Break capability apart and only one area has actually converged

"They caught up" can be proven and disproven at the same time because capability was never one thing. A model can match its rivals at writing code while lagging generations behind elsewhere. Lay out every domain this analysis touches and the unevenness jumps out.

Coding happens to be the most visible domain: it has the most public leaderboards, and it's the capability everyone touches daily. The rows below it either require a dedicated evaluation lab to run or never made the self-reported list at all. The public "they caught up" narrative grows almost entirely out of that first row.

And coding is exactly the row that matters most

The best-measured trend in the field lives in software engineering: the length of task an AI can complete on its own roughly doubles every 4 months. The author notes this likely holds in other domains such as math too — software engineering is just where the measurement is most thorough. That "task length" has a proper name: time horizon.

Extend that trend two or three years and you land at models taking on work that would occupy a person for months or years. Even if the trend only holds for coding, whoever first controls a superhuman-programmer model controls a strategic position that's very hard to erase. That's why it's worth doing the arithmetic on those few months of gap.

AI Futures Model (December 2025 build, parameters updated 2026-04-02). The two dates at bottom left are what the simulation produces at median parameters: automated programmer in June 2028, superintelligence in May 2029. In the key parameters column, the coding time-horizon doubling period is set to 4 months. This is a forecasting model's output, not measured data. Source: AI Futures Project

The comparison covers only the US and China because that's where the best models, the most AI research, and the most compute all sit. Europe has the money and the people in principle, but so far has shown no real intent to compete.

Which raises the question: if the per-domain gaps are this obvious, how does it all come out as "parity" on the leaderboards? Two more layers explain it. The first is the word "gap" itself — it's quietly doing two jobs, and most people only use one of them.

The word "gap" is doing two jobs at once

The definition fits in one sentence: how long it takes one lab to build a capability another lab already has. But that sentence contains two questions pointing in opposite directions.

Illustrative diagram, not real data. The 6.3 months and 10.0 months shown are made-up figures drawn to explain the concept: the left span measures when the leader hit the trailer's current score, the right span measures when the trailer will hit the leader's current score. Source: scaling01

One more layer gets skipped a lot: capability is domain-specific. A model can match another at writing code and still lose elsewhere, or lose on aggregate. So "caught up" always demands a follow-up — caught up at what?

Models ship in batches, so there's no continuous curve

Try to compute any of this and you hit a practical problem: progress is discrete. Models appear every few weeks to months, and two models never score exactly the same, so there's no way to measure the interval directly. Three approaches, each with a trade-off.

The author picked the third: it's symmetric, it better reflects the underlying trend, and it reveals whether the gap is widening or narrowing. The cost is stated plainly too — the result depends on which regression model you choose, and it smooths away discontinuities in the trend. That works against Kimi K3: if K3 genuinely is a break in the trend, the regression will push it back down.

That said, over roughly the past five years only one or two events qualify as trend breaks: reasoning models in Q3 2024, and possibly the surge in agentic models after Claude Opus 4.5 shipped in November 2025. Neither had anything to do with model size, so the idea that a model less than twice the size of DeepSeek-V4-Pro breaks the trend on scale alone doesn't hold up well.

Every conclusion below therefore comes with two numbers: an adjacent-model range (the instantaneous gap right now, with an interpolated value inside the range on the charts) and a trend-line estimate (the long-run picture). With the method settled, we reach the second layer of the disagreement — the one that matters most: which questions you use to measure.

Which benchmarks make the list decides the conclusion

"Won 30 of 35" carries exactly as much weight as the selection process behind those 35. Break the composition down and the missing categories say more than the present ones.

Vision and agentic together account for 24 benchmarks — nearly seventy percent of the list. The coding group does contain a few the author considers high signal, such as ProgramBench and SWE-Marathon, which cover broader task scopes and longer horizons.

Third-party leaderboards carry the same skew

What about the third-party Artificial Analysis Index? It's one of the most popular indices today, but the author argues it shouldn't be used to judge a model's overall strength, for two reasons.

First, the weighting. The index composition tilts heavily toward agents and short-horizon coding.

