Compute multipliers explained: what Anthropic's steep curve shows, and what it doesn't

What's going on

Jacob Coxon was a pre-training researcher at OpenAI from 2023 to May 2026, then spent about four months at Anthropic. On November 13 he said a slide from Anthropic's IPO roadshow was a major reason he resigned. According to The Decoder, the slide shows the "pre-training compute multiplier" for code from July 2024 to October 2026 on a logarithmic axis. On an axis like that, each tick mark stands for a tenfold increase. A straight line means steady exponential growth; a curve bending upward means acceleration. That's exactly what the line does. Coxon reads it as evidence that Mythos wasn't a fluke and could now be retrained "with a fraction of the compute originally required." He's calling for a pause on frontier models next year.

So far, most of the attention has gone to interpreting the slide. We explain the metric behind it, and what it does and doesn't prove.

What a compute multiplier is

Dario Amodei explained the term publicly in January 2025 in his essay on DeepSeek and export controls. A compute multiplier is an improvement that lets you reach the same result with less compute. Amodei's example, fittingly, comes from coding: instead of $10 million, a model hits 40 percent on a coding task for just $5 million. That's a 2x multiplier. Typical ones, he said, are around 1.2x, sometimes 2x, and rarely 10x.

In the roadshow video, according to Simon Willison's transcript, Anthropic describes "total effective compute" as the product of physical compute capacity and every multiplier it has found. At 4x, 25 units are enough for the same model instead of 100 — or you can use 100 units to train a model that would once have needed 400. A quote from the video:

"Compute multipliers give us a competitive advantage that is hard to copy."

You don't measure this kind of number on the finished giant model. You train smaller models with the old and the new recipe and compare them at the same test scores. Then you extrapolate using a scaling law — a formula that predicts how performance rises with more compute. That tells you how much more compute the old recipe would need for the same result. Epoch AI did this publicly with 231 language models, using perplexity as the performance measure. It captures how well a model predicts text.

What public research says

According to Epoch's 2024 study, the compute needed for the same performance halves roughly every 8 months (range: 5 to 14 months). That works out to about 3x per year, with a 95 percent interval of 1.5x to 64x. Amodei estimated around 4x per year in 2025. In February 2026, Epoch's Anson Ho summarized that pre-training estimates typically fall between 3x and 20x per year.

The MIT paper by Gundlach et al. (arXiv 2511.21622) is much more cautious. The authors tested key innovations from 2012 to 2023 in ablations — experiments where you remove individual components. At small compute scales, these innovations explain less than 10x of gains, not the 22,000x from earlier estimates. The gap only closes once you account for the fact that some innovations pay off more and more with scale. Extrapolated to 2023 levels, about 72 percent of the gains (measured on a log scale) come from the switch from LSTMs to Transformers, and about 16 percent from the Chinchilla rule, i.e. a better ratio of model size to data. At very small budgets, total progress might only be around 20x. The authors also stress that such multipliers are "reference-dependent." Against LSTMs, the lead of mixture-of-experts models grows by about 63 percent a year; against standard Transformers, it stays constant at 2x.

For context, we calculated what overall factor the public rates would imply over the slide's time span (27 months):

This is exactly where the slide stops being useful. As Nathan Lambert points out, the y-axis has no numbers. You can see the curve getting steeper, but not where it falls in this table.

Why "for code" is a special case

Code can be checked automatically. Either the tests pass or they don't. That makes it easy to filter training data for code and to machine-verify synthetic examples written by a strong model. It's also the basis for reinforcement learning with verifiable rewards (RLVR) — training with rewards that a program hands out automatically. The slide does say pre-training. But synthetic data from RL-trained models can feed into pre-training.

That's why Adam Karvonen suspects terms like "compute multiplier" obscure the fact that most of the progress comes from data. Epoch's Anson Ho also thinks the term "algorithmic progress" may be misleading, pointing to distillation, where a large model teaches a smaller one. DistilBERT reached about 97 percent of BERT's capabilities with roughly 3 percent of the compute. In a comparison with the Nanochat training project, Steven Byrnes attributes a factor of about 8 to the better dataset alone.

Our take: a multiplier for code tells you how cheaply a model reaches certain coding results. It doesn't follow that training as a whole has become that much more efficient.

