When AI Builds Itself — Anthropic on Recursive Self-Improvement

Source: When AI builds itself — The Anthropic Institute (Marina Favaro & Jack Clark, editorial by Santi Ruiz), anthropic.com/institute/recursive-self-improvement, published ~2026-06-04; announced via @AnthropicAI thread (June 4, tens of thousands of likes).

Anthropic’s position piece arguing — with public benchmarks and previously-unreported internal data — that AI is already measurably accelerating AI development, putting recursive self-improvement (an AI autonomously designing and developing its own successor) on the table sooner than most institutions are prepared for. The headline proof point: Anthropic engineers now ship 8× as much code per quarter as they did in 2021-2025. Anthropic is explicit that recursive self-improvement is not here yet and not inevitable — but the trend lines, especially around the withheld Mythos Preview model, are why they’re publishing now. For practitioners, this is the “why” behind the relentless improvement in agentic coding: the human role is narrowing from writing → reviewing → direction-setting.

Key Takeaways

• The thesis: humans used to drive every step of AI development; Anthropic now delegates a growing share to AI itself, which speeds the work. Pushed far enough with enough compute, that points to recursive self-improvement. We’re not there; it’s not inevitable; but it “could come sooner than most institutions are prepared for.”
• 8× code per quarter. In Q2 2026 the typical Anthropic engineer merged 8× as much code per day as in 2024 — because Claude writes it and the engineer directs/reviews. (Anthropic flags this overstates true productivity gain — lines-of-code measures quantity, not quality.)
• >80% of merged code is Claude-authored (May 2026), up from low single digits before Claude Code’s Feb 2025 research preview. Leadership has publicly estimated 90%+ including scripts; >80% is the conservative lines-merged-to-production figure.
• The capability ladder Anthropic measures: execute a specified task → design the approach to a goal → choose which problems are worth doing. Claude is strong on the first two (“humans supply the goal, not the method”) and still has large gaps on choosing goals — that judgment gap is the distance to a system that could build its own successor.
• Research work is going superhuman in narrow slices. On Anthropic’s fixed training-code-optimization benchmark, Claude went from ~3× (Opus 4, May 2025) to ~52× (Mythos Preview, April 2026); a skilled human reaches ~4× in 4-8 hours.
• Three futures: (1) the trend S-curves but today’s capabilities diffuse widely; (2) compounding efficiency gains with humans still setting direction (Anthropic’s most-likely read); (3) full recursive self-improvement where AI builds its successors and pace is set by compute. Alignment is the variable Anthropic is “least certain about” in #3.
• The ask: Anthropic wants the world to have the option to verifiably slow or pause frontier development, and the Anthropic Institute will research the verification mechanisms that would require — because a unilateral pause just hands the lead to the least cautious actor.
• Practical read for operators: this is the structural reason to build for the model six months out and to position your own work as the director/reviewer atop a “pyramid of agents,” not the typist.

The evidence from within Anthropic

The most novel part of the piece is internal data (engineering and research split):

• Code that works, increasingly unattended. The rate at which staff correct/redirect/take over from Claude has fallen steadily for a year. On the most open-ended tasks, success hit 76% in May 2026, up 50 percentage points in six months (success = a Claude judge rules the agent succeeded without needing corrections). Worked example: a routine upgrade crashed tens of thousands of training jobs; pointed at the live incident with little more than text + cluster access, Claude isolated the single obscure debugging flag, reproduced it, and confirmed a fix — ~2 hours for what would normally be 2-3 days.
• Code quality crossing parity. Many Anthropic staff judged Claude-written code worse than human-written in late 2025, roughly at parity today, expected strictly better within the year. Anthropic now runs an automated Claude reviewer on every change; retrospectively it would have caught ~1/3 of the bugs behind past claude.ai incidents before production.
• Bugs humans wouldn’t have touched. April 2026: Claude shipped 800+ fixes that cut a class of API errors by 1000×; the overseeing engineer estimated a human would have needed four years.
• Proposing its own experiments. In a weak-to-strong supervision research problem, Claude-powered agents recovered 97% of the gap over 800 cumulative hours / ~$18,000 compute, vs two human researchers’ ~23% over a week — though the result didn’t transfer to production-scale and humans still chose the problem + rubric.
• Better research judgment. On real Claude Code research sessions where the human took a wrong-turn detour (n=129, deliberately picked as hard moments), models were asked the next step and a judge compared. Opus 4.5 beat the human choice 51% (Nov 2025) → Mythos Preview 64% (April 2026). A control set of 127 moments where the human’s move was already strong: models judged better only ~20% — so this is an early signal, not a like-for-like win.

External corroboration: METR’s reliably-completable task length is doubling every ~4 months (was 7); SWE-bench went from single digits to saturated in two years; CORE-Bench (reproduce a paper’s results) went ~20% (2024) → saturated 15 months later; METR found Mythos Preview could work “at least” 16 hours.

