Verse III · Evidence

Evidence

A thesis this blunt deserves a test, not just an argument. The cleanest one running right now is not a small model against a large one. It is the open-weight Chinese labs against the Western frontier: less capital, chips a generation behind and deliberately throttled by export controls, and architectures built for efficiency rather than scale for its own sake, measured against Anthropic, OpenAI, and Google on the same public benchmarks. As of June 2026, that test is not going the way the trillion-parameter story predicts it should.

The numbers

On BenchLM's June 2026 composite ranking, DeepSeek V4 Pro tops the Chinese field at 87, against Gemini 3.1 Pro at 93, GPT-5.4 Pro at 92, and Claude Opus 4.6 at 88. A six-point gap to the best closed model and a one-point gap to the third-best is not parity, but it is a long way from the gap a model trained on restricted hardware for a fraction of the budget was supposed to leave.

On SWE-bench Verified, the benchmark the field treats as the closest thing to a real coding exam, DeepSeek V4 Pro Max scored 80.6 percent in June 2026, tied exactly with Gemini 3.1 Pro's 80.6 percent, with Kimi K2.6 close behind at 80.2 percent and Qwen3.7 Max at 80.4 percent. Claude Opus 4.8 and GPT-5.5 still lead that benchmark outright, at 88.6 and 88.7 percent. But a Chinese open-weight model tying one of the three labs this manifesto names as the frontier, on the benchmark that matters most for the work this manifesto cares about, is not a footnote. On vendor and third-party tracked SWE-bench Pro numbers, GLM-5.2 has been reported at 62.1 percent, ahead of GPT-5.5's own reported 58.6 percent. Code Arena's agentic web-development leaderboard, tracking 67 models as of April 2026, places GLM-5.1 third in the world and Kimi K2.6 sixth.

The price

None of that competitiveness costs what frontier inference costs. DeepSeek V4 Flash prices at fourteen cents per million input tokens and twenty-eight cents per million output tokens. Claude Opus 4.8 prices at five dollars and twenty-five dollars for the same. GPT-5.5 runs five dollars and thirty dollars. Gemini 3.1 Pro runs two dollars and twelve dollars. A model tying or trailing the frontier by single digits on the hardest available coding benchmark, at roughly a thirtieth to a hundredth of the price per token, is the unit economics this manifesto's case for small and efficient models rests on, demonstrated at a scale no single engineer's benchmark could prove on its own.

The capital, and the hardware it had to work with

The headline number attached to DeepSeek's rise was 5.6 million dollars, the GPU cost of the V3 pretraining run, plus 294,000 dollars for the R1 reinforcement learning pass. That figure is true and also misleading, in the way a single line item is misleading when presented as a whole budget. A more careful accounting puts DeepSeek's total server capital expenditure near 1.6 billion dollars and its operating costs near 944 million dollars, with hardware spending alone exceeding 500 million dollars over the company's history once research, failed runs, and infrastructure are counted rather than just the final successful training run. That correction matters, and it cuts both ways. It is intellectually dishonest to wave away DeepSeek's cost advantage as a marketing number. It is just as dishonest to repeat the marketing number as the whole truth. The honest comparison is harder and more favorable to this manifesto's case than either: a company that has spent on the order of a few billion dollars in total, competing against labs that have raised, between them, tens of billions of dollars in venture capital to reach a similar place on the scoreboard.

That capital gap exists despite a hardware gap working against the Chinese labs, not for them. DeepSeek trained V3 on Nvidia H800 chips, a version of the H100 deliberately degraded to comply with US export controls, and has been reported using restricted H100s and H20s alongside them because the legal supply was not enough on its own. DeepSeek's own CEO, Liang Wenfeng, has said publicly that Chinese labs need two to four times the computing power to match what an unrestricted lab can do with modern Nvidia hardware. Huawei's Ascend chips, the domestic alternative Beijing is now pushing with a 295 billion dollar plan to build a computing grid that excludes Nvidia almost entirely, deliver only around sixty percent of an H100's performance for inference, and by DeepSeek's own evaluation are not yet good enough to train a frontier-class model on at all. Older chips, fewer of them, a fraction of the legal supply, and a domestic replacement that is not ready. That is the hardware budget behind a model now tying a Google frontier release on the field's hardest public coding benchmark.

