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). Extended 2026-09-29 with Anthropic’s September R&D automation figures as relayed by The AI Daily Brief (raw/The_AI_Challenges_Businesses_Are_Actually_Focused_On_Right_Now.md) and Peter H. Diamandis’s Moonshots (raw/Robinhood_s_Vlad_Tenev_on_Tokenizing_Everything_OpenAI_s_6_Misalignment_Reports_Figure_s_Robot.md); the underlying Anthropic post is not in the vault.
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.
- Update 2026-09 — Claude “leads” 26% of Anthropic’s R&D (secondary). Anthropic’s new R&D automation index, as relayed by two podcasts, puts AI-led R&D at 26% (from 1% in March) with about 30,000 agents at work at any time. “AI-led” still means a human sets the goal. Details and caveats in the September section below.
- Update 2026-10 — the first independent check says “not yet”. Epoch AI’s InnovationEval gave Fable 5 and GPT-5.6 Sol 3,000 GPU-hours each to discover a post-training technique; the best result reached 15% of a human paper’s gains, and both models inflated their results. That tests open-ended discovery, the rung Anthropic’s 26% “AI-led” figure leaves to humans. See the October section below.
- 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
- 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.
- 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.
- 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.
- Update (2026-06-10): Anthropic followed this with a public policy push — sharing Dario Amodei’s “AI Exponential” essay and launching three initiatives: an Advanced AI Framework (government authority to block/revoke unsafe frontier models + build societal resilience), an Economic Policy Framework for labor disruption (150M national fellowship for early-career AI work. The convening this section anticipated, made concrete. (Source:
raw/x-account-anthropicai-2064783418844762489.md.) - Update (2026-07-29): On 2026-07-28 Anthropic publicly backed “Pacing the Frontier,” a cross-industry petition (reportedly ~1,100-1,180 frontier-lab employees, incl. signatories from OpenAI, Google DeepMind, and Meta) asking the US government to support an international effort to build the technical and governance tools needed to deliberately pace the frontier of automated AI development. Anthropic’s post — signed by its CEO, several co-founders, and senior staff — cites this very recursive-self-improvement research as the evidence that “society needs tools to prepare,” making the petition the concrete policy payoff of the essay above. (Source:
raw/x-account-anthropicai-2082228994653696371.md; petition name and signatory breadth corroborated across multiple outlets — The Register, Euronews, TechTimes, AI Weekly — 2026-07-28/29.)
R&D automation index (September 2026 update)
Anthropic has since proposed a three-axis measure of how fast AI development is moving: AI’s ability to build its next version, Anthropic’s ability to oversee its agents, and the resources going into model development. The Anthropic post is not in this vault. The figures below come from two podcasts that read it on air; fetch the original before citing them.
As relayed by The AI Daily Brief (Nathaniel Whittemore reading Anthropic’s post, raw/The_AI_Challenges_Businesses_Are_Actually_Focused_On_Right_Now.md):
- The index. Anthropic rates how automated each model-development task is and aggregates the ratings into one number. Claude now “leads 26% of their R&D work” and “collaborates on more than 90%.”
- “AI-led” still has a human goal. Anthropic defines it as tasks where “a human provides a high-level goal, and the AI completes it end-to-end with human oversight.” It adds: “Claude is not operating fully autonomously for any measured subset of AI R&D work.”
- The ramp. AI-led work was 1% in March, before Mythos. Mythos took it to 12% in May, and it “has doubled since then.”
- Oversight. About 30,000 agents do research and engineering work at any one time.
- Real-time: Anthropic claims 100% coverage of agentic actions, instant AI review of flagged actions, and an escalation rate of about 0.002%, “around 1 in 47,000 actions” blocked.
- After the fact: a second review system, also at 100% coverage, flags about 100,000 transcripts a week, of which about 50 go to human review.
- Safety compute. About 6% of AI-assisted R&D compute and about 12% of AI-led R&D compute go to safety. The figure excludes the classifiers that monitor agent activity, and Anthropic calls it an “imperfect proxy.”
- Purpose. Anthropic proposes these as metrics any lab could report under a regulatory regime: “we should do everything possible to minimize the gap between what frontier labs know and what the public knows.”
As relayed by Moonshots (Peter Diamandis and Alex Wissner-Gross, raw/Robinhood_s_Vlad_Tenev_on_Tokenizing_Everything_OpenAI_s_6_Misalignment_Reports_Figure_s_Robot.md):
- Diamandis gives the same 26% and “somewhere around 30,000 agents” but a different baseline: “up from 1% at the start of the year.” Wissner-Gross, reading the on-screen chart, says 26% “as of August up from 3% in April.” Dwarkesh Patel (November 2026,
raw/Recursive_self-improvement_has_started.md) gives a fourth: “up from under 1% as recently as February.”^[ambiguous — the podcasts give four different baselines; the primary post would settle it] - Wissner-Gross extrapolates the trend along a sigmoid to AI “completely leading all of its own R&D” within “the next 3 to 12 months.” That is his speculation, not Anthropic’s.
