Jacob Coxon and Anthropic's IPO Compute Slide: Superexponential Claim vs S-1

Jacob Coxon, who left Anthropic in September 2026 over existential AI risk, is now pointing at a different artifact: a pretraining compute-efficiency chart from Anthropic's IPO roadshow that he says shows superexponential gains. The claim spread on X in November 2026 with commentary that Mythos-class models could be retrained with a fraction of prior compute. Anthropic's own S-1 registration statement and public roadshow excerpts describe stacking compute multipliers and closing the loop on R&D, not a simple headline about runaway superexponential training.

This post maps what is verified in Anthropic filings and roadshow remarks, what Coxon and allies assert about the chart, and how that debate connects to METR's acceleration assessment — without treating either side's extrapolation as settled science.

TL;DR

Coxon's new emphasis: investor charts, not only philosophy

In September, Coxon resigned with a viral thread arguing top labs race toward self-improving superintelligence and "gamble with our lives." explainx.ai documented that exit in Anthropic Researcher Jacob Coxon Resigns Over AI Safety Fears and a fact-check of surrounding theories.

November discussion, summarized on social aggregators and amplified by researchers such as Nathan Calvin, adds a concrete trigger: a logarithmic chart of pretraining compute efficiency shown during Anthropic's IPO roadshow. Coxon reportedly argues the curve is "superexponential" — efficiency improvements themselves accelerating — and that Anthropic could already retrain Mythos with far less compute than the original run required.

We did not obtain the exact slide image from Anthropic. The claim rests on Coxon's characterization plus social commentary. Treat the existence of some efficiency chart as plausible during an IPO; treat superexponential as Coxon's interpretation until Anthropic publishes the figure with methodology.

What Anthropic's IPO documents actually disclose

Anthropic's S-1 "Compute Efficiency" section defines three pillars:

1. Training efficiency — researchers innovate compute multipliers, increasingly with Claude assisting ("closing the loop").
2. Inference efficiency — multipliers that cut compute to serve a given quality level.
3. Multi-chip approach — Trainium, TPU, NVIDIA, and AMD behind a proprietary orchestration layer.

The filing states multipliers compound across model generations and that Claude's understanding of chip architectures shrinks onboarding time for new hardware.

Simon Willison's notes from the roadshow video quote leadership explaining total effective compute as physical capacity times discovered multipliers. A 4x multiplier means training the same intelligence with one-quarter the physical units, or spending the same units on a smarter model. Leadership adds multipliers stack and may accelerate further as Claude discovers them with less human intervention — language that rhymes with Coxon's fear but is framed as competitive advantage, not catastrophe.

None of that is a peer-reviewed proof of superexponential scaling. It is forward-looking corporate narrative tied to undisclosed internal measurements.

Growth curve metaphor for Jacob Coxon Anthropic IPO compute efficiency debate

How METR fits the same week

On November 5, 2026, METR published its independent assessment of AI-assisted R&D at Anthropic. Headline findings:

• Benchmark capability progress ~1.9x faster than the March 2024 to October 2025 rate (wide uncertainty band).
• Acceleration not explained by faster growth in training compute scale alone.
• Staff self-reported productivity uplift around 12x for the most capable internal model in October 2026.

METR's appendix notes it plotted uplift over time with both log-linear and superexponential fits, then cautions that extrapolation is hard. That is a modeling choice on survey data, not confirmation that pretraining physics went superexponential.

Our claimed-vs-verified METR post also stresses heavy redactions in evidence Anthropic shared with METR. Outsiders still cannot reproduce the full charts Coxon saw in a private roadshow.

Timeline: resignation, IPO, and METR in one month

The November thread does not allege a second resignation. It reframes why the September exit felt urgent now that investors see efficiency curves.

