Anthropic's Secret Model 4: What a Leaked Report Reveals
METR's independent report found Anthropic runs an unreleased internal Model 4 that tops every benchmark, weeks ahead of public Claude releases.
What is Anthropic’s Model 4?
Model 4 is an unreleased internal AI system built by Anthropic, identified by the independent research group METR in a report published November 5, 2026. METR found that Model 4 finished training between September and October 2026 and scored above 50% on all 15 benchmarks it was tested against, the first internal model to clear that bar across the board, with some scores reportedly near 90%. It has never been released publicly.
TL;DR
- METR, a nonprofit known for tracking how long a task AI agents can autonomously complete, got rare inside access to Anthropic from August 12 to October 29, 2026, including internal benchmarks, compute figures, and staff interviews.
- Anthropic’s internal model lineup includes at least three systems the public has never seen, called Model 2, Model 3, and Model 4 in METR’s report, with Model 4 finishing training just weeks before the report came out.
- Model 4 is the strongest model METR analyzed, clearing 50% on every one of 15 benchmarks and reportedly pushing past 90% on several.
- METR estimates Anthropic’s internal capability gains have been running at roughly 1.9 times the historical rate since October 2025, with a plausible range of 1.5x to 2.8x.
- Anthropic told METR it found no convincing evidence this speedup comes from a jump in training compute, and METR’s own estimate attributes around 80% of the acceleration to AI assisting the research process itself.
- Anthropic staff self-reported productivity gains that rose from roughly 1.2x with Claude 3.7 Sonnet in early 2025 to about 12x by the time Model 4 existed, though METR flagged this data as weak, self-reported evidence.
- Large portions of METR’s underlying evidence, including the analysis behind the “no bottlenecks” finding and the compute data, were redacted by Anthropic before publication.
| Vendor | Tier | Risk | ApprovalNEW |
|---|---|---|---|
| Northwind Supply | Tier 1 | Low | Approved |
| Acme Logistics | Tier 2 | Medium | Approve |
| Helios Data | Tier 1 | High | Pending |
| Granite Freight | Tier 3 | Low | Approved |
| Pinecrest Labs | Tier 2 | Low | Approve |
How did METR get access to Anthropic’s internal models?
METR is a nonprofit that evaluates frontier AI systems, best known for a benchmark tracking how long a task an AI agent can complete unsupervised, a metric that has been doubling at a steady clip over the past few years. Because of that track record, when METR publishes findings about the pace of AI progress, the industry pays attention.
For this specific engagement, METR placed a small team inside Anthropic’s own systems, working on Anthropic’s devices from August 12 to October 29, 2026. The team got internal benchmark scores, training data, compute figures, internal forecasts, staff surveys, and interviews with around 15 researchers. The result was a roughly 100-page internal report for Anthropic, condensed into a shorter public version. The core question METR set out to answer: is AI meaningfully speeding up the process of building better AI. That question matters because if AI starts materially accelerating AI research, it creates a feedback loop, often called recursive self-improvement, that safety researchers have worried about for years.
What did METR actually find?
Three findings drove most of the reaction. First, as of October 2026, Anthropic appears to be improving its models’ capabilities at roughly twice the rate it did between March 2024 and October 2025. Second, METR’s rough estimate is that about 80% of that acceleration comes from AI itself assisting the research process, not from simply buying more compute. Third, METR said it found no clear bottlenecks, whether in data, compute, or algorithmic progress, that would obviously slow this down within the next 12 months.
Each of these came with real uncertainty attached. METR’s own confidence interval on the 1.9x acceleration figure runs from 1.5x to 2.8x, and the group said it wouldn’t be shocked if the true number were as low as 1.3x or as high as 3x. When METR broadened its analysis to include other labs like OpenAI and Google, the industry-wide acceleration estimate came out lower, around 1.4x, suggesting Anthropic’s unreleased internal models are a big part of why its own number runs higher.
What do we know about Models 2, 3, and 4?
METR’s report references a sequence of internal checkpoints the public has never seen. Model 2 finished training in July 2026. Model 3 finished training between August and September. Model 4 finished training between September and October, essentially right before METR’s report came out.
Model 4 stands out because it’s the first internal model in METR’s dataset to score above 50% on all 15 benchmarks tested, with performance on several benchmarks reportedly reaching around 90%. Commentary circulating after the report’s release pointed to a chart plotting Anthropic’s internal capability frontier against what’s actually available to the public: the internal line consistently runs ahead, and Model 4 reportedly sat more than 40 days ahead of anything Anthropic customers could access. Some online speculation has tied Model 4 to a future “Claude 6” release, but that link is not confirmed by METR’s report and should be treated as guesswork rather than fact.
Is the acceleration driven by more compute?
