Former Anthropic Researcher Says the Company Could Retrain Mythos on a Fraction of the Compute
Former researcher Jacob Coxon has claimed that Anthropic may already be able to retrain its Mythos model on a small fraction of the compute spent the first time. His basis is a slide from the company's IPO presentation, which shows how training efficiency for code improved from July 2024 to October 2026. Anthropic describes the data points as internal estimates shown for illustration, so this is an opinion, not a fact. Plan for rapid cost declines in models at today's capability level.
AI-processed from The Decoder; edited by Hamidun News
Training a large language model costs hundreds of millions of dollars in electricity and GPUs, and how fast that cost falls shapes both API pricing and the balance of power between labs. Now that pace is being argued about in public. Jacob Coxon, a former researcher at OpenAI and Anthropic, has stated that Anthropic could already retrain its Mythos model using only a fraction of the compute it needed the first time. He backs this up with a slide from Anthropic's investor presentation ahead of its IPO.
What the slide shows
The chart plots a "pretraining compute multiplier" for code. Put simply, it answers the question: how many times less compute is needed today to reach a model of the same level as before. The period covers July 2024 to October 2026.
The value axis is logarithmic. That compresses the scale: a 10x gain and a 100x gain look almost identical on it. Even so, the line climbs steeply. This is what Coxon considers important. If a curve rises sharply on a scale that compresses magnitude, the real growth is very large and accelerating. He calls this regime "superexponential": each successive improvement arrives faster than the last.
"This is a crazy chart, and I'm surprised they published it," he wrote on X. By his account, this slide was one of the main reasons he left Anthropic. He worked there for about four months.
Anthropic's caption says the points are based on internal estimates and are included for illustration only. The source gives no exact values for the multiplier, so the scale of the improvements cannot be verified from the outside.
What this means for Mythos
Mythos is the model that delivered a notable jump in capabilities this spring. Many saw it as a one-off breakthrough. Coxon sees it differently: in his view, Mythos sits on the same curve as previous improvements and does not deviate from it.
An analogy: you spend six months building a complex service, stumbling over pitfalls and trying different architectures. A year later, you can rebuild it in a few weeks because you know which decisions work. If the same thing happens with every model generation, a level of capability that required a huge cluster yesterday will cost many times less tomorrow. Only the frontier will remain expensive.
An honest caveat applies here. These are the words of one person, a former employee at that, and they rest on a chart labeled "for illustration." There is no independent confirmation that Mythos can actually be retrained at a fraction of the cost.
Why Coxon wants a pause
He draws a radical conclusion from his assessment. Algorithmic improvements, he says, are "on the verge of automation": AI will start accelerating AI research itself. That is why he wants development of frontier models suspended next year, giving competitors time to catch up with the leaders.
His logic runs like this: a pause would "turn human-level intelligence into a commodity" and would "only hurt OpenAI and Anthropic." If others reach the same level, the two leaders' advantage in the segment of the most expensive models would disappear. This is an argument against the common view that slowing down benefits the labs themselves.
Coxon drew wide attention in September with public warnings about the risks of advanced AI, and other industry figures have repeated them. One of the main concerns is self-improving AI that could slip out of control.
What developers and founders should do
First, build rapid declines in the cost of current-level models into your product's financial model. If training efficiency grows the way the slide suggests, today's flagship capabilities will soon become mass-market.
Second, do not build a business on exclusive access to a single model. An advantage that depends on a model being unique disappears quickly.
Third, treat claims like this as a signal, not as proof. The chart is public, but Anthropic's internal data is not. Until the company discloses its figures, the real pace of progress can only be inferred from indirect signs: API prices and the cadence of new model releases.
- Nov 21, 2026Anthropic mathematician says Claude helped prove the Riemann hypothesis in six weeks
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
- Nov 21, 2026Chinese Tech Giants Double AI Spending to $139B, While Labs Recover Less Than 8%
- Nov 21, 2026Ego lite Gives Each AI Agent Its Own Workspace Space in the Browser
Key Questions Answered
What the slide shows?
The chart plots a "pretraining compute multiplier" for code. Put simply, it answers the question: how many times less compute is needed today to reach a model of the same level as before. The period covers July 2024 to October 2026. The value axis is logarithmic. That compresses the scale: a 10x gain and a 100x gain look almost identical on it. Even so, the line climbs steeply. This is what Coxon considers important.
What this means for Mythos?
Mythos is the model that delivered a notable jump in capabilities this spring. Many saw it as a one-off breakthrough. Coxon sees it differently: in his view, Mythos sits on the same curve as previous improvements and does not deviate from it. An analogy: you spend six months building a complex service, stumbling over pitfalls and trying different architectures. A year later, you can rebuild it in a few weeks because you know which decisions work.
Why Coxon wants a pause?
He draws a radical conclusion from his assessment. Algorithmic improvements, he says, are "on the verge of automation": AI will start accelerating AI research itself. That is why he wants development of frontier models suspended next year, giving competitors time to catch up with the leaders.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.