Three numbers that keep getting mixed up

According to Tokenpost, the revenue metric covers training and inference (running the models in production) as well as Anthropic's approach of using multiple chip families. It isn't labeled as pre-training efficiency. Nor has it risen steadily. In Q2 2026 it stood at 2.08x; in Q3 it dropped to 1.72x. According to the prospectus, that was due to $5 billion of short-term rented compute that "costs significantly more." Bret Kerr distinguishes five denominators for "cheaper": training capital, internal cost per token, customer price per token, cost per completed task, and revenue per compute dollar. They can't be converted into one another. He puts the lead over OpenAI at 1.5x to 3x depending on the denominator, though with very low confidence about the magnitude. Constellation Research called Anthropic's claim that a dollar of compute goes further at Anthropic than at any other frontier lab the "second most absurd" statement of the roadshow.

What this means for export controls

In 2025, Amodei argued that efficiency gains don't make chip bans obsolete. Labs reinvest the savings into even smarter models at the same high price. On ChinaTalk, Lennart Heim and Sihao Huang stressed that the road to fewer chips often runs through a great many chips first — for trial and error. On top of that, 60 to 80 percent of compute goes into running models.

In October 2026, the Institute for Progress modeled a US compute lead of about 16x. Without distillation, that would equal roughly 13 months of lead; with distillation, about 6.5 months. The model explicitly assumes US and Chinese labs are equally efficient, and calls this a "key limitation" itself.

Anthropic's roadshow claims the opposite: multipliers that are "hard to copy." Our take: if that's true, the lead lies in recipes, and protecting them becomes as important as the chips themselves. That fits with the fact that, according to Reuters (cited by Let's Data Science), Anthropic halted a pre-training run for more than a month to better protect against theft of model weights and intellectual property. If the multipliers spread, on the other hand, chips remain the main deciding factor. That Chinese labs are training on Nvidia's restricted Blackwell chips, according to a forensic analysis, shows how sought-after those chips remain.

Where "effective compute" becomes a safety threshold

Anthropic's Responsible Scaling Policy (RSP), the company's safety policy, has counted in effective compute since 2023. Version 2.0 from October 2024 required a comprehensive assessment after 4x more effective compute (excluding post-training) or after six months of post-training improvements. By our math, with 4 to 5x hardware growth per year (per Epoch) and roughly 3x from algorithms, the 4x mark is hit after about six months. The faster the multipliers grow, the sooner. Whether this rule still applies in the current version 3.4 isn't clear from our sources.

The threshold for AI-accelerated research used to be defined in terms of effective compute too. According to an analysis on LessWrong, Anthropic estimated the pace from 2018 to 2024 at about 35x per year: 4.2x hardware, 2.8x algorithms and roughly 3x post-training. The threshold was a doubling of that pace, i.e. "roughly 1000x" per year (mathematically 35² ≈ 1225). The author criticized that effective compute can't be measured directly. It's inferred from noisy test results. Since version 3.1, the threshold refers to a doubling of overall capability progress — two years of progress at the 2018–2024 pace in a single year. On November 10, Anthropic stated in the prospectus that this threshold has been reached. The reasoning cites expectations for future models and the general pace. Anthropic doesn't say whether the slide is part of that. Either way, a multiplier for code alone isn't the same as overall capability progress.

Coxon's pause and a contradiction

Coxon wants to pause frontier models for a year so competitors can catch up and human-level intelligence becomes a commodity. According to him, that would only hurt OpenAI and Anthropic. When he resigned in September, he already spoke, per TechCrunch, of "gambling with our lives" and of pace agreements between the labs. In June 2026, Anthropic itself proposed a verifiable international pause in its report "When AI Builds Itself" — but only if other labs follow suit.

The timing is an open question. Coxon resigned on September 8, the roadshow materials are dated November 10, and the curve runs through October 2026. Tokenpost points out that the date says nothing about when the slide was created. Which version he saw is unclear.

The bottom line

What's established: Anthropic is showing investors an internally estimated, accelerating multiplier for code. The revenue metric has risen 4.3x, and the threshold from the safety policy is considered reached. What's not established: the size of the multiplier, whether it transfers beyond code, and the "superexponential" trajectory. The missing axis values, the "for illustration only" footnote and the reference dependence shown by Gundlach et al. all argue against it. Coxon's line about a "fraction" would also hold true at the normal trend of about 3x per year. What matters is how small that fraction is.

Our take: the slide is a strong signal of what Anthropic itself believes, but it isn't a measurement yet. For policy, what matters is the contradiction it exposes. If the multipliers really are hard to copy, the equal-efficiency assumption in export control models loses its footing, and a pause would freeze the lead rather than shrink it. If they spread quickly, a pause would do exactly what Coxon wants.