Not according to METR. The group stated it received convincing evidence that the capability acceleration is not due to an increase in the growth of training compute, though the specific evidence behind that conclusion was redacted by Anthropic before publication. This is a meaningful detail: for years, the standard explanation for better AI models was “more chips, more data, more scale.” METR’s report suggests something different is now doing a lot of the work, namely AI systems themselves helping researchers do research faster, which METR pegs at roughly 80% of the current acceleration, with real uncertainty around that number.
How much faster is Anthropic’s AI research actually moving?
METR’s headline chart uses something called AECI, an Anthropic-specific version of the Epoch Capabilities Index that rolls multiple benchmark scores into a single comparable number over time. Plotting model releases (including internal ones) against this score shows a clear kink around October 2025: before that point, scores were climbing at roughly 15 points per year; after it, the climb roughly doubled to around 29 points per year. METR summarized this by noting Anthropic appears to have made as much benchmark progress in the past year as it did in the two years before that, a period that spans the public jump from Claude 2.1 to Sonnet 4.5.
Separately, METR gathered informal survey data and Slack polls asking Anthropic staff how much AI was speeding up their own work. Self-reported multipliers rose from around 1.2x in early 2025 (around the time of Claude 3.7 Sonnet) to above 5x by the time of a model internally called Mythos 5, to roughly 12x by October 2026 with Model 4 in use. METR was explicit that this is soft, self-reported evidence, not a rigorous measurement, and one researcher interviewed pushed back hard, estimating only 15-20% of their speedup was attributable to AI and saying they distrust self-reported productivity claims generally. Interview responses centered closer to a 2x multiplier when averaged out, with one researcher noting they were now bottlenecked mainly on ideas and compute rather than on writing code.
Does this mean recursive self-improvement has started?
That’s the debate the report triggered. Demis Hassabis, CEO of Google DeepMind, has described early-stage recursive self-improvement (RSI) as AI making human engineers several times more productive while humans stay in the loop reviewing and directing the work, versus a riskier “closed loop” RSI where humans are removed from the process entirely. By that framing, what METR documented inside Anthropic looks like the early, human-in-the-loop version: AI is still mainly assisting researchers rather than running experiments fully autonomously, though METR noted early signs of more autonomous AI contributions emerging.
Run it on our own servers.
✓SOC 2 and HIPAA from day one.
✓Log every action.
✓Only the right people get in.
✓The context that makes this notable is a comparison to Anthropic CEO Dario Amodei’s own public comments earlier in 2025, when he estimated AI coding tools were giving Anthropic something like a 15-20% total research speedup, up from around 5% six months prior, calling it a “gradually growing” advantage. METR’s report, released months later, describes something that looks considerably larger than gradual.
Is this something to actually worry about?
The honest answer is that the evidence is suggestive but incomplete. Anthropic redacted large portions of METR’s underlying analysis, including detail behind the “no significant bottlenecks” finding and specifics of the compute evidence, so outside researchers can see METR’s conclusions without being able to independently verify most of the supporting data. METR itself flagged substantial uncertainty in nearly every number in the report, from the 80% AI-attribution estimate to the acceleration multiplier itself. What is clear is that Anthropic has at least three internal models, unreleased to the public, that outperform anything currently shipping, and that the gap between what Anthropic can use internally and what the public can access appears to be measured in weeks, not days.
Frequently Asked Questions
What is Anthropic’s Model 4?
Model 4 is an internal, unreleased Anthropic AI model identified in METR’s October 2026 assessment. It finished training between September and October 2026 and was the first model in METR’s test set to score above 50% on all 15 benchmarks evaluated.
Has Anthropic confirmed Model 4 publicly?
No. Model 4, along with internal predecessors METR calls Model 2 and Model 3, comes from METR’s independent report rather than an Anthropic product announcement. Anthropic gave METR access to internal data but redacted significant portions of the underlying evidence before the report was published.
What does METR mean by “AI R&D acceleration”?
It refers to AI systems helping speed up the process of researching and building better AI, as opposed to speedups from simply buying more computing power. METR estimated that around 80% of Anthropic’s recent capability gains come from this kind of AI-assisted research rather than compute growth, though it stressed this estimate carries real uncertainty.
Is this the same thing as recursive self-improvement?
It’s an early form of it, by the definition Google DeepMind CEO Demis Hassabis has used publicly: AI making researchers several times more productive while humans remain in the loop reviewing and directing the work. A more concerning “closed loop” version, where AI does the research without human review, has not been confirmed in METR’s findings, though the report notes early signs of more autonomous AI contributions appearing.
Why does the compute question matter?
If Anthropic’s progress were mainly due to spending more on chips, the acceleration would be explainable by scale alone. METR instead found convincing evidence that growth in training compute was not the driver, pointing instead toward AI assistance in the research process itself as the larger factor, a dynamic that’s harder to predict or bound than compute spending.
