METR Report Reveals Anthropic's Unreleased Model 4 Passing All 15 Benchmarks

An unreleased Anthropic model cleared 50% on all 15 benchmarks in METR's October 2026 assessment — the first model in the test set to do so. The model, identified only as Model 4, finished training between September and October 2026 and has never been announced by Anthropic.

The disclosure comes from METR's independent report, not from an Anthropic product launch. Model 4 sits alongside two other undisclosed predecessors, which METR labels Model 2 and Model 3. Anthropic granted METR access to internal data but redacted significant portions of the underlying evidence before publication. So the assessment exists, but you cannot read most of what backs it.

One detail matters for buyers tracking the vendor's public line: the model marketed as Fable 5.2 corresponds to what METR calls Model 3 Early. In other words, the product customers can licence today is an early-stage cut of a model the lab has since improved on at least twice.

What does 'AI R&D acceleration' actually mean?

METR uses the term to describe AI systems speeding up the research and engineering behind building better AI — gains distinct from those bought with more computing power. The report puts a number on it: roughly 80% of Anthropic's recent capability gains came from AI-assisted research rather than compute growth. METR stressed the estimate carries real uncertainty.

That figure deserves attention. If algorithmic progress, not capital expenditure on chips, drives most capability gains at one leading lab, compute budgets become a weaker proxy for competitive position than many enterprise AI strategies assume.

Is this recursive self-improvement?

Partly. By the definition Google DeepMind CEO Demis Hassabis has used publicly, the pattern METR describes is an early form of recursive self-improvement. AI makes researchers several times more productive while humans stay in the loop, reviewing and directing the work.

The more concerning variant — a closed loop where AI conducts research without human review — has not been confirmed in METR's findings. The report does note early signs of more autonomous AI contributions appearing. That is a warning light on the dashboard, not an alarm.

What does this change for buyers and auditors?

Three practical points stand out:

• Redacted evidence means external validation of frontier claims rests on trust in METR, not reproducible data.
• A capability gap between shipped product (Fable 5.2, i.e. Model 3 Early) and internal frontier (Model 4) is now documented by a third party.
• Capability gains decoupled from compute spending complicate any procurement model that benchmarks labs by infrastructure spend.

Anthropic has confirmed none of this itself. The entire picture — Model 2, Model 3, Model 4 and the 80% figure — comes from a single assessor working with a partially redacted dataset.

For CIOs, the takeaway is procedural rather than technical. Model roadmaps at frontier labs now change faster than release cycles disclose, and independent evaluation may know more about a vendor's inventory than the vendor's own announcements reveal. Procurement due diligence should price that gap in.