The headline claim
The story now circulating under the banner “Anthropic’s New Model 4 Beats Opus 5.5” centers on an internal capability comparison rather than a public product launch. According to recent reporting on METR’s assessment, Anthropic gave evaluators visibility into unreleased systems labeled Model 2, Model 3 and Model 4, with Model 4 scoring 178.6 on AECI, Anthropic’s version of the Epoch Capabilities Index . That score is reported as roughly 3.9 to 4.0 points above Opus 5.5, which is placed around 174.5 to 174.56 on the same chart .
That is the hard kernel of the story: Model 4 is not described as a consumer release, not priced, not documented as an API endpoint, and not available for outside testing. It is an internal frontier checkpoint, apparently trained in the September-to-October 2026 window, and its public significance comes from what it implies about Anthropic’s development curve . The AI Revolution video indexed on November 8 framed the same finding as “Model 4 beats Opus 5.5” and linked it to ASI timelines, while also noting that full autonomous self-improvement remains unproven .
What AECI says, and why it matters
AECI matters because it converts results across many benchmarks into a single latent capability estimate. Recent summaries of the METR work say Anthropic’s models improved at about 15 AECI points per year before October 2025 and about 29 points per year afterward, producing a central estimate of roughly 1.9 times faster progress . The reported 95 percent confidence interval runs from 1.5 to 2.8 times, while broader commentary places the plausible range around 1.3 to 3 times .
This is why a 3.9-point gap between Model 4 and Opus 5.5 is not being treated as a trivial leaderboard bump. TPS, citing QbitAI’s coverage, says the gap corresponds to roughly three months of Anthropic’s past rate of progress . In a slow-moving software market, three months might be routine; in frontier AI, where model releases already compress product cycles, an internal system apparently leapfrogging the latest public flagship within weeks becomes strategically meaningful.
The most important caveat is that AECI is not a universal IQ scale. It is a fitted index that depends on which benchmarks are included, how scores are transformed, and how saturated those tests have become. AF’s summary says Model 4 was the first model in the METR dataset to clear 50 percent across all 15 evaluated benchmarks, with some results approaching 90 percent . That sounds impressive, but it also raises a measurement problem: if the tests are running out of headroom, the index may become less stable just when it is most needed.
The acceleration claim
The second layer of the story is bigger than Model 4. METR’s assessment reportedly covered 11 weeks of access, from August 12 to October 29, 2026, with internal metrics, benchmark results, employee surveys and roughly 15 interviews . Its central conclusion, as reported by several outlets, is that Anthropic’s capability progress on well-specified tasks has accelerated to about twice its previous rate .
Oossa reports that METR estimated roughly 80 percent of the acceleration may come from AI helping Anthropic researchers, rather than from faster growth in training compute . AF similarly states that the audit found the acceleration was not primarily driven by expanded training compute and that METR saw no immediate bottleneck in data, compute or algorithms expected to stop the trend over the next 12 months . That is the part that makes this story feel different from a normal benchmark race: the alleged driver is not merely more chips or more data, but AI assistance inside the research loop.
The reported employee-productivity numbers are even more dramatic, but also weaker evidence. Oossa says informal October 2026 polls had Anthropic employees reporting average productivity gains of about 12 times with the latest internal model . The same source cautions that METR treated such self-reports as loose evidence that may overstate gains . In other words, the 12x number is useful as a signal of perceived workflow change, not as a settled measurement of company-wide output.
Recursive self-improvement, but not yet the sci-fi version
The phrase “recursive self-improvement” can mislead. The strongest version would mean an AI system autonomously designs, trains, evaluates and deploys a superior successor, which then repeats the cycle with diminishing human involvement. The reporting around Anthropic’s Model 4 does not establish that scenario. TPS says METR’s view is that the speedup comes mainly from AI assisting human researchers, while fully autonomous AI R&D is not yet the main source of progress .
That distinction is crucial. A lab where researchers use Claude agents to write code, generate experiment scaffolds, analyze logs and run parallel ideas is already a recursive-looking system in the economic sense: better AI helps make the next AI faster. But it is still human-directed. The current evidence points to a human-in-the-loop acceleration machine, not a self-contained artificial scientist.
Even so, the risk question changes. If AI assistance compresses experiment cycles, then model developers may need faster safety evaluations, faster red-team loops and stronger rollback procedures. If the same family of models helps write the systems and evaluate the systems, evaluation independence also becomes a live governance issue. The worry is not simply “ASI tomorrow”; it is that ordinary safety processes may be calibrated for yesterday’s development speed.
Redactions and the trust gap
The most serious limitation is the redaction problem. TPS reports that Anthropic co-founder Sam McCandlish acknowledged that much information was removed from the public report, making the analysis harder to verify, while arguing that the missing details could reveal a commercially sensitive “acceleration roadmap” . Oossa likewise reports that Anthropic redacted much of the evidence METR considered most useful, including forecasts and bottleneck analysis .
That creates a difficult balance. Anthropic has a plausible reason not to publish operational details that could help rivals accelerate. But the more the key evidence is hidden, the more outsiders must rely on METR’s judgment and Anthropic’s selective disclosure. For a finding as consequential as “AI may be doubling AI-development speed,” that is uncomfortable.
ASI in two years?
The “ASI in 2 years” framing is best read as a provocative forecast, not a conclusion of the audit. The 36Kr Europe version quotes METR researcher Ryan Greenblatt as saying the report is a strong signal for ASI within one or two years, but that claim is a timeline interpretation layered on top of the acceleration evidence . The same body of reporting also stresses uncertainty, redactions, benchmark limitations and the absence of fully autonomous research .
A cautious reading is therefore stronger than a sensational one. The Model 4 result suggests Anthropic’s internal frontier may already be ahead of its latest public flagship. The AECI trend suggests the company’s measured capability progress may have nearly doubled since October 2025. The workflow evidence suggests Claude-like systems are now materially speeding the work of building successor systems. But none of that proves ASI is inevitable by 2028.
Why this story matters
The real development is not that a secret model scored a few points higher than Opus 5.5. It is that those points sit inside a broader pattern: internal models are arriving quickly, evaluation headroom is narrowing, AI-assisted research is becoming central, and outside verification remains constrained by redaction.
If Anthropic’s forthcoming self-analysis confirms the core claims, the policy debate will move from abstract concern about recursive self-improvement to concrete questions: how should labs measure AI-caused acceleration, when should responsible-scaling thresholds trigger, and how can safety work keep pace when the research engine itself is being automated? For now, Model 4 is less a public model than a warning light on the dashboard. It does not prove the arrival of ASI, but it does suggest that the clock governing frontier AI development may be ticking faster than expected.
Comments
Be the first to comment.