Where AI Research Actually Happens ended on one bar chart: the best open-weight models trail the closed frontier by four months on Epoch's index, ten on private benchmarks, sixteen on compute. A single bar hides the interesting part, which is how that lag has moved. This post is the lag over time, and what it says about a sovereign stack that runs on weights it did not train.
The data is Epoch AI's Capabilities Index (ECI), a composite of more than fifty benchmarks fitted onto one scale, anchored so that Claude 3.5 Sonnet scores 130 and GPT-5 scores 150. I took the published scores on 30 October 2026: 273 models with a score and a release date, 143 of them closed, 123 with downloadable weights, and seven that Epoch files as neither (a limited-access Gemini preview, hosted-only Qwen variants, and Mistral Large 4 itself, scored as an API preview), which I left out. The file is under every chart, and every number in this post comes from one script over it.
The highest ECI score available with closed weights and with open weights, by release date. Each dot is a new record. Epoch AI Capabilities Index, 30 October 2026 fit.Download the data (CSV)
The closed frontier has gone from GPT-4 at 126 to GPT-6 Astra at 169 in three and a half years. The open-weight frontier has gone from LLaMA-65B at 110 to DeepSeek V4.1 Flash at 159. Both lines climb at about the same rate, which is the first thing worth knowing: the open models are not falling further behind in capability terms, they are climbing a staircase that someone else built a few months earlier.
The second thing is where the two lines nearly meet. In January 2025 DeepSeek-R1 landed at 139, three points below o1, which had been the closed frontier for a month. That is the closest the open line has come to the closed one in the whole series.
For each day, the best open-weight model available, and how many months earlier the closed frontier first reached its score. Dots are the open-weight records, coloured by the country of the lab that released them. Derived from the Epoch data above.Download the data (CSV)
The measure is simple. Take the best open-weight model available on a given day. Find the first day the closed frontier reached that score. The lag is the time between the two. It climbs by one day per day between open releases and drops at each new record, which gives the sawtooth.
The sawtooth has three regions. Through 2023 and the first half of 2024 the closed frontier was GPT-4 and its successors, stuck within three points of 126 for fourteen months while the open models crawled up from 110. By June 2024 Qwen2-72B had got to 125 and the lag was fifteen months, and the day before Llama 3.1 405B it touched sixteen. Then the 405B model matched GPT-4o and the lag fell to two months in one release.
The second region is the eighteen months from July 2024 to the end of 2025, when the lag never got above nine months and averaged four to six. DeepSeek-R1 in January 2025 brought it down to one month, the lowest point in the data.
The third region is 2026, where the lag has settled into a band between four and nine months and the sawtooth teeth have the same shape every time: a Chinese lab releases, the lag drops to four or five, two or three months pass, it climbs to eight or nine, another Chinese lab releases. The mean for 2026 to date is 6.7 months. On 30 October it was 7.9, because GPT-6 Astra had gone out in September and nothing open had answered it yet.
The half-year means, for the record:
Half year
Mean lag, months
Mean gap, ECI points
2023 H2
6.6
9.2
2024 H1
12.6
6.4
2024 H2
5.9
5.2
2025 H1
3.8
5.5
2025 H2
6.2
6.4
2026 H1
7.0
10.7
2026 H2 to 30 Oct
6.3
8.9
Epoch's own figure is lower than mine, and the difference is the method. Their May 2026 analysis counts an open model as having caught up with a closed one when it beats it in at least five per cent of bootstrap resamples, which is a generous reading of two overlapping error bars. On that reading the lag since January 2026 is four months. On a strict reading, where the open model's central score has to match, they get six, and my day-by-day version of the same strict reading gets between six and seven. Their October 2025 analysis of the two years before that gave 3.5 months with a 90% interval of 1.1 to 5.3, which matches the middle region of the chart. Both pieces say the gap is slightly wider than it was. The chart agrees, and adds that the widening happened in capability points before it happened in months: the ECI gap was five or six points through 2025 and has been nine to eleven through 2026. The closed labs have found a faster gear, and the open labs are matching their pace from further back.