The index's four category weights: Agents 34%, Coding 24%, General 18%, Scientific Reasoning 24% — the first two summing to 58%. The largest single component, GDPval-AA v2, is a 220-task agent evaluation. Source: Artificial Analysis

Second, difficulty. The index does nothing with how hard each evaluation is or how much information a given score carries. Everything gets averaged under hand-set weights, so a benchmark anyone can pass and a benchmark almost no model can solve count exactly the same.

That doesn't make it useless. Its rankings match how models feel in daily use, because it measures work with genuine economic value. The author's fix is to relabel it: treat it as a measure of everyday usefulness rather than general intelligence. That's why his piece calls it the Artificial Analysis Index, with the word Intelligence dropped from the official name. So is there a yardstick that skips hand-picked weights and factors in difficulty? There is — and Anthropic uses it too.

How ECI folds question difficulty into the score

Epoch AI's Capability Index (ECI) targets exactly those two flaws. It aims to measure the underlying general capability of a model and place every model on one shared scale. The technique is called Item Response Theory (IRT).

Applied to evaluations, each benchmark gets two parameters:

With those parameters estimated, a single unified statistical model gets fitted to find the capability value that best explains all observed scores. That value is the model's ECI.

Anthropic recently adopted the same approach for measuring model strength, calling it AECI in its own system cards. Their chart also plots benchmark difficulty along the way.

Working through one cell shows what the method is doing. GPQA Diamond's 50-point mark lands near 129, meaning a model with capability around 129 should score roughly 50 there. Claude 3 Opus has an AECI of about 126 — slightly below 129 — so the model predicts it scores a bit under 50. It actually scored 50.4, a touch above expectation.

AECI chart drawn from Anthropic system card data. Red dots are the capability values of successive Claude models; colored boxes are difficulty ranges per benchmark (each box spans the capability needed for 25% to 75% scores, with the center line at the 50% mark). Left to right, GPQA Diamond's line sits lowest and GSO-Bench highest, in ascending difficulty. Source: Mythos Preview system card

Capability values on one shared scale, with less bias than a hand-weighted index — that's what ECI buys you. With both yardsticks in hand, it's time to measure the same set of models with each.

Two yardsticks, same models, a gap that differs by more than 3x

Same models, same period. Change the measuring stick and both the size and the direction of the US-China gap change with it.

First, the AAI yardstick

Kimi K3 scores 57.1 on this index. Because the index is bounded between 0 and 100, the author fits each country's frontier trend with a sigmoid curve.

Frontier trends fitted on AAI. Info box at bottom right: adjacent-model range 37 to 49 days (1.22 to 1.61 months), local interpolation 45 days (1.48 months), trend-line basis 88 days (2.89 months), US-China crossover 2027-11-09. Source: scaling01

By this stick, the Chinese frontier trails the US frontier by a bit over a month, and extending the trend line puts the overtake on November 9, 2027.

Switch to ECI and the same models flip the conclusion

Once Epoch published preliminary scores for Kimi K3, combining them with Moonshot's published results gives 9 benchmarks — enough to fit its ECI directly: 155.53, with a 90% interval of 153.87 to 158.21. That figure is the author's own calculation, not an officially published Epoch ECI. It lands between GPT-5.3 Codex and GPT-5.4 Pro, closer to the former.

Frontier trends fitted on ECI. Info box at bottom right: adjacent-model range 133 to 161 days (4.37 to 5.29 months), local interpolation 160 days (5.27 months), trend-line basis 185 days (6.08 months), and the US-China crossover field reads none. Annual progress rates: US 14.6, China 11.0. Source: scaling01

Drop to the lab level and the ranking nearly inverts

Looking only at national frontiers hides the pace of individual labs, so both charts have per-lab versions. Put the annual progress rates side by side and you can see just how far apart the two methods land.

Moonshot falls from first on the left to below OpenAI on the right; Anthropic climbs from fifth to a commanding first. Only Google holds its position among the seven: 21.4 and 6.3, last on both.

On the AAI per-lab trend, Moonshot has the steepest slope and crosses the US frontier line on February 8, 2027 — the original text describes this as passing Anthropic. On the ECI per-lab trend, Moonshot's crossover field also reads none, and even with Mythos Preview removed from Anthropic's data, Anthropic's rate stays at 18.1 points per year, still the fastest.

