---
id: "lib-2026-11-05-independent-assessment-of-ai-r-and-d-acceleration"
type: "library"
title: "Independent Assessment of AI R&D Acceleration at Anthropic in October 2026"
lang: "EN"
org: "METR"
source_type: "Report"
date: "2026-11-05"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-05-independent-assessment-of-ai-r-and-d-acceleration/"
original_url: "https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/"
tier: "S"
tier_name: "Shapes the field"
tier_reason: "METR, the leading independent evaluator of frontier AI risk, reports the first high-access assessment of AI-driven R&D acceleration inside a frontier lab, with its method, uncertainty ranges and redactions disclosed."
summary_basis: "full-text"
original_minutes: "50"
tags: ["ai technology", "ai r&d automation", "ai safety", "business transformation"]
---

# Independent Assessment of AI R&D Acceleration at Anthropic in October 2026

Tier S (Shapes the field): METR, the leading independent evaluator of frontier AI risk, reports the first high-access assessment of AI-driven R&D acceleration inside a frontier lab, with its method, uncertainty ranges and redactions disclosed.

Summary based on the full text.

## In short

METR staff (2.5 full-time equivalents) spent 12 August to 29 October 2026 inside Anthropic assessing how much AI is speeding up AI R&D. Their judgement: Anthropic is likely improving capabilities on well-specified tasks at about twice its March 2024-October 2025 rate, mostly because of its own AI use, and METR found no clear bottleneck for the next 12 months. Anthropic held redaction rights and removed a majority of the evidence METR found most informative, so readers have to rely on METR's judgement more than METR considers ideal.

## Key numbers

- around twice: Anthropic's current rate of capability improvement on well-specified tasks vs March 2024-October 2025 (plausible 1.3x to 3x)
- 1.9x [1.5-2.8x]: acceleration of AECI benchmark growth since October 2025, 95% confidence interval
- ≈80%: METR's guess at the share of current acceleration due to AI use in R&D
- ≈12x: average self-reported staff uplift from the most capable internal model, October 2026 (loose evidence)
- 1.1x-2.9x: range of acceleration estimates across robustness checks

## What it reports

- AECI (Anthropic's version of the Epoch Capabilities Index) has grown 1.9x faster since a breakpoint estimated at 7 October 2025, roughly two years of historical progress in one year. The Haiku 5.5 system card reported 1.7x; the difference is newer unreleased models.
- Robustness: acceleration ranges 1.1x-2.9x depending on the link function, 1.9x-2.2x across snapshot and benchmark choices.
- METR is confident the acceleration is not due to faster growth in training compute.
- Interviews centred on a 2x acceleration; Slack-poll self-reported uplift reached about 12x, which METR treats as very loose evidence.
- R&D is accelerated mainly through AI assistance, not fully autonomous work, with "signs of life" of autonomous contributions.
- Anthropic cancelled a planned component verifying compute allocation between capabilities and safety.
- METR took no payment; Anthropic provided tokens and devices.

## How they got it

High-access review of internal benchmark scores for unreleased models, training compute data, algorithmic efficiency metrics, headcount, code and PR counts, uplift surveys and about 15 staff interviews. Acceleration is estimated with piecewise linear fits with one breakpoint and bootstrap intervals. A roughly 100-page full report went to Anthropic; this is the redacted public version.

## Limits

- Most decisive evidence is redacted; the table of shared versus published evidence shows several categories withheld.
- AECI covers verifiable tasks and internal benchmarks are near saturation; acceleration may not carry to messy real-world work.
- Anthropic's pre-2025 growth was below the industry frontier, which may inflate the measured acceleration.
- METR reports reversals in its own interpretation during the engagement and lacked staff-level access.

## Why it matters for companies

If frontier capability gains are compounding at about twice the 2024-25 pace, planning cycles built on annual model upgrades are too slow: tooling, evaluations and policies need refresh on a quarterly rhythm. The report also sets a template for independent assurance with disclosed redactions, a useful model for any firm asking vendors for evidence behind safety claims. Treat the self-reported 12x uplift as a warning about survey-based productivity claims.

Original: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/