Anthropic's Hidden Models: Mythos, Model 2, 3, and 4 Explained
Anthropic's IPO filing reveals unreleased internal models beyond Mythos, including a Model 4 that broke internal benchmarks researchers struggled to measure.
What did Anthropic’s IPO filing reveal about its unreleased models?
Anthropic’s IPO filing and a related assessment from the research group Meter disclosed that the company has at least three unreleased internal models beyond Mythos, the largest model it has given to trusted partners but not the public. These are referred to by the shorthand names Model 2, Model 3, and Model 4. According to the filing, Model 2 finished training in July 2026, Model 3 in August or September 2026, and Model 4 in September or October 2026. Model 4 reportedly broke internal benchmarks that no previous model, including Mythos, had been able to touch, which made it hard for Anthropic’s own researchers to measure how capable it actually is.
TL;DR
- Anthropic’s IPO documentation at ipo.anthropic.com shows a model hierarchy that goes beyond what the public has access to: Fable is the biggest model available to everyone, Mythos is a step above that with limited release, and behind Mythos sit Model 2, Model 3, and Model 4.
- Model 4 finished training in roughly September to October 2026 and reportedly exceeded internal benchmarks so thoroughly that researchers lost a clear way to track how fast it was improving.
- Anthropic amended its S1 filing to say that recursive self-improvement (AI meaningfully accelerating its own research and development) is approaching, and it set a formal safety tripwire at the point where AI-driven acceleration doubles research progress.
- The company’s own tracking puts current acceleration at around 1.9x of that 2x threshold, with a wider confidence range that could run as high as 3x, meaning the estimate itself is becoming less reliable near the cutoff.
- Meter’s independent 11-week audit found a sharp speed-up tied to a specific, redacted part of Anthropic’s training pipeline, something tied to a secret agent scaffold used alongside Mythos, and Anthropic pulled back on sharing compute-allocation data mid-audit.
- Anthropic has stated the acceleration is not driven by scaling up training compute (more chips, more data centers), which points toward an algorithmic or process-level change instead.
- None of this confirms “hyperbolic” runaway growth yet. The benchmark-breaking performance of Model 4 is striking, but it is not by itself evidence of accelerating acceleration.
Nothing gets lost.
What is Anthropic’s model hierarchy, and where does Mythos fit?
The IPO page lays out a tier structure. Fable is described as the most capable model generally available to the public. Mythos sits above Fable in the same general class but has only been shared with select trusted organizations rather than released broadly. Behind Mythos, the filing and Meter’s report point to three further models that the public has never interacted with: Model 2, Model 3, and Model 4.
These aren’t necessarily the models’ internal project names, just the labels that became visible through the disclosure. Model 2 and Model 3 are mentioned mainly as context: they’re newer and presumably more capable than Mythos, but they weren’t the focus of the disclosures. Model 4 is the one that stands out, because it reportedly did something none of the earlier models, including Mythos, could do: it broke past benchmarks Anthropic uses internally to measure capability, to the point that the usual yardsticks stopped being useful.
That detail matters because it signals a capability jump large enough to outrun Anthropic’s own measurement tools, not just a modest score improvement.
What is recursive self-improvement, and why did Anthropic flag it now?
Recursive self-improvement (RSI) describes a condition where AI systems contribute meaningfully to researching and building better AI systems, creating a feedback loop. Anthropic’s own data, as described in the filing, shows this has been a gradual slope rather than a sudden switch. The company tracks AI’s role in its research and development on a scale from no AI involvement, to AI assisting, to AI collaborating, to AI leading certain projects, up to fully automated research and development.
By Anthropic’s account, AI was leading about 1% of relevant research projects in March of the same year, and that figure rose to roughly 26% by August. That’s a meaningful jump in a matter of months. Anthropic also set a formal threshold: the point where AI-driven acceleration doubles the pace of research progress (a 2x multiplier) compared to a world without AI assistance. Crossing that line is treated as a safety tripwire under the company’s responsible scaling policy, a signal that development has entered territory that is harder to predict and needs a concrete response.