Which country's lab released the best open-weight model available, day by day. Same data.Download the data (CSV)
Of the 1,344 days from the first LLaMA to the end of October 2026, an American lab held the open-weight record for 237, a European lab for 234 and a Chinese lab for 873. Europe's share is Mistral for 129 days over the winter of 2023 to 2024, and 105 days in 2023 when Stability AI's Stable Beluga 2 was the best open model in the world, which is the only time a British lab has appeared on this strip.
Since 26 December 2024, the day DeepSeek-V3 went past Phi-4, the record has been Chinese without a break: 673 days. DeepSeek, Alibaba, Moonshot and Z.ai have passed it between them fourteen times. Meta's best 2025 release, Llama 4 Maverick, scored 132 in April against DeepSeek-R1's 139 from January. OpenAI's gpt-oss-120b in August 2025 scored 140 against a Chinese record of 144. The American open frontier is now Meta's Muse Spark Open at 155, released in October 2026 and ten months behind the closed line.
The same strip, coloured by the licence the record holder was released under, using Epoch's three categories.Download the data (CSV)
The licence strip matters more than the country strip if you are the one deploying. The record holder was unrestricted (Apache 2.0, MIT or equivalent) for 803 of those days, under a use-restricted licence of the Llama kind for 199, and non-commercial for 342. The non-commercial days are the first LLaMA, Stable Beluga 2, and then ninety-three days from July to October 2026 when the best open-weight model in the world was Kimi K3, released under Moonshot's own licence: free until your group revenue passes twenty million dollars a year, then a separate agreement. Epoch files that as non-commercial. For most of the firms I would expect to be reading this, it is.
So the model you would have picked for a production stack in September 2026, on capability alone, was one you could not legally run in production. DeepSeek V4.1 Flash has since taken the record back under MIT, and the strip is green again. It was red for a quarter, and nothing stops it going red again.
The best open-weight score from a Chinese, American and European lab, each as its own staircase, against the closed frontier. Mistral Large 4 is shown hollow at its weights release on 29 October 2026, at the score Epoch gave its API preview three weeks earlier.Download the data (CSV)
A sovereign stack argument usually stops at “open weights”, as if open weights were one thing. The chart splits them by where the lab is.
The Chinese line is the open frontier. It has been the top open line since the end of 2024 and sits just under eight months behind the closed frontier today.
The American line is Meta, Microsoft, OpenAI's one open release, Nvidia, Thinking Machines and Meta again. It fell behind the Chinese line in December 2024 and has stayed behind it, by about a year for most of 2026: Gemma 4 in April was twelve months behind the closed frontier, Nemotron 3 Ultra in June was nearly fourteen, and Muse Spark Open in October is ten.
The European line is Mistral and only Mistral. It was the top open line for part of 2023 and 2024, when Mixtral 8x7B and 8x22B were the best models anyone would let you download. Since then every Mistral release has been well below both other open lines. Mistral Medium 3.5, the best European open-weight model Epoch had scored at the end of October, is at 141 and sixteen months behind the closed frontier.
Mistral Large 4 changes this, and the timing is why it is a hollow dot. Epoch scored the API preview at 150.5 on 6 October. The weights went up on Hugging Face on 29 October under Mistral's new Ecosystem Open-Weight licence, a modified MIT whose one restriction is that selling the model as a hosted service needs a separate commercial licence. If the downloadable weights keep the preview score, which they should, Europe's open frontier jumps from sixteen months behind to under thirteen. That puts it level with the American line, which only passed 150 itself in October, and six months behind the Chinese one. The closed model it matches is GPT-5 Pro, the frontier of October 2025.
If “sovereign” means weights from a lab under European law, the stack post already said you pay a capability price for it. The chart puts a number on it: a year behind the frontier for Mistral Large 4, sixteen months for anything before it, against seven or eight for Chinese weights you accept the provenance of.
