The DeepSeek Moment Was the Peak, Not the Start

The popular reading of DeepSeek R1's January 2025 release was that open-weight models had finally begun a permanent catch-up with closed frontier labs. That reading was wrong. The data now show the opposite: DeepSeek R1 marked the point at which the gap was smallest, and it has widened since. This matters differently for capability forecasting, enterprise routing decisions, and closed-model valuations — and those three stories should not be collapsed into one.

The V-shape nobody named. Two LessWrong analyses published in May 2026 aggregate performance across 17 benchmarks, including 8 that are private and contamination-resistant, to track the open-weight versus closed-frontier gap over time.¹² The headline finding is clean: the gap narrowed into January 2025, then widened. Stanford's AI Index 2026 corroborates this independently via Arena Elo (a measure of user preference in head-to-head model comparisons), showing the closed-model lead growing from 0.5% to 3.4% between August 2024 and early 2026 — a 6.8x widening in roughly nine months.³

The specific gap estimates are: open-weight models are 4–6 months behind on public benchmarks and 8–10 months behind on private, contamination-resistant evaluations.¹ The divergence between those two numbers is itself informative, and I will come back to it.

The narrative that was wrong. The claim that circulated through 2025 was not merely "DeepSeek R1 is impressive." It was "open-weight development has hit an inflection; architectural replication is now fast enough that open models will converge to the closed frontier on a sustained basis." That is a different and stronger claim, and the data do not support it.

Release timing complicates this somewhat. Closed labs ship on irregular schedules; a single large release can spike the gap. The widening since January 2025 could partly reflect GPT-5-class and Gemini Ultra-successor releases creating a temporarily elevated frontier, rather than open-weight models stalling. This is a legitimate counterpoint. But the Stanford Arena Elo data covers a nine-month window and shows a directional trend, not a single spike — and the LessWrong finding that private-benchmark gaps are wider than public ones is consistent with structural divergence rather than timing noise.

The DeepSeek R1 moment was the peak of open-weight convergence, not the beginning of a trend. That distinction changes almost every downstream inference people made from it.

The measurement story is the most underreported part. The 4–6 month public lag versus 8–10 month private lag is not just a methodological footnote. It tells you something specific: contamination inflates open-model public benchmark scores more than closed-model public scores.

Contamination here means training data that includes benchmark questions or close variants — so a model can appear to perform better than it functionally does on tasks a user would actually run. The LessWrong analysis identifies two benchmarks (FrontierMath and ARC-AGI) where contamination most clearly inflates scores, and both inflate closed model scores, not open ones.¹ That is the opposite of what you would expect if closed labs were gaming their own private benchmarks. It complicates clean capability claims in both directions.

The practical implication: the true capability gap is probably narrower than the private-benchmark 8–10 month figure suggests, because some of those private benchmarks favour closed models in ways that may not reflect deployment reality. But it is almost certainly wider than the public-benchmark 4–6 month figure suggests, because contamination is inflating public open-model scores. The honest range is somewhere between those two numbers, and the directionality — closed models widening their lead since January 2025 — holds on both.

It is also worth flagging that private benchmarks are, by definition, not publicly verifiable. LessWrong posts are not peer-reviewed. I am treating the directional finding as credible because it is consistent with independent data (Stanford Arena Elo) and because the methodology described — using contamination-resistant evaluations to separate signal from noise — is sound in principle. But the specific month-gap estimates should be held loosely.

The capability story and the commercial story are not the same story. Here is where most coverage conflates two things that should be kept separate.

Even if the capability gap is widening, the price differential between open-weight and closed-frontier models remains extreme.

A 100x price differential means the commercial question is not "which model wins." It is "which tasks route to which tier." And that routing question does not require capability parity to resolve. A legal firm that runs 80% of its document review on a cheap open-weight model and reserves the remaining 20% of genuinely hard reasoning tasks for a closed frontier model is not waiting for convergence. It is already doing this.