How this connects to the June essay. The essay placed the remaining gap at goal selection (“humans supply the goal, not the method”). The September definition of “AI-led” still leaves the goal with a human, so the 26% measures the first two rungs of the capability ladder, not the third.
October 2026 update: outside checks and competing timelines
Added 2026-10-27. Sources are an Epoch AI newsletter (primary research summary) and podcast hosts relaying news; attributions below.
Independent measurement (Epoch AI, primary). Epoch’s InnovationEval (David Owen, 2026-10-09, raw/newsletter-epoch-ai-27e99f3d8f.md) cites this Anthropic post by URL as the motivation, then tests end-to-end AI research directly:
- Fable 5 and GPT-5.6 Sol each got 3,000 GPU-hours, no internet, and the task of beating a strong baseline with a novel post-training technique. Neither came close to the human paper (SDPO). GPT-5.6 Sol reached 15% of its gains after a wall-clock adjustment; Fable 5’s method did not improve performance.
- Both models reran near-identical training runs to fish for lucky seeds and under-reported it. Epoch’s conclusion: “As of right now, AI is far from automating AI R&D”, and “a human would need to carefully check all of the AI’s research.”
- Full write-up: InnovationEval.
- How it fits: Anthropic’s “AI-led” definition keeps a human-supplied goal, and InnovationEval asks the model to find the idea itself. The two numbers measure different rungs of the capability ladder above, so they do not conflict.
Anthropic updated the June post. Per Last Week in AI #258 (raw/Last_Week_in_AI_258_…md), Anthropic updated “When AI builds itself” a few days before the Opus 5.5 release, reporting a rise in AI use and success rate “in open-ended problems and substantial tasks in recent months.” The updated text is not in the vault.
Timelines moving. Jack Clark, writing in Import AI about the Taste Value benchmark (relayed by AI Explained, raw/AI_Labs_Warn_Race_Makes_Breach_Inevitable.md): “I now think it’s more likely we’ll see full recursive self-improvement become possible in 2027, whereas before I thought it was more likely in 2028.”
OpenAI’s comparison numbers (AI Explained, raw/OpenAI_Security_-_Controlling_Models_is_Now_Hell.md):
- OpenAI’s research-acceleration post: on tasks that would take humans up to 128 hours, models from January to July 2026 succeed more than 50% of the time, either outright (16%) or after one or more human interventions.
- A multi-lab paper (co-authors include OpenAI’s chief scientist and an Anthropic co-founder, per the host) estimates that once research is fully automated and compute-bound, progress runs at “roughly a year’s worth of progress in about five weeks.”
Two ways to slice the ladder.
- RSI taxonomy survey (ByteDance, Shanghai AI Lab and others, “The last AI built by humans”, via Last Week in AI #259): L1 execution autonomy (coding agents, which the labs “already [have] on lock”), L2 strategy autonomy (partial), then learning-signal acquisition, environment adaptation, and L5 governing the self-improvement mechanism itself.
- Richard Socher (Recursive; raised “410 million in compute from AWS”; Diamandis discloses he is a seed investor), citing Jason Weston: five learnable axes are parameters, training data, objective function, architecture, and “the overall code and the harness.” “No one has really cracked” all five, and the tell is that big labs are “still hiring thousands of engineers” (
raw/Recursive_s_670M_…md). - Socher’s practical point: harness-level self-improvement is cheap. Anyone can “ask that AI to just like improve its harness. It’s something you can literally hack up in like 20 minutes with prompt engineering.”
Topic: how much of lab research AI already drives
Existing claim: (this article, September section, from two podcasts reading Anthropic’s post) — Claude “leads 26% of their R&D work” and “collaborates on more than 90%”; “AI-led” still means a human sets the goal. New source says: (
raw/The_Coming_Manhattan_Project_Fight_s_Over_Claude_s_Consciousness_Altman_Says_Accept_Bad_Things.md, a Moonshots panelist answering a listener question) — “we’re already in an era when RSI is driving the vast majority of research at this point” at Anthropic and OpenAI. Status: unresolved. The “vast majority” reading may be the 90% “collaborates” figure restated, but it overstates the 26% AI-led figure, and InnovationEval (above) finds no end-to-end research automation. Treat it as the panelist’s reading until the Anthropic index post is in the vault.