The objective with no terminal point

Every number in this chapter so far explains a price difference. This section explains why one side of that comparison keeps paying it. A small model trained for a defined task has an external target: clear a benchmark score, hit a latency budget, handle a domain at an acceptable error rate. Once that target is cleared, more training is measurable waste, and a lab can point at the number and say spending further here no longer pays. Artificial general intelligence has no equivalent target, because general intelligence has never been given an operational definition. There is no benchmark a frontier lab clears that ends the project, because the field invents a harder benchmark the moment the old one stops discriminating between models. Humanity's Last Exam exists for exactly this reason. It was built in 2025 because models had pushed past ninety percent on the benchmarks that came before it, MMLU among them, and a saturated benchmark stops being able to tell anyone anything. The new exam will saturate too. Another one will follow it. That is not a flaw in the benchmarking process. It is what an open-ended objective looks like from the outside.

This manifesto has already made the economic version of this argument twice, without naming the mechanism behind it. The introduction calls scale-chasing a bill that arrives every quarter and never comes due. The definition of a frontier model earlier in this manifesto calls it a moving target by definition, never a fixed product. Both observations describe the same fact from different angles. An objective with no stopping rule produces a capital commitment with no stopping rule. A lab chasing a fixed target can be defunded the moment the target is reached, or the moment it becomes clear the target will not be reached on the money already spent. A lab chasing general intelligence has no such moment available to it, because there is no fixed target whose achievement, or failure, a board or an investor could point to and call it settled.

The honest counter to this argument is that a hundred small models with a hundred bounded objectives are not obviously different from one large model with none, since in aggregate the spending might look similar from a distance. It is different, and the difference is the whole point. Each of those hundred bounded efforts can be judged, funded, and killed on its own terms, the moment it proves out or fails. None of that capital is hostage to a definition that does not exist. The trillion-parameter program has no equivalent off-ramp. It can only ever be funded again, on the promise that the next run is the one that finally arrives.

What the trillion-parameter frontier actually spent

The other half of this comparison deserves the same rigor applied to it. OpenAI has never published the architecture or parameter count of GPT-4. Its own GPT-4 Technical Report says outright that it withholds "further details about the architecture (including model size)... due to the competitive landscape and the safety implications." What follows is leak and informed estimate, not OpenAI disclosure, and it is presented that way.

A widely cited SemiAnalysis report from July 2023 described GPT-4 as a mixture-of-experts model exceeding one trillion parameters in total, commonly cited at 1.8 trillion across 16 experts of roughly 111 billion each, with around 220 billion of those parameters active on any single forward pass, trained on roughly 13 trillion tokens. OpenAI has confirmed none of this, but the figures are detailed enough, and corroborated widely enough since, that this manifesto treats them as the working estimate.

That training run is the one with public numbers attached to it, and a rough check holds up against them. Counting only the 220 billion active parameters as the ones doing work on each token, the standard compute estimate is six times parameters times training tokens.

That puts the run at roughly 1.7 times ten to the twenty-fifth floating point operations. Spread across an A100 cluster running at a realistic 35 percent of peak throughput, that works out to roughly 44 million GPU-hours, close enough to the figure reported elsewhere, 25,000 GPUs run for 90 to 100 days, around 55 to 57 million GPU-hours, to treat both as the same neighborhood. At a cluster power draw near 25 megawatts sustained for that stretch, the run consumed somewhere around 50 to 60 gigawatt-hours of electricity, enough to power several thousand homes for a year, for one training run of one model. At roughly one to two dollars an hour for that much A100 time, the run cost on the order of 60 to 110 million dollars. Sam Altman has himself acknowledged, in public remarks the same month reporters first floated the 100 million dollar figure, that the true number was higher, not lower, and that he already considered the era of simply building bigger models to be ending.

That is the spending this manifesto is arguing against, confirmed close enough by the founder's own account, for one model, one training run, before a single dollar of inference, staffing, or the failed runs that never made it into a press release. Measured against that, a team spending a few billion dollars in total, training on chips a generation behind, and tying the result on a real public benchmark, is not the underdog story it is usually told as. It is the more disciplined one.