The widening capability gap actually clarifies enterprise segmentation rather than collapsing it. If open-weight models were within a few percentage points of closed-frontier on everything, the routing decision would be genuinely difficult. The 100x price gap with a real (if contested) capability difference makes the segmentation legible: cost-route what you can, frontier-route what you must.

What this means for the "open-weight will commoditise closed pricing" thesis. The bear case on closed-model commercial positions relied on a specific mechanism: open-weight capability convergence creates a credible alternative, compressing closed-model pricing toward marginal inference cost. That mechanism requires the capability gap to narrow, not widen.

The data since January 2025 weaken that mechanism. They do not eliminate the commercial pressure from open-weight models — the 100x price differential is real and enterprises are using it — but they separate that pressure from the capability story. The commercial threat to closed models is cost-substitution for mid-tier tasks, which does not require capability parity and was always going to happen regardless of whether the capability gap converged. The specific claim that open-weight would eliminate the closed-frontier premium is the one the data are pushing back against.

I want to be careful about closed-model valuations here. The research file I was given included Anthropic and OpenAI valuation figures that appear speculative or projected rather than confirmed by primary sources at this run date. I am not going to assert those numbers. What I can say is that the "open-weight will compress closed-model pricing" narrative was one of the more serious structural challenges to any premium valuation of closed-frontier labs — and the evidence that the capability gap is widening rather than narrowing is relevant to that thesis, in the direction of strengthening the closed-lab position.

What the gap means for compute-versus-algorithmic-progress decomposition. This is the angle that matters most for forecasters, and it is the one that gets least attention in commercial coverage.

When DeepSeek R1 appeared to match or approach closed-frontier performance at substantially lower compute cost, the inference many drew was that algorithmic progress (better architectures, better training recipes, more efficient use of compute) was eating into the advantage closed labs derived from raw scale. If that were true, the gap between "what you can do with a lot of compute" and "what you can replicate with good algorithms and less compute" should be closing.

The data since January 2025 suggest the opposite is happening. Closed labs have continued to benefit from advantages that architectural replication does not capture. Those advantages could be raw compute scale, proprietary training data, post-training techniques that are not published, or some combination. The widening gap — especially on private benchmarks — is evidence that the compute-side advantage is not being dissolved by algorithmic catch-up at the rate the DeepSeek R1 moment implied.

This is directly relevant to AI capability forecasting models that use the open/closed gap as a proxy signal for how quickly algorithmic progress is compressing the value of scale. If the gap is widening, those models should update toward "scale advantage more durable than post-R1 estimates."

I hold this interpretation with appropriate uncertainty. The alternative reading — that the widening gap reflects release-timing effects rather than structural divergence — is plausible and I have not been able to rule it out from the available sources. But the nine-month Arena Elo trend is not a one-event artifact, and the contamination analysis pointing to private-benchmark gaps being wider than public ones is consistent with a structural rather than cyclical story.

What to watch. Three things determine whether the widening-gap trend continues or reverses.

First, whether the next generation of open-weight flagship releases (Llama 5-class, whatever Mistral and Qwen produce at scale) close the private-benchmark gap or leave it intact. If open-weight models post-2026 close to within 2–3 months on private benchmarks, the convergence thesis gets a second life.

Second, whether enterprise routing patterns start to reveal a revealed-preference signal. If high-stakes professional-services workflows (legal, medical, financial analysis) show meaningful adoption of open-weight models at the frontier rather than just in cost-routing roles, that is evidence the capability gap is less functionally significant than the benchmarks suggest.

Third, whether the contamination-resistance methodology in these LessWrong analyses gets picked up by more formal evaluation infrastructure. The divergence between public and private benchmark gaps is important enough that it deserves more than two community posts to establish. If Stanford, Scale AI's HELM (Holistic Evaluation of Language Models, a multi-task benchmark suite), or Epoch reproduce the finding with peer-reviewed methodology, it becomes load-bearing for forecasting. Right now it is credible but unconfirmed.

The comfortable 2025 narrative — DeepSeek proved open models are catching up permanently — deserves to be retired. The data show a V-shape, not a slope. That is a more interesting and more actionable picture than the convergence story it replaces.

Footnotes

¶