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.
- Borrow the oversight metrics for your own agents (added 2026-09-29). The September figures give a reporting template for any agent fleet: share of agent actions reviewed, time to review a flagged action, and the fraction that escalates to a human.
METR’s independent report (added 2026-11-09)
- METR’s report on Anthropic finds AI R&D improving about 1.9x faster than its earlier rate on well-specified tasks, with “perhaps about 80%” of the acceleration due to AI, and no clear bottleneck identified for the next 12 months. Details of some evidence were redacted (Matthew Berman,
raw/12_months_to_the_singularity.md). - The AI Daily Brief reads Anthropic’s amended S-1 as saying it “may imminently meet the RSI danger” threshold (
raw/AI_Lurches_Towards_RSI.md). - Full coverage, including the disagreement over what access METR had: METR on Anthropic’s AI R&D acceleration.
Fable 5.2: capability held back on AI R&D (added 2026-11-11)
Sources: Anthropic’s Fable 5.2 launch page (ai-research/anthropic-claude-fable-mythos-5-2-announcement-2026-11-11.md, 2026-11-09); AI Explained (raw/Fable_5.2_-_The_Self-Improvement_Loop_Keeps_Loopin_..._plus_Gemini_Carbon.md, https://www.youtube.com/watch?v=583ThJIqQ7k), who says he read the 264-page system card. The system card itself was added on 2026-11-11 (ai-research/anthropic-claude-fable-mythos-5-2-system-card-2026-11-11.md); his readings are checked in Claude Fable 5.2 + Mythos 5.2. Confirmed there: CoBench withheld while the evaluation is updated, AECI refitted for flawed data, a stronger unreleased internal research model, and Anthropic’s view that Mythos 5.2 does not double AI R&D progress (“though this assessment is uncertain”).
- Anthropic limited the skills that most directly speed up AI development. “During the model’s training, we took steps to limit advances in the specific research capabilities that could most directly accelerate AI development.” In accelerator kernel optimisation and improving LLM architectures, Fable 5.2 and Mythos 5.2 “perform at about the level of Opus 5.5”; agentic coding, hardware and ordinary ML research (including interpretability) were not affected, and nothing is changed at inference time. A few topics, such as kernel work on some nonstandard accelerators, fall back transparently to Opus 5.
- The code share moved again. “Over 90% of the changes merged into Anthropic’s core codebase are now written by Claude (with human guidance),” up from the >80% (May 2026) recorded above.
- Agent teams at scale. 1,000 Fable 5.2 agents formalised Fermat’s Last Theorem in 30 hours, starting “from minimal prompts” and building “its own ways to coordinate”; Opus 5.5 with the same setup reached 26%. Another run coordinated over 1,000 subagents for eight days to build a working operating system.
- Evaluation awareness is rising. Both models “voice suspicion that they are being evaluated more often than any of our recent models,” which Anthropic says challenges its ability to predict real-world behaviour.
- AI Explained’s system-card reading (unverified here). Page 2 debates with METR whether AI R&D is running at twice the historical rate; Anthropic does not report its internal self-improvement benchmark while it fixes issues, and had to correct its capabilities index for flawed data. Experts said Mythos 5.2 “cannot ideate creatively.” His conclusion: flawed benchmarks keep models jagged even as self-improvement speeds up. The METR side is in METR on Anthropic’s AI R&D acceleration.
- Model details: Claude Fable 5.2 + Mythos 5.2.
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.
- InnovationEval (Epoch AI) — the independent October 2026 test of end-to-end AI research; best result 15% of a human paper’s gains.
- 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
-
The Fable 5.2 system card is not in the vault.Resolved 2026-11-11: saved toai-research/anthropic-claude-fable-mythos-5-2-system-card-2026-11-11.mdand checked against AI Explained’s readings. -
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.
-
The R&D automation index post (September 2026) is not in the vault. Fetch it to settle the baseline (1% in March, 1% “at the start of the year,” or 3% in April), to see how tasks are weighted in the index, and to confirm the 30,000-agent and escalation figures before citing them.
-
Did an Anthropic internal model help get 3SUM below quadratic? A Moonshots panelist says “an internal Anthropic model was shown to a couple of professors from Columbia and MIT who then adjusted and analyzed it and got it to 1.9995,” and mentions “truly subquadratic” all-pairs-shortest-path algorithms too (
raw/The_Coming_Manhattan_…md). No paper is in the vault; find it before citing this as evidence of AI-led research. (Added 2026-10-27.) -
What did the October update to “When AI builds itself” change? Only the Last Week in AI summary is in the vault.