The two per-lab trend charts, and how Kimi K3's ECI was estimated

AAI per-lab trends, with annual rates for all seven labs and Moonshot's crossover date of 2027-02-08 at bottom right. Source: scaling01

ECI per-lab trends. Same labs, and a rate ranking almost exactly opposite to the chart above. Source: scaling01

Before Epoch's scores came out, the author ran his own estimate. The community poll put Kimi K3's ECI at 158 to 159; he guessed 156 to 157; analyst Teortaxes guessed 157 to 158.

To avoid relying on intuition, he used the high correlation between AAI and ECI to convert between them. A straight line doesn't work: AAI is capped between 0 and 100, so scores bunch up near both ends, while ECI is unbounded and linear. The result is a line that overestimates at the weak end around Claude 2 and Claude 3, underestimates through the middle, and overestimates again among strong models.

The fix is to apply a logit transform to AAI first — the inverse of a sigmoid — which pulls the compressed ends back apart before fitting.

The fit after the logit transform, with mean absolute error down 37.2% from a straight linear fit. Source: scaling01

The conversion produced an estimate of 158.33, with a bootstrap 80% interval of 154.49 to 162.02. Epoch's scores then arrived, and fitting 9 benchmarks directly gave 155.53 — inside the interval, but toward the low end. The rest of the analysis uses 155.53; the conversion formula stays on the shelf for later.

Mythos Preview needed an estimate too. Anthropic only publishes AECI on its own scale, so conversion is required: the Substack Point Estimate computed 161.5 using the ECI method, Ramez Naam got 161 via least squares, and the author also landed on 161, 90% interval 158 to 166. That's close to Claude Fable 5's actual ECI of 160 (interval 158 to 165), which makes sense if Fable 5 is a smaller model distilled from Mythos Preview with extra safety training — both of which would nudge the score down a little.

The forward gap: how long to reach today's best

Flip to the forward question and the answer runs like this. On the national trend line, Chinese models reach Mythos Preview's 161 on March 7, 2027 — about 11 months out. The 90% confidence band spans December 28, 2026 to June 18, 2027, which converts to a lag of 8.72 to 14.38 months.

On the hardest questions, the gap widens

One result in Epoch's preliminary numbers is especially telling. On FrontierMath Tier 4, a set of research-grade math problems, Kimi K3 scores about 39 — still behind GPT-5.2 Pro (around 45), released seven months earlier. On the same chart, Claude Fable 5 scores 88.

FrontierMath Tier 4 (v2), Epoch AI internal scores across 40 models, topping out at 88%. The green dot in the red box is Kimi K3 at about 39; the purple dot above is Claude Fable 5 (max) at about 88. Source: Epoch AI

Five things this calculation leaves out

The author's own read is that both gap figures run small, because several high-impact factors never enter the formula at all — usually because the numbers aren't public.

One: model size. Kimi K3 is 2.8 trillion parameters. Cursor's Michael Truell has said the Opus and GPT families are comparable in size to Cursor's newly released 1.5-trillion model, most likely referring to Claude Opus 4.8 and GPT-5.4; GPT-5.5 and GPT-5.6 are estimated closer to 3 trillion. So a 2.8-trillion model beating a 1.5-trillion-class Claude Opus 4.8 on some benchmarks doesn't say much architecturally.

Two: inference efficiency. None of the 35 benchmarks measures it, and by the author's read Kimi K3 remains less efficient than US models. Burning more tokens at the same parameter class to reach the same score is a cost no leaderboard records.

Three: real serving cost and total compute. Whether you hold a million of the newest Blackwells or a pile of lower-throughput older Hoppers determines how many instances you can run at once and how fast you can serve users. In a scenario where both countries compete at full tilt, this weighs no less than benchmark scores.

Four: time spent on safety testing. US labs test far more thoroughly than Chinese labs, and that work may add another month to the gap.

Five: how representative the evaluations are. Most benchmarks don't reflect real work, and cost and time limits mean the scoring runs never extract a model's peak performance either.

