AI Accelerates AI: Anthropic's Internal Research Hits Warning Threshold

A recent audit by the nonprofit METR, which gained unprecedented access to Anthropic's internal data, reveals a significant acceleration in the pace of the company's AI model improvements. The most striking finding is that roughly 80 percent of this speedup is attributed to AI being used in Anthropic's own research, potentially triggering a critical warning level outlined in the company's own Responsible Scaling Policy.

What happened

According to The Decoder, citing METR's audit, Anthropic's model test scores have been climbing about 1.9 times faster since October 2025 compared to the preceding period. This measurement is based on the AECI, a composite metric bundling various benchmarks. The Decoder reports that this means Anthropic achieved in one year what previously took two, comparable to the jump from Claude 2.1 to Sonnet 4.5. METR's analysis included three unreleased models trained between July and October 2026.

METR estimates that approximately 80 percent of this acceleration stems from AI being employed in Anthropic's internal research, not just increased computing power. Employees reported in interviews that experiments that once took days to set up are now "~free," and informal Slack polls suggested a 12x productivity boost from the latest internal model, though METR treats these self-assessments as "very loose evidence." While AI is primarily acting as an assistant, METR noted "signs of life" that independent AI contributions could soon play a larger role. The auditors found no bottlenecks that would slow this pace over the next year, indicating a "significant risk" of continued acceleration.

This acceleration is particularly significant because Anthropic's Responsible Scaling Policy flags a doubling of AI progress due to automated AI research as a risk indicator. The Decoder, citing METR's companion review, notes that while Anthropic's public risk report stated acceleration was "less than a factor of 2," the internal, unredacted version implied a value of 1.97. METR considers it plausible the warning level was reached as early as July 15, accusing the public report of understating both the acceleration and the uncertainty.

However, the results are imprecise, with the true acceleration factor potentially ranging from 1.3 to 3. When models from other major players like OpenAI, Google, and xAI are included, the industry-wide acceleration drops to 1.4, possibly due to Anthropic catching up from a slower 2025. Views within Anthropic are also not uniform, with some researchers attributing a much smaller share of productivity gains to AI or crediting new hires as the primary factor.

Crucially, Anthropic exercised unilateral redaction rights over "the majority of the evidence the auditors rated as most informative," including future model forecasts and bottleneck analyses. This means the public must rely on METR's judgment "much more than would be ideal," according to the organization, with one graph in METR's report explicitly labeled "FAKE DATA" due to redactions.

Why it matters for small companies

This audit isn't just about Anthropic; it's a stark indicator of an accelerating trend that will reshape every industry. The fact that AI is now significantly contributing to the development of better AI means the pace of technological advancement is no longer linear but exponential. For small companies, this isn't a distant concern; it means the AI tools you adopt today will be superseded by vastly more capable versions at an ever-increasing rate. The competitive advantage will shift from merely using AI to rapidly integrating and adapting to its evolving capabilities. This isn't just an AI revolution; it's an AI-driven innovation feedback loop, demanding unprecedented agility from every business.

What to do this week

1. Audit AI-assisted workflows: Review your current business processes and identify areas where AI tools could act as powerful assistants, mirroring how Anthropic uses AI in its own research to reduce setup time and initial effort.
2. Dedicate experiment time: Allocate a few hours this week for your team to experiment with a new AI feature or tool relevant to their roles, focusing on how it could streamline a specific, repetitive task.
3. Prioritize continuous learning: Encourage team members to spend a small, consistent amount of time (e.g., 30-60 minutes weekly) exploring new AI capabilities and best practices, fostering a culture of rapid adaptation to emerging technologies.

Sources

• METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July — The Decoder