METR report: Anthropic's unreleased Model 4 scores 178.6 on AECI, ahead of Opus 5.5

Third-party evaluator METR has disclosed three unreleased Anthropic models, and the strongest of them, Model 4, scores 178.6 on Anthropic's AECI capability index, about 3.9 points above Opus 5.5. The report was covered on November 6, 2026 by Chinese outlet QbitAI (量子位), whose Chinese-language article is the basis for this piece.

The three models

The models are called Model 2, Model 3 and Model 4. Their final names have not been made public. According to METR, training was completed in July (Model 2), August to September (Model 3) and September to October (Model 4) of this year.

AECI is Anthropic's version of the Epoch Capabilities Index. It combines results from multiple internal and external benchmarks to track capability over time. In the AECI chart in the report, Model 2 scores close to Opus 5.5, while Model 3 and Model 4 score higher.

• Model 4 AECI: 178.6, the highest of the group and about 3.9 points above Opus 5.5.
• Equivalent progress: QbitAI says that, at Anthropic's past rate of improvement, 3.9 points corresponds to roughly three months of progress.
• Opus 5.5 timing: QbitAI says Opus 5.5 was released a month and a half ago.
• Trend: METR estimates that AECI growth for Anthropic models since October 2025 has run at about 1.9 times the earlier rate.
• Benchmark saturation: all 15 benchmarks included in Model 4's AECI evaluation show scores above 50%. Opus 4.6, from about nine months ago, scored below 50% on many of them.

METR says it is concerned that hard tests able to fully challenge these models are running short, which will make future capability measurement harder.

Anthropic's response

Anthropic co-founder Sam McCandlish responded publicly. He agreed with several of METR's judgments: AI is speeding up Anthropic's R&D, but how much is hard to measure precisely.

He said Anthropic gave METR benchmark results for unreleased models, estimates of model progress, and access to researchers so the evaluator could reach independent conclusions. He also acknowledged that the public report removed a large amount of information, which makes METR's analysis harder for others to verify. In his view, details of the R&D speedup are commercially sensitive and could give other companies an "acceleration roadmap."

McCandlish said Anthropic will publish its own progress analysis in the coming weeks. It will cover the pace of internal AI R&D, where the company stands relative to its Responsible Scaling Policy, and which new safety measures it plans to adopt.

AI in Anthropic's R&D

In September, Anthropic published data showing that, as of August, Claude had led 26% of model R&D work under human supervision, up from under 1% six months earlier.

The METR report includes interview excerpts from Anthropic researchers. One said the time from generating an idea to running an experiment has dropped by about 80%, and that ideas and compute, not code, are now the main bottlenecks. Others said Claude and agent tools let teams try more ideas at once, including directions they previously would not have explored.

METR leans toward the view that broad AI use accounts for most of Anthropic's current R&D acceleration. It stresses that the speedup mainly comes from AI assisting human researchers, and that fully autonomous AI R&D is not yet the main source of progress.

In September, Dario Amodei published an essay, "We must pace the frontier," arguing that AI capability growth must slow. By his assessment, AI progress has clearly accelerated since around this past summer, driven mainly by AI helping to build the next generation of AI. A New York Times article dated November 5 reports that his concern about AI risk motivates his push for technical leadership. Under his vision, per that account, training continues but the company leaves enough time for alignment, safeguards and third-party evaluation.

What this means

The report puts a number on the gap between Anthropic's unreleased and public models: Model 4 leads Opus 5.5 by about 3.9 AECI points, which QbitAI equates to roughly three months of past progress. The figures come with limits. Anthropic itself says the public report omits significant detail, so the AECI numbers cannot be independently checked from the published material. METR's finding that AI assistance drives most of the current speedup, while autonomous AI R&D does not, is the clearest description yet of how Claude is contributing to its own successors. Anthropic's promised analysis, covering internal R&D speed and its position under its Responsible Scaling Policy, is the next source of detail.