As of the most recent estimate referenced in the filing, Anthropic places current acceleration at around 1.9x, just under that 2x line. But the confidence interval around that number widens the closer it gets to the threshold, with some estimates running as high as 3x. In other words, the company isn’t just saying “we’re close,” it’s saying its own ability to measure how close is getting shakier.
What did Meter’s audit find, and why was so much redacted?
Without shadow IT.
Meter, an independent group that evaluates AI systems, ran an embedded 11-week audit inside Anthropic looking specifically at how AI is contributing to AI research. According to Meter’s report, the audit uncovered a clear acceleration tied to a specific part of Anthropic’s training pipeline, but large portions of the explanation were redacted at Anthropic’s request. Meter’s own language noted that some redactions were themselves hard to describe without revealing the sensitive information behind them.
Two things were singled out repeatedly in connection with this speed-up: an unnamed “agent scaffold” (described as something other than Claude Code) and Mythos itself. The combination of the two appears to be what’s driving the jump. Meter’s report also noted that a planned compute-allocation audit, something Anthropic had reportedly agreed to let Meter examine, was pulled back partway through, with historical data on algorithmic efficiency, training and inference compute use, and data efficiency withheld.
Anthropic has been explicit on one point: it says it’s confident the acceleration is not the result of scaling up training compute, meaning it isn’t simply a function of buying more GPUs or building more data centers. That leaves an algorithmic or process-level explanation, which is consistent with the redacted references to a new agent scaffold and shifts in how compute is allocated across pretraining, reinforcement learning, and inference.
Is this evidence of runaway or “hyperbolic” AI growth?
Not on its own. A model breaking internal benchmarks is a sign of a real capability jump, but it doesn’t by itself demonstrate that the rate of improvement is itself accelerating, which is what “hyperbolic growth” would mean. Anthropic co-founder Daniel Kokotajlo has publicly responded to the amended S1 filing by saying that if the disclosed timeline holds, it suggests the path to full recursive self-improvement and more advanced systems could be shorter than previously expected. But that’s a statement about trajectory, not confirmation that growth has gone hyperbolic.
The redactions make it hard to independently verify how much of the acceleration comes from a one-time breakthrough (a new scaffold or technique) versus a sustained, compounding trend. Anthropic’s own confidence interval around its 1.9x estimate, which stretches toward 3x, is itself an admission that uncertainty increases right at the point where it would matter most.
Frequently Asked Questions
What is Anthropic’s Model 4?
Model 4 is the internal name that emerged from Anthropic’s IPO filing and Meter’s audit for an unreleased model that finished training around September to October 2026. It reportedly surpassed internal benchmarks that no earlier Anthropic model, including Mythos, had been able to beat, making it difficult to measure its full capability with existing tools.
What is Mythos, and is it available to the public?
Mythos is described as Anthropic’s largest model in the same general class as Fable, the biggest model available to the general public. Unlike Fable, Mythos has only been shared with select trusted organizations, not released broadly.
What is recursive self-improvement in AI development?
Recursive self-improvement refers to AI systems meaningfully contributing to the research and development of future AI systems, creating a feedback loop where better AI helps build even better AI. Anthropic tracks this on a scale from no AI involvement to fully automated research and development, and its filing noted AI was “leading” around 26% of relevant projects as of August, up from about 1% in March.
Why did Anthropic redact so much of Meter’s audit report?
Without shadow IT.
Anthropic has said that detailing exactly what is driving the acceleration risks providing a roadmap that competitors could use to replicate the same gains. The redactions cluster around a specific, unnamed agent scaffold and changes in compute allocation across training and inference.
Does this mean AI has crossed a dangerous threshold?
Not according to Anthropic’s own figures. The company set a 2x acceleration threshold as a safety tripwire and currently estimates it’s at around 1.9x, just under that line, though the uncertainty range around that estimate extends as high as 3x.