GPT-4 against DeepSeek V3, run cost to run cost

Set the two training runs already detailed in this chapter side by side, on the same accounting basis, and the point lands harder than either figure does alone. DeepSeek's own technical report puts V3 at 671 billion total parameters, 37 billion of them active per token, trained on 14.8 trillion tokens using 2.788 million GPU-hours on the export-compliant H800 chips described above. The R1 reinforcement learning pass added 294,000 dollars on top of V3's 5.576 million dollar pretraining run, putting the full disclosed cost at roughly 5.87 million dollars. GPT-4's leaked figures describe a model with more than twice the total parameters, roughly six times the active parameters, trained on a comparable volume of tokens, using around twenty times the GPU-hours, at an estimated cost on the order of 100 million dollars.

Both numbers describe the same thing: dollars of GPU time spent to produce one finished, trained model. Neither side here is a marketing estimate dressed up as fact. DeepSeek's figure comes from its own published technical report. GPT-4's comes from a leaked architecture analysis, checked earlier in this chapter against the number OpenAI's own chief executive declined to dispute. On that basis, training a model that would go on to tie a current Google frontier release on SWE-bench Verified cost roughly seventeen times less than training the model OpenAI has never confirmed the size of. That is not a rounding difference. It is this manifesto's whole argument, expressed as one number divided by another.

None of this is abstract to me. I write code on Claude Code most days, and I also keep a Qwen coding model running locally on an Nvidia Spark on my desk. The local model takes the routine work, scaffolding, mechanical refactors, the parts of the job that do not need a frontier model's judgment, and every task it handles is a task Claude Code never bills me for. The savings are not a thought experiment. They show up on my own invoice.

None of this proves small models win every argument. It proves the bigger one: that capital and chip count were never the only path to competence, and a team with less of both has already shown, in public, on benchmarks anyone can check, how much further efficiency can carry a model than the frontier labs' own roadmaps assume.

Sources

1. BenchLM. "Best Chinese LLMs in 2026: DeepSeek V4, Kimi K2.6, GLM-5, Qwen, and Every Model Ranked." June 2026. benchlm.ai. The composite benchmark ranking placing DeepSeek V4 Pro, GLM-5.1, Kimi K2.6, and Qwen3.5 against Gemini, GPT, and Claude.
2. Morph. "Best AI Model for Coding (June 2026): 12 Models Ranked by SWE-bench Pro Score and Cost per Task." June 2026. morphllm.com. The SWE-bench Verified and SWE-bench Pro scores, and the per-token pricing, cited above.
3. SemiAnalysis. "DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts." 2025. newsletter.semianalysis.com. The fuller accounting of DeepSeek's total capital expenditure and operating costs behind the headline training-run figure.
4. Center for Strategic and International Studies. "DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race." 2025. csis.org. The chip restrictions DeepSeek trained under, Liang Wenfeng's public comments on the compute penalty, and Huawei Ascend's current limits for training.
5. OpenAI. "GPT-4 Technical Report." 2023. arXiv:2303.08774. States directly that architecture, model size, hardware, and training compute are withheld from the report.
6. Patel, D. and Wong, G. "GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE." SemiAnalysis, July 2023. newsletter.semianalysis.com. The original leak behind the widely cited 1.8 trillion parameter, mixture-of-experts estimate for the original GPT-4.
7. Fortune. "OpenAI's Sam Altman says giant A.I. models not necessary." April 2023. fortune.com. Reports Altman's remarks on GPT-4's training cost and the end of scale-first model design, from an MIT event the same month.
8. DeepSeek-AI. "DeepSeek-V3 Technical Report." December 2024. arXiv:2412.19437. The company's own disclosure of V3's 671 billion total and 37 billion active parameters, its 14.8 trillion training tokens, and its 2.788 million H800 GPU-hours.
9. Phan, L. et al. "Humanity's Last Exam." Center for AI Safety and Scale AI, January 2025. arXiv:2501.14249. Built after MMLU and similar benchmarks saturated above ninety percent accuracy, the example this chapter uses for benchmark goalposts that move as soon as they are cleared.