{"slug": "anthropics-secret-model-4-raises-the-stakes", "title": "Anthropic’s Secret Model 4 Raises the Stakes", "summary": "Anthropic's model-capability progress accelerated by an estimated 1.9 times its previous pace, with a wide uncertainty range of 1.5 to 2.8 times, according to an independent assessment by the non-profit research group METR published as \"Independent Assessment of AI R&D Acceleration at Anthropic in October 2026.\" METR attributed roughly 80% of the observed acceleration to AI itself, though it flagged substantial uncertainty and redacted key supporting details, and reported that an unreleased Anthropic model trained between September and October 2026, dubbed Model 4, scored above 50% on all 15 benchmarks in the tested set. Internal Slack surveys cited by METR suggested AI tools boosted productivity on specific tasks by up to 12 times, while interviews with approximately 15 Anthropic researchers put general efficiency gains closer to two times.", "body_md": "## The acceleration claim that set off alarms\n\nAnthropic’s model-capability progress accelerated by an estimated **1.9 times** its previous pace, according to an independent assessment by METR. While this figure carries a wide uncertainty range, from 1.5 to 2.8 times, it suggests the company achieved roughly as much benchmark progress in the past year as in the prior two years. This acceleration describes measured model capabilities, not every dimension of AI advancement.\n\nMETR, a non-profit research group, published these findings in a public report following temporary, rare access to Anthropic's internal systems and materials. Their full assessment, titled \"Independent Assessment of AI R&D Acceleration at Anthropic in October 2026,\" distinguishes METR’s rigorous findings from the more alarmist framing found in some online videos.\n\nResearchers had access to internal benchmark scores, training data, compute figures, forecasts, and conducted interviews with approximately 15 Anthropic researchers. This unusual level of access allowed for a granular analysis of the company's R&D trajectory, providing an unvarnished look at the pace of frontier AI development.\n\n## Model 4 is powerful—and still a black box\n\nMETR’s assessment introduced **Model 4**, an unreleased Anthropic model trained between September and October 2026. This internal system achieved scores above 50% on all 15 benchmarks in the tested set, marking it as the highest-scoring model in the report. While impressive, these scores reflect performance on specific tasks and do not inherently establish broad or human-level competence.\n\nThe existence of such powerful, unreleased models highlights a critical **internal-public capability gap**. Progress in AI development can outpace what customers and independent observers can access, meaning the frontier of AI capability is often further advanced than publicly available models like [Claude](https://www.stork.ai/en/claude-ai-workspaces). Speculation about Model 4's lineage or its potential release as a future Claude version remains unconfirmed.\n\nHigh scores on these benchmarks also raise questions about **benchmark saturation**. As models consistently clear 50% and even approach 90% on some tests, the utility of existing benchmarks for distinguishing capabilities may diminish. A benchmark score provides evidence of performance on defined tasks, but it is not a complete measure of real-world reliability or general intelligence.\n\n## AI may be speeding up the people who build AI\n\nMETR cautiously attributed roughly 80% of the observed acceleration to AI itself, though it underscored substantial uncertainty in this estimate. While faster growth in training compute did not explain the change, key supporting details regarding this finding were redacted from the public report. For additional detail, consult the full [Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 - METR](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/).\n\nInformal employee self-reports, gathered via internal Slack surveys, suggested AI tools boosted productivity on specific tasks by up to **12 times**. However, this figure represents anecdotal feedback on particular workflows, not a controlled measurement of overall company output. Interviews with approximately 15 researchers offered varied estimates, with many centering closer to a **two-times** improvement in general efficiency.\n\nThis suggests a nascent **feedback loop**: AI tools assist researchers in coding, testing ideas, and streamlining workflows. These improvements may, in turn, contribute to the development of better models, which then further assist future research. METR’s account still places humans in the loop for much of this work, distinguishing it from an unconstrained intelligence explosion.\n\nAI-driven advancements are effectively compressing the R&D cycle. Researchers leverage models to - generate code - design experiments - analyze results - refine training processes\n\nThis iterative enhancement accelerates the discovery and implementation of new capabilities.\n\nEnjoying this? Get one like it in your inbox each morning.\n\none email a day · unsubscribe in two clicks · no third-party tracking\n\n## The unanswered question is what can slow it down\n\nMETR’s assessment did not identify convincing evidence of major constraints likely to emerge within the next 12 months. This finding, however, comes with limited confidence; specific analyses of potential bottlenecks were among the details withheld from the public report.\n\nTransparency remains a central challenge. Anthropic’s extensive redactions leave readers unable to inspect some of the most informative evidence. This includes the specifics behind the assessment that faster growth in training compute did not explain the acceleration, as well as the full bottleneck analysis.\n\nPublicly, METR states only that it found no clear near-term limits to the observed acceleration. What cannot be verified publicly are the underlying data and methodologies that led to this conclusion, making it difficult for external parties to scrutinize or replicate the findings.\n\nIf AI-assisted research continues to accelerate, the pressure on independent evaluation and safety work will intensify to keep pace. The practical stakes are high: clearer disclosures, more frequent independent assessments, and rigorous audits are critical next steps to understand the true trajectory of AI progress and its implications.\n\n## Frequently Asked Questions\n\n### What did METR find about Anthropic’s AI progress?\n\nMETR estimated that Anthropic’s model capabilities were advancing at roughly 1.9 times their previous rate, with substantial uncertainty around the estimate.\n\n### What is Anthropic’s Model 4?\n\nModel 4 is METR’s label for an unreleased Anthropic model it evaluated. The report describes it as the strongest model in its tested set.\n\n### Does the report prove that AI is building better AI on its own?\n\nNo. METR attributed much of the acceleration to AI assistance, but described current use mainly as helping human researchers, not replacing them.\n\n### Why are parts of METR’s report redacted?\n\nAnthropic reportedly withheld sensitive details, including evidence about compute and bottlenecks, limiting outsiders’ ability to independently assess some conclusions.", "url": "https://wpnews.pro/news/anthropics-secret-model-4-raises-the-stakes", "canonical_source": "https://www.stork.ai/blog/anthropics-secret-model-4-raises-the-stakes", "published_at": "2026-11-06 23:53:31+00:00", "updated_at": "2026-11-07 00:51:53.048250+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-safety", "large-language-models", "ai-policy"], "entities": ["Anthropic", "METR", "Model 4", "Claude"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/anthropics-secret-model-4-raises-the-stakes", "markdown": "https://wpnews.pro/news/anthropics-secret-model-4-raises-the-stakes.md", "text": "https://wpnews.pro/news/anthropics-secret-model-4-raises-the-stakes.txt", "jsonld": "https://wpnews.pro/news/anthropics-secret-model-4-raises-the-stakes.jsonld"}}