An understated gap doesn't mean Chinese labs aren't genuinely speeding up, though. They are — unusually so.

And one more layer: US labs are sitting on unreleased models

Mythos is rumored to be a 10-trillion-parameter model. SpaceX AI, OpenAI and Meta are all building their own 10-trillion-class models. The author's view is that Kimi K3's apparent parity is likely Chinese models catching the previous generation while US labs have already moved to the next one — holding their strongest models back because of legal uncertainty around cyber and bio-chem capabilities.

If Anthropic and OpenAI actually shipped Claude Fable 5.1 and GPT-6, there would be no catch-up discussion at all, because these 10-trillion-parameter models are something else entirely.Lisan al Gaib, scaling01

Why Chinese models suddenly got faster last August

Chinese labs really have accelerated, and the starting point is specific. The author tries to locate the inflection by taking the derivative of the index with respect to time, while acknowledging it's not a great estimator.

Instantaneous progress rate (AA index points per year). The two vertical lines mark overtakes: Moonshot passes Anthropic from 2025-08-06, GLM from 2025-10-19. At peak, Moonshot hits about 39.8 and Z AI about 36.5, versus roughly 31 for Anthropic and 30.3 for OpenAI. Source: scaling01

The pattern itself is strange: US labs have more compute, more data, more talent, and stronger models to accelerate their own research — yet they're progressing more slowly. The author first checked a technical explanation. Could it just be that the index is capped, so the US, already scoring high, has a harder climb and only looks slower? He applied the logit transform to strip that compression out, and Moonshot and Z AI still peak higher.

Distillation is part of it, but not all of it

Distillation in the strict sense requires access to the model itself, but using a stronger model to grade a weaker model's outputs already counts as a version of it. Anthropic has repeatedly reported tracing distillation attacks back to Chinese labs. So it's real; outsiders just don't know the scale or how much it helped. The author's phrasing: some portion of the capability Chinese models have gained comes from distillation.

Beyond that, several other explanations exist, and the author's view is that reality is a combination of them:

One caveat has to travel with all of this: the catch-up narrative holds on AAI only. It does not hold on ECI.

Counter-evidence: the harder and longer the task, the wider the gap

The UK AI Security Institute's cyber range is among the hardest evaluations around — the tasks are brutal, and it takes 100 million tokens to push a model to its limit. Epoch's MirrorCode goes further still, burning a billion tokens on a single model.

In this 32-step enterprise network attack simulation, GLM-5.2 averages about step 11, on par with Claude Opus 4.5. GPT-5.6 Sol reaches about 29, and Anthropic's newer Claude Mythos 5 about 27. Yet on the self-reported coding benchmarks, GLM-5.2 scores close to Claude Opus 4.8.

UK AISI's "The Last Ones" cyber range, a 32-step enterprise network attack simulation. The horizontal axis is cumulative tokens consumed (log scale); each model runs 10 times at 100 million tokens per run. GLM-5.2 (green) and Opus 4.5 both stall around step 11; DeepSeek-V4-Pro (blue) reaches about 8. Source: UK AI Security Institute

The same model scores near the latest Claude Opus 4.8 on public, easily-targeted coding leaderboards, then falls back to the seven-or-eight-month-old Claude Opus 4.5 on hard tasks requiring sustained multi-step work. That drop is itself a signal.

The piece closes on an observation: if Chinese models were genuinely as strong as the US frontier, they ought to be cracking some previously unsolved math and physics problems. So far, nobody has seen it.

Run the same models over the same period through two evaluation systems and it's not just the numbers that differ — the direction flips too.

Even the acceleration story inverts. On AAI, all four Chinese labs post faster annual progress than the three US labs, with Moonshot leading everyone at 39.4. On ECI, Anthropic leads by a mile at 25.0 while Moonshot drops to 13.8, behind OpenAI's 14.3.

The method is borrowed from educational testing and called Item Response Theory. In a sentence: two students both score 60, one on a grade-school test and one on an olympiad paper, and those describe wildly different ability — so to compare them, you first need to know how hard each paper was.

"Three times bigger" is abstract; days are not. Same question — how many days ago did the US frontier sit where the Chinese frontier sits today — and here's what each yardstick says: