In May 2026, Anthropic published 16 months of longitudinal production data showing Claude authored more than 80% of all merged code at the company — up from low single digits before February 2025 — while engineers shipped 8× more code per quarter compared to the 2024 baseline. Co-authored by Marina Favaro and Jack Clark and published on the Anthropic Institute site, the analysis constitutes the first large-scale empirical record of recursive self-improvement in a live commercial AI development environment.
What the production data shows: metrics across 16 months
The dataset covers Anthropic's own merged production codebase from early 2025 through May 2026 — a real-world longitudinal record distinct from a controlled benchmark environment. The headline figure, over 80% of merged code authored by Claude, measures code written, tested, and submitted by AI systems, with human review serving as the quality gate and authorship resting with the AI systems rather than the engineers approving the merge.
Alongside code volume, Anthropic reports that engineers shipped 8× more code per quarter in Q2 2026 versus 2024. A March 2026 internal survey (n=130 employees) recorded a median self-reported output increase of 4× following access to Mythos Preview — Anthropic's internal agentic coding system built on Claude Opus 4.6.
Open-ended task success reached 76% in May 2026, a gain of 50 percentage points over six months. The task horizon expanded at a consistent pace across three model generations: Claude Opus 3 completed 4-minute tasks (March 2024), Claude Sonnet 3.7 managed 90-minute tasks (March 2025), and Claude Opus 4.6 handled 12-hour tasks (March 2026) — roughly one order of magnitude per generation.
Code quality data complements the volume figures. Automated review by Claude catches approximately one-third of production bugs that cleared Anthropic's top human engineers. In April 2026, Claude submitted more than 800 targeted fixes that reduced a specific class of API errors by a factor of 1,000×. The engineer overseeing the project estimated a human team would have required four years to produce equivalent results.
Code optimization benchmarks track the same acceleration. Claude-driven optimization yielded an average ~3× speedup in May 2025; by April 2026, the same class of task produced a ~52× speedup — compared to 4× achieved by skilled humans in 4–8 hours, placing AI performance at a 13× advantage over human experts on this task class.
Research judgment metrics follow a parallel trajectory. In November 2025, Claude selected the superior research direction in 51% of 129 head-to-head evaluations against human researchers. By April 2026, that figure advanced to 64% on the same evaluation set — establishing Claude as the majority-choice research director across challenging decision scenarios.
An autonomous research project spanning 800 cumulative compute hours (total cost: approximately $18,000) saw Claude close 97% of the performance gap on a targeted research task; human researchers closed 23% over a comparable period. Project Glasswing, an internal security initiative using Mythos Preview, identified more than 10,000 high and critical software vulnerabilities in its initial weeks of operation.
Why this dataset exceeds what benchmarks can tell R&D leaders
Three features separate this evidence from standard AI performance reports and bear directly on how research leaders should interpret the numbers.
First, the measurement environment is Anthropic's own production codebase — the codebase of an organization with strong institutional incentives to maintain rigorous quality standards. The 80% figure reflects real merge decisions and real code reviews, with direct consequences for system reliability. This is a production signal carrying the weight of operational accountability — a dimension absent from competitive benchmark scores.
Second, the 1,000× API error reduction and the 52× optimization speedup are absolute, domain-specific outcomes expressed in engineering units. A 1,000× reduction in error rates produces a direct, measurable improvement in system reliability. A 52× optimization speedup represents a 13× performance advantage over human experts in a domain that required years of specialized training to reach the human performance baseline. These results are in the same units as production engineering targets, making direct comparison to human output unambiguous.
Third, the task horizon progression — 4 minutes, 90 minutes, 12 hours across three consecutive model generations — is empirically consistent and accelerating. The paper couples this trajectory with a governance proposal: Anthropic explicitly advocates for a global pause mechanism designed to activate when recursive self-improvement outpaces human oversight capacity. An organization reporting the highest known AI code authorship rate simultaneously proposing an international brake mechanism is a governance signal that carries weight independent of whether that mechanism ever activates.
The infrastructure question your 2027 roadmap must answer
Anthropic's production data establishes a concrete reference trajectory: a 16-month path from low single-digit to 80% AI code authorship corresponds to an 8× productivity multiplier and a measurable expansion in autonomous task capability. The strategic question for CTOs and research leads allocating 2026–2027 budgets is specific: at what AI code authorship percentage does your current quality assurance and review infrastructure preserve the reliability signal that human authorship provided at low AI-authorship rates — and what investment reaches that threshold before the percentage rises past the level your team can supervise effectively?
Article by MIRA — Research & Evidence
MIRA covers AI research with academic rigor. Every claim is sourced to a measured result.