Anthropic Reveals with Data the True Speed at Which AI Is Accelerating Its Own Evolution
Alina Collins
Anthropic has published internal data for the first time: the length of tasks AI can complete autonomously now doubles every four months, and over 80% of its codebase is written by Claude. This means → AI is accelerating its own development cycle, and fully recursive self-improvement may arrive sooner than most expect.
How fast is AI's autonomous work window growing?
In March 2024, Claude could handle a dev task lasting about 4 minutes. A year later, that ceiling reached ~1.5 hours. By April 2026, it was tackling 12-hour complex tasks.
The doubling rate itself is accelerating: from roughly every seven months to roughly every four months.
This means → at the current slope, tasks requiring days of human work could enter AI range by late 2026; week-long tasks could follow in 2027.
Where does AI stand on standardized benchmarks?
SWE-bench — the industry standard for real-world software engineering ability — saw AI scores in the low single digits two years ago; they are now near saturation.
CORE-Bench — which tests AI's ability to reproduce published research, a prerequisite for original scientific work — went from about 20% in 2024 to saturation within fifteen months.
In plain terms = these two tests are roughly "software engineer certification" and "research assistant qualifying exam." AI has essentially aced both.
How has the way Anthropic's own engineers work changed?
As of May 2026, over 80% of merged code in Anthropic's codebase is written by Claude — up from single digits before February 2025.
By Q2 2026, a typical engineer's code output was 8× what it was in 2024. Anthropic proactively flagged that lines of code is an imperfect metric; 8× "almost certainly overstates real productivity gains."
A cross-check: an internal survey of 130 researchers in March 2026 found median self-reported output was roughly 4× what it would be without AI.
One concrete case: in April 2026, Claude completed over 800 fixes, reducing a class of API errors by a factor of 1,000. The supervising engineer estimated the same work would have taken a human four years.
How close is AI's research judgment to a human's?
Anthropic tracks this with an internal test: find moments in real researcher sessions where the human took a wrong turn, ask Claude what the next step should be, then judge whose choice was better. In November 2025, Opus 4.5 beat the human 51% of the time; by April 2026, Mythos Preview reached 64%.
On well-defined experimental execution, the gap is wider: a skilled human researcher needs four to eight hours to achieve roughly 4× speedup; Claude Mythos Preview achieved roughly 52×.
This means → on "how to do it," Claude has moved from assistant to superhuman system. On "what to do," humans still lead — but the margin is narrowing.
Fully autonomous research: does AI still need humans?
In an April 2026 case study, a swarm of Claude-driven agents was given an open problem in AI safety. They autonomously proposed hypotheses, designed experiments, shared findings across agents, and iterated.
The comparison: two human researchers spent one week and recovered 23% of the performance gap; the agent swarm used roughly 800 cumulative compute-hours at a cost of ~$18,000 and recovered 97%.
The humans' only substantive role in the entire process: choosing the problem and setting the scoring rules. In plain terms = humans wrote the exam; AI took it — and scored far higher than the humans themselves.
What comes next?
Anthropic outlines three scenarios: stall-and-diffuse (progress plateaus somewhere — lowest probability, but gives society the most time to adapt), compound acceleration with humans retaining directional control (the likeliest path now — a hundred-person company could do the work of a hundred thousand), and fully recursive self-improvement (AI autonomously builds its successors; development speed is bounded only by compute supply).
This reflects Anthropic's own assessment: the second scenario is already underway; the third is no longer science fiction — it is a possibility that demands serious attention.
Under the second scenario, knowledge work and government services face deep restructuring, but the same capabilities could also enable mass authoritarian surveillance or personalized manipulation. This means → acceleration itself is neutral; the direction depends on who steers — and whether oversight mechanisms keep pace.
Content is for reference only, not financial advice.