ECI against estimated training compute for the models Epoch has both numbers for. Compute estimates from Epoch's open-versus-closed dataset and the Notable AI Models table; most recent closed models have no public estimate and are missing.Download the data (CSV)
Compute Isn't the Constraint used DeepSeek-V3 as the exception: near-frontier capability for 2.8 million H800 hours, which the technical report prices at about five and a half million dollars. The scatter says it stopped being an exception the moment it happened.
Every open-weight model above 140 on this chart sits between two and twenty times 1024 FLOP. Every closed model above 140 with a public estimate sits at 1025 or above. GPT-5 at 150 used 6.6 times 1025. Kimi K2.5 reached 148 five months later on 5.8 times 1024, an eleventh of it. GPT-6 Astra at 169 is estimated at 9.4 times 1026. DeepSeek V4.1 Flash at 156 is on 4.3 times 1024, which is about one two-hundredth.
The November 2024 Epoch report measured a fifteen to sixteen month lag in training compute, with Llama 3.1 405B as the open model that finally spent what GPT-4 had spent. Nobody has released weights trained on more compute than that model since. Epoch's largest open estimate after it is Kimi K3 at 2 times 1025, half of Llama's budget, scoring 158 to Llama's 129. The compute lag is now a measure of something the open labs have decided not to do, and the capability lag of seven months is what their alternative costs them.
Which is the point for the research argument from June. The thing that put DeepSeek, Moonshot and Z.ai seven months behind GPT-6 rather than three years behind it is not a cluster. It is the set of techniques that make a 4 times 1024 FLOP training run score 156, most of which they published, and all of which were developed by people who could not buy their way out of the constraint.
The index is one lens. Here are the others, with my reading of each.
Measure
Lag
Source and date
Epoch Capabilities Index, allowing a draw within the error bars
4 months
Epoch AI, May 2026, from 1 January 2026
Epoch Capabilities Index, strict
6 to 8 months
This post, 2026 to date; Epoch's strict figure is 6
Cyber ECI
6 to 7 months
Epoch AI, October 2026
Public benchmarks
4 to 6 months
Håvard Tveit Ihle, May 2026, 9 benchmarks
Private benchmarks
8 to 10 months
Same, 8 benchmarks
Training compute
15 to 16 months
Epoch AI, November 2024; no open model has passed Llama 3.1 405B since
The cyber row is new since the six charts post. Epoch's cyber-only index puts Z.ai's GLM-5.3 above GPT-5.4 and below Claude Mythos Preview, so the open cyber frontier in October 2026 was where the closed one stood six or seven months earlier. If you are picking open weights for security work, that is the number.
The public and private benchmark rows come from Håvard Tveit Ihle's analysis of Epoch's benchmark data, not from Epoch itself, which the six charts post had wrong and now says. His method is different from the index: for each of seventeen benchmarks he sets score thresholds and times how long after the first closed model each threshold was crossed by an open one. The private set lags the public set by about four months, which he reads as the open models tuning harder on the benchmarks they can see. He also notes that several of his “private” benchmarks are semi-private at best, that exposure cuts both ways, and that his guess is the real gap is larger than ten months, not smaller.
The June post said the achievable sovereign win was operational: freeze open weights, host them under your own law, accept that the training was not yours. The data adds three qualifications.
The lag is seven months, and it is stable. A stack built on the best open weights today runs what the closed frontier was running in March. For most production work that is fine, and the staircase shows no sign of the open line stalling. It is a lag, not a ceiling.
The best open weights are Chinese, and have been for 673 days. The post on securing the open-source supply chain covers what that means in practice. The capability data says the alternative, American or European open weights, costs you another three to nine months, and that the European option costs the most. Mistral Large 4 narrows that from sixteen months to thirteen. It does not close it.
The licence is the variable nobody charts. Kimi K3 held the open record for a quarter under terms that excluded most of the firms who would want it. The strip above is green most of the time, and the exceptions are recent. Pin the licence of the weights you deploy as carefully as you pin the weights.
Every chart has a CSV link, and they all come out of one script over Epoch's published scores. The Epoch file itself is here as taken on 30 October 2026, so the derivation can be checked. ECI scores shift slightly each time Epoch refits the index, so the figures here will not match a later download to the decimal.