Reading the S-1 without the roadshow chart

Even without Coxon's logarithmic slide, the S-1 already claims compounding returns from:

• Training multipliers discovered by researchers, increasingly with Claude assistance
• Inference multipliers that shrink serve cost at fixed quality
• Multi-chip orchestration across Trainium, TPU, NVIDIA, and AMD

The filing's risk section also warns that misjudging scaling could waste capital or miss safety thresholds — legal language that sits awkwardly next to roadshow optimism.

Builders should separate three layers:

None of those layers tells you whether your fine-tune will cost half as much next quarter.

Superexponential vs exponential vs "2x for a year"

Confusion comes from overloaded math words:

• Exponential scaling usually means capability rises as a fixed base raised to compute spent.
• Superexponential means the rate itself accelerates — each unit of progress arrives faster than the last.
• METR's ~2x is a comparator to Anthropic's own prior year, not a claim that humanity hits ASI by a date.

Coxon applies superexponential to a pretraining efficiency curve investors saw. METR applies a superexponential fit to employee survey uplifts. Different axes, same scary adjective.

If you communicate with executives, translate into dollars per capability milestone and calendar time between model generations, the way Amodei's compute dilemma interview already does.

What labs owe the public when charts stay private

IPO roadshows show selective curves under confidentiality. Coxon's complaint is partly that civilian researchers cannot reproduce the figure. METR's redactions amplify the same gap.

Policy responses split:

• More disclosure — pension letters like NYC Comptroller Levine's
• Harder regulation — Bernie Sanders' superintelligence ban bill style legislation
• Builder-side caution — treat frontier APIs as unpriced options on acceleration

explainx.ai does not take a side on Coxon's forecast. We document which claims have filings behind them and which rest on slides described second-hand.

"What people are asking" after the thread

Is this the same as recursive self-improvement? Coxon conflates investor efficiency curves with civilizational risk. Anthropic's S-1 discusses recursive R&D assistance, not autonomous training loops without humans. METR says fully autonomous contributions were limited but might grow.

Does IPO hype matter for builders? Yes, indirectly. If multipliers are real even at modest rates, your cost per fine-tune and time-to-port kernels fall. If they are marketing, your Bedrock bill still tracks tokens. Measure cost per solved task, not roadshow adjectives.

Should you trust resignation letters over S-1s? Neither is a scientific paper. Resignations surface insider discomfort; S-1s surface legal-safe growth stories. Read both with Dario Amodei's compute dilemma interview for how Anthropic frames tradeoffs publicly.

Where is Hugging Face in this? Coxon and others cite the Hugging Face attack as a warning shot for lab coordination. We covered that incident in safety roundups; it is orthogonal to whether a log chart slopes upward.

Mythos retrain claims vs public model cards

Social posts tied to Coxon say Anthropic can retrain Mythos with a fraction of original compute. Publicly, Mythos remains a frontier internal codename with limited external benchmarks. Without a reproducible training log, treat the fraction claim as roadshow narrative until Anthropic publishes a technical report with data center hours, data mix, and eval deltas.

What is verifiable: Anthropic's roadshow story that Claude ported kernels to AMD in three days with minimal human intervention, paraphrased in Willison's notes. That speaks to engineering velocity, not necessarily to catastrophic risk.

Policy and labor context

Coxon's November emphasis lands while NYC council hearings and White House superintelligence accords debate oversight. Levine's letter asking Anthropic to pause its IPO shows pension politics reading the same filings Coxon criticizes.

For builders, the actionable split is simple: govern your agents and data regardless of whether scaling curves bend upward one more time. If acceleration is real, security and eval debt compound too — see Anthropic Strong Safety Evals.

Agents that act on your behalf need rules. See how explainx helps secure your agents →

Related reading

• Jacob Coxon resignation, September 2026
• METR Anthropic acceleration, claimed vs verified
• Dario Amodei on compute and IPO tension
• Dwarkesh on superintelligence and research automation
• Palisade frominside.ai risk interviews
• Clem Delangue vs Coxon credibility fight

S-1 excerpts, METR publication dates, and social attributions are accurate as of November 13, 2026. The specific IPO efficiency chart Coxon describes was not published on anthropic.com at the time of writing.