Abstract
For most of AI’s history humans drove every step of its development cycle; Anthropic is now delegating a growing share of that work to AI systems, which is speeding up its own progress. Taken far enough, and given enough compute, that trend points to a system capable of fully autonomously designing and developing its own successor — recursive self-improvement. This has not happened and is not inevitable, but it could arrive sooner than most institutions are prepared for, and the response advocated is not a unilateral pause but the construction of verification systems that would make a coordinated, multi-lab slowdown credible.
Evidence from the outside world
- Task-length trend. Claude Opus 3 (March 2024) handled software tasks taking a human roughly four minutes; Claude Sonnet 3.7 (March 2025) roughly ninety minutes; Claude Opus 4.6 (2026) twelve-hour tasks. The doubling time for reliably completable task duration is now approximately four months, down from a previous seven-month cycle.
- SWE-bench — real open-source codebases and genuine bug reports requiring fixes that pass project tests — went from single-digit scores to saturation within two years.
- CORE-Bench — whether models can reproduce published research findings by rerunning code and data from papers — went from roughly 20% success in 2024 to saturation fifteen months later.
- METR’s measurement of long-duration task completion found Claude Mythos Preview could operate for “at least” 16 hours, hitting the limits of measurement without exhausting capability.
Evidence from within Anthropic
Volume
- Lines of code merged per engineer per day held stable across 2021–2024, began growing in 2025 after Claude Code launched in research preview (February 2025), and accelerated again in 2026 as models began working autonomously for multiple hours. By Q2 2026 a typical engineer merged 8× more code per day than in 2024. The authors flag that this measures quantity rather than quality and is “almost certainly an overstatement of the true productivity gain.”
- As of May 2026, over 80% of merged production code originated from Claude, up from single-digit percentages before the February 2025 Claude Code launch.
- A March 2026 poll of 130 research staff put the median self-reported productivity multiplier from working with Mythos Preview at roughly 4× versus no AI assistance; the authors expect the true uplift in March was somewhat lower.
- In April 2026, Claude shipped over 800 fixes that reduced one class of API error a thousandfold; the supervising engineer estimated the human equivalent at about four years of work.
Quality
- Two dimensions are distinguished: whether code works, and whether it is maintainable.
- The rate at which staff correct, redirect, or take over from Claude has declined steadily over a year, including on complex open-ended problems with no clear specification. Success rates rise across trivial, routine, and substantial task tiers; on open-ended problems Claude reached 76% success in May 2026, up 50 percentage points in six months.
- Example: a training job crashing across tens of thousands of operations, where Claude isolated an obscure flag and delivered the fix in about two hours against a typical human two to three days.
- On readability, opinion varies: many staff judged Claude-authored code inferior to human-written code in late 2025 and at approximate parity now, with superiority anticipated within a year.
- Claude now reviews proposed changes before they merge into the production codebase, catching bugs and security flaws. Retrospective analysis found this automated review would have caught roughly one-third of the bugs behind past production incidents.
Research execution
- A standard test is run for each model release: optimise training code for speed while preserving correctness. Claude Opus 4 (May 2025) achieved roughly 3× speedup; Claude Mythos Preview (April 2026) roughly 52×. Skilled human researchers take four to eight hours to reach 4× on the same task.
- Autonomous research project (published April 2026): Claude-powered agents were given an open AI safety question — can weaker models reliably supervise stronger ones? — and proposed hypotheses, tested them, shared findings, and iterated. Two human researchers over one week recovered roughly 23% of the performance gap between weak supervision and strong-model ground truth; the agents, over 800 cumulative hours and roughly $18,000 of compute, recovered 97%. Caveats: the result does not transfer cleanly to production-scale models, and humans selected the problem and wrote the scoring criteria — direction-setting was the only meaningful human contribution.
- Research session navigation: real Claude Code sessions from January–March 2026 on open-ended investigative problems were mined for moments where a human investigator went down a detour before redirecting. Models were shown only the pre-detour work and asked what to do next. Claude Opus 4.5 (November 2025) beat the human choice 51% of the time; Mythos Preview (April 2026) 64%. The authors note the sample (n=129) deliberately selects suboptimal human moments and is not a parity comparison.
Three levels of capability
- Executing specified tasks — Claude excels.
- Running experiments toward predetermined goals — strong performance demonstrated.
- Choosing which problems matter — a significant human advantage remains, though it is narrowing.
What the future of work at Anthropic might look like
- As human- and AI-authored code quality reaches parity, humans stop writing code and move entirely to review — but if Claude generates code faster than humans can review it, review becomes the organisational bottleneck.
- With execution cheap, the valuable question shifts from running experiments to deciding which experiments deserve running. The “doing” phase costs minimal human time even where compute costs stay substantial.
