3.5 Sonnet scores 130 and GPT-5 scores 150, which lets models from different years be compared even after individual benchmarks top out. Each step in the chart is a model that beat every earlier model, and the strip beneath it shows how long each lab held the record.

The other cards show where each lab stands today, how far open-weights models trail, how little labs now publish about how their models were trained, and which countries release open-weights models.

OpenAI has held the top spot most of the time

Since GPT-4 launched in March 2023 with an ECI of 121.87, the record has been broken 20 times. An OpenAI model held it on 1,084 of 1,333 days, or 81%. Anthropic held it for 210 days, most recently with Claude Fable 5.1 in September 2026, and Google DeepMind for 39 days with Gemini 2.5 Pro and Gemini 3 Pro. The current record is GPT-6 Astra at 169.07.

The lead is narrow. Each lab's best model sits within a few points of the others: OpenAI 169.07, Anthropic 168.36, Google DeepMind 166.45 and Meta AI 162.02. Epoch's 90% confidence intervals are about three points wide on either side, so the top three are statistically close.

Open models trail by about seven months, and most are Chinese

The best model with downloadable weights is DeepSeek V4.1 Flash 1017 at 159.31. The closed frontier first passed that level on March 5, 2026, with GPT-5.4 Pro at 160.21, about seven months earlier. Five of the six highest-scoring open-weights models come from Chinese labs: DeepSeek, Moonshot and Z.ai. Of the 71 notable open-weights language models released since January 2025, 48 came from Chinese developers.

Labs say less about how their models were built

Epoch rates how precisely it knows each model's training compute. In 2021, 71.1% of notable language models had a figure Epoch called Confident, meaning accurate within a factor of three. For models released in 2026 the share is 20.0%. The largest training run with a published, precise figure is still Meta's Llama 3.1-405B from 2024, at 3.8×10²⁵ FLOP. Epoch's estimate for GPT-6 Astra is 9.4×10²⁶ FLOP, about 25 times larger, but it rates that estimate only Likely, meaning accurate within a factor of ten.

Methodology & caveats

Capability scores are Epoch AI's ECI values as downloaded on November 6, 2026 (253 models with release dates from February 2023 to October 2026). The record steps, tenure shares and the open-weights lag are data.co derivations from that file. A model sets a record if its score is higher than every model released before it. Records start with GPT-4 and run to November 6, 2026. The lag is the time between the best open-weights model's release and the first closed model to score at least as high. ECI scores carry 90% confidence intervals of roughly ±3 points; anchor models have none. Models Epoch has not scored do not appear, and ECI is revised as Epoch adds benchmarks.

Model facts come from Epoch AI's Notable AI Models dataset (models in the language domain; Epoch's notability criteria include state-of-the-art results, high citation counts and wide use). The disclosure percentages divide the number of models rated Confident by all models released that year, rounded to one decimal place; 2026 covers releases through October 28. The roughly 25× compute ratio divides Epoch's Likely estimate for GPT-6 Astra by Meta's published figure for Llama 3.1-405B. Counts by country use the developer's country as Epoch records it. Epoch's data is published under CC BY 4.0.