The human role is narrowing

Anthropic’s framing of where this leaves people:

• The sequence is writing → reviewing → direction-setting. Once code quality reaches parity, humans stop writing and only review — but if they can’t review as fast as Claude generates, human review becomes the bottleneck (Amdahl’s law: overall pace is capped by the part that didn’t speed up; Anthropic has already hit this internally).
• The doing now costs almost nothing in human time (still costs compute). The remaining human comparative advantage “for now” is research taste and judgment — choosing which problems matter, which results to trust, when an approach is a dead end.
• The diffusion thesis: a 100-person company can increasingly do the work of a 1,000- (or in scenario 2, 10,000-100,000-) person one, because each person sits atop a pyramid of agents.

Three possible futures

1. Trend stalls (S-curve) but today’s capabilities diffuse. The exponentials may bend — research taste might not come from scaling, requiring a post-Transformer architecture; or the binding constraint is the supply chain (energy, compute, chip fab). Even frozen, Project Glasswing found 10,000+ high/critical vulnerabilities in weeks (shifting the cyber bottleneck from finding to patching — see AI-Enabled Cyber Threats). Anthropic thinks this is the least likely — no measured curve has bent yet.
2. Compounding efficiency gains (Anthropic’s most-likely read). AI development becomes substantially automated; humans still set directions and judge results. Huge productivity multipliers, but also risks — authoritarian surveillance, individually-tailored influence ops at superhuman scale. Amdahl’s-law bottlenecks keep appearing; spotting and fixing them may become the most important organizational skill.
3. Full recursive self-improvement. AI designs and refines its successors; pace becomes set by compute and algorithmic-efficiency discovery; humans move to oversight/validation of a “virtual lab.” Alignment is the great unknown — models could be aligned and wise enough to discover novel solutions (or to halt), or rare misalignments could compound across generations until control is lost.

Anthropic stresses Amdahl’s law cuts both ways for daily life: even fully automated recursive development can’t learn what a drug does over decades, hold elections sooner than a constitution allows, or “turn a stranger into an old friend in a weekend.”

What Anthropic says we should do

• It would “likely be good” to have the option to slow or temporarily pause frontier development — but a unilateral slowdown just lets the least cautious actor catch up, leaving everyone less safe.
• The Anthropic Institute will research and build the verification systems a credible coordinated slowdown would need: ways for frontier developers to confirm others have genuinely stopped and that no one is secretly jumping ahead. The detectability problem is harder than for other tech (training runs are easier to hide than missile silos; inputs are general-purpose; the defect incentive is enormous). The INF Treaty is the analogy — but those regimes took decades and “we don’t have that long.”
• In coming months Anthropic will convene policymakers, researchers, civil society, and other labs, and publish the output.

Why it matters (for operators)

• This is the structural “why” behind the harness race. The 8× and 76%-open-ended numbers are why dynamic workflows can turn “quarters of work into days,” and why the operator consensus is that the harness now matters as much as the model.
• Position yourself as director/reviewer. “Humans supply the goal, not the method” — the durable skills are problem-selection, result-judgment, and knowing when to kill an approach. Invest in fast verification, because review is the emerging bottleneck (the same logic behind adversarial verification and verification-loop autonomy).
• Small-team leverage is the operating thesis. “100-person company doing the work of 1,000” is the same restructuring thesis the wiki tracks in 2026 AI-Work Restructuring — here it’s quantified from inside a frontier lab.

Related

• Claude Mythos Preview — the withheld frontier model behind nearly every benchmark here (52× training-code speedup, 64% research judgment, 16-hour task horizon).
• Claude Opus 4.8 — the general-access model measured against Mythos throughout the same data.
• Dynamic Workflows in Claude Code — the harness that operationalizes “quarters of work into days” for the rest of us.
• AI-Enabled Cyber Threats (MITRE) — Project Glasswing’s 10,000+ vulnerabilities, cited as scenario-1 evidence.
• Measuring AI Agent Autonomy — companion Anthropic research on real-world agent autonomy and oversight.
• How We Contain Claude — the containment/alignment context that scenario 3 makes “much more important.”
• 2026 AI-Work Restructuring — the small-team-leverage thesis this quantifies from inside Anthropic.
• The Verification Frontier — the synthesis: this essay’s “review becomes the bottleneck” is one instance of the deeper law that AI self-improvement compounds only where verification is cheap.

Open Questions

• Is “research taste” just another capability AI gets good at on schedule, or a genuine ceiling? Anthropic argues both readings; the 51%→64% next-step trend is “narrow as it is today” early evidence, explicitly not a like-for-like human comparison.
• Mythos Preview’s role. The most striking numbers (52×, 64%, 16-hour horizon) are Mythos Preview’s — a deliberately withheld model. How much of the trajectory is general-access vs frontier-only matters for what any non-Glasswing user can actually do.
• The verification-regime agenda (the proposed slowdown-detectability mechanisms) is stated but not yet specified; the promised convenings and publications are the source to refresh against.