- Research taste and judgment — problem selection, interpreting results, recognising dead ends — is the current human comparative advantage.
- Amdahl’s Law signatures are already appearing in practice: human code review as a new bottleneck, and an explosion of new ideas, initiatives, tools, and simulations far exceeding organisational capacity to pursue them. Recognising and removing such bottlenecks may become the most important organisational skill.
- Employee testimony is quoted on several sides. One engineer: “I started leaning hard into Claudifying about a year ago. That’s been ~5 months since I last wrote any code myself.” On what is lost: “Work (and life) ran on a gift economy of small favors between humans… [Claude is] faster, it creates zero debt, but each of these is a lost bid for human collaboration.” On residual human advantage: “The comparative advantage of humans as of right now is still in seeing the bigger picture and thinking beyond the confines of the immediate task.” On ambivalence: “On days where everything works well, I can’t help but think nothing I do matters… But then everything breaks and I realize I have no idea what I’ve been up to anymore.” And on the current division of labour: “The shape of stuff today is roughly ‘humans have ideas, and the models are able to implement, test and evaluate them an [order of magnitude] faster than before.’”
What if we’re wrong?
- The work still in human hands — choosing which problems to pursue — may be the most important part, and it is genuinely unclear whether current training methods can unlock it.
- The counter-argument offered: most AI progress comes through incremental iteration rather than breakthrough insight, so much of what advances the frontier is automatable, and research advances depend heavily on tools and resources that enable faster experimentation.
- Even if Claude never develops strong research judgment, humans spending most of their time on direction-setting while Claude handles execution means each researcher steers far more work than before.
- The less conservative reading: research judgment, already improving in early results, may follow the pattern of other qualitative skills where AI systems initially fail and then improve substantially.
Possible futures
1. The trend stalls, but today’s capabilities diffuse widely. Exponentials may be S-curves approaching diminishing returns; research judgment might require architectural breakthroughs beyond Transformers rather than emerging from scale. Supply-chain constraints — chip fabrication, grid expansion, compute availability — could bind before intelligence does, and exogenous shocks could slow progress. The authors do not believe this likely, given that measured capabilities continue following established curves. It would give governments and societies the most time to adapt.
2. Compounding efficiency gains with humans still directing. AI development becomes substantially automated while humans set research directions and judge results; 100-person companies could do the work of 10,000- or 100,000-person organisations. Benefits span knowledge work and government services; risks include authoritarian surveillance and influence operations that tailor manipulation to each individual. Human roles shift toward partnering with AI to scale research and building the systems needed to verify that AI outputs can be trusted. The authors believe the evidence points to this scenario.
3. Full recursive self-improvement. AI systems design and refine their own successors, and the pace of AI progress becomes determined entirely by the availability of compute. Humans shift to oversight, validation, and verification of an expanding “virtual lab” run by AI systems, which could also revolutionise other fields. This is where the authors are least certain, particularly about how the alignment problem gets solved or not: systems might be aligned enough to discover novel alignment solutions or to halt development if needed, or misalignment present in today’s models could compound as models build their successors, growing more frequent but less understood until control is lost. Economic consequences are hard to predict if human labour stops being competitive, and robotics may follow a similar pattern. Amdahl’s Law constraints persist regardless: more intelligence “can’t learn what a drug does over decades of use, can’t hold elections sooner than a constitution dictates, and can’t turn a stranger into an old friend in a weekend.”
What should we do?
- Slowing frontier development would likely be a good thing, but a unilateral slowdown could let less cautious actors catch up and so decrease safety overall. A unilateral pause by one lab is achievable immediately and accomplishes much less: it would change who the front-runner is without creating the wider deliberative process that is currently missing.
- The proposal is to build the systems a credible slowdown or pause would require — so that frontier developers can verify that others globally have actually stopped or slowed, and that a bad actor cannot use a coordinated slowdown as cover to jump ahead in secret.
- A meaningful slowdown would require multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions, each able to verify the others have actually stopped. A credible pause also requires specifying what triggers it, what lifts it, and who adjudicates.
- Detectability is harder than for other technologies: training runs are far easier to conceal than missile silos, the inputs are general-purpose, and the incentive to defect quietly is enormous because whoever continues while others pause could inherit the lead. Nuclear verification regimes took decades to build both the infrastructure and the trust — “We don’t have that long.”
- Anthropic’s stated commitment: if such systems existed, it expects it would slow down or temporarily pause, provided other developers at or near the frontier did so in a verifiable manner.
- Next step: in the coming months, convening conversations where policymakers, researchers, civil society, and other AI companies help answer the questions raised — especially around full recursive self-improvement and how to create better options for coordination and deliberation — with the results to be published. People outside AI companies should be involved in the deliberation.