Recursive self-improvement (RSI): Anthropic's definition, its own numbers, and what the labs officially said in September 2026
Anthropic defines RSI as "a model fully autonomously building its successor" and states that as of August 2026 no measured subset of its AI R&D runs fully autonomously; on 2026-09-06 OpenAI said it had reached its automated research intern goal, with an automated AI researcher as the next node in March 2028. Sakana AI announced an RSI Lab, and Schmidhuber joined on 2026-09-24. Every figure here is the vendor's own tally, and each one carries its source and check date.
Updated 2026-09-29
Four self-reported numbers
Claude's lead share
Anthropic's own metric: Claude leads on 26% of its AI R&D work (through August 2026, self-measured).
Collaborates or above
Same page, verbatim: "The share of work at or above “AI collaborates” is above 90%" — same data window.
Code volume
Anthropic says its engineers on average ship 8x as much code per quarter as from 2021-2025 (own figure).
OpenAI's next target
OpenAI claims strong progress toward an automated AI researcher by March 2028 — a target date, not one met.
What RSI actually means
Anthropic's verbatim definition is recursive self improvement, "a model fully autonomously building its successor". On that reading, work where a human still supervises, or AI only leads part of the chain, does not count. The dispute is the definition itself: per the AP report of 2026-09-19, some companies call any feedback from AI on model improvement RSI, others require full autonomy — settle the definition before judging any already-achieved claim.
What each lab said in September 2026
Anthropic (data through August 2026): Claude leads 26% of its AI R&D work, more than 90% of the work sits at AI collaborates or above, and no measured subset runs fully autonomously. OpenAI (2026-09-06): says it has reached the automated research intern goal, heading for an automated AI researcher in March 2028. Sakana AI: announces the establishment of an RSI Lab, with Schmidhuber joining as Chief Scientific Advisor on 2026-09-24.
Official-wording timeline
The data window of Anthropic's measurement page (through that month, no publication date printed): Claude leads 26% of its AI R&D, more than 90% is at AI collaborates or above, and no subset is fully autonomous.
OpenAI publishes two official articles the same day: the automated research intern goal is reached, with strong progress toward an automated AI researcher by March 2028. Chief scientist Pachocki, writing "An Alien Mind", expects this pace to hold into recursive self-improvement and is concerned no one is prepared.
Sakana AI announces that Jürgen Schmidhuber is officially joining as Chief Scientific Advisor, alongside his current positions, and will be involved in its Recursive Self-Improvement (RSI) Lab.
Confirmed vs unverified
Confirmed (official wording, dated)
Confirmed: Anthropic officially states the RSI definition and writes "Claude is not operating fully autonomously for any measured subset of AI R&D work" for August 2026; OpenAI said on 2026-09-06 that its automated research intern goal has been met, with March 2028 the next node; Sakana's page announces its RSI Lab but prints no founding date, and Schmidhuber joined it on 2026-09-24.
Unverified / not said officially
Unverified: that RSI is already achieved, or that AI builds its own successor unaided — this circulates as rumor, no official statement says so, and Anthropic's wording is the opposite. Reading OpenAI's March 2028 automated researcher as RSI is speculation; its text goes only as far as "strong progress". Sakana prints no founding date for its RSI Lab, so this page does not guess one. Warning posts on X are not evidence and are not repeated here.
Two definitions, two kinds of progress
Strict reading: full autonomy
On Anthropic's definition (a model fully autonomously building its successor), its measurement says no measured subset runs fully autonomously — 26% is leads, not autonomy.
Loose reading: any feedback
Per the AP report of 2026-09-19, the looser reading counts any feedback from AI on model improvement as RSI — hence the already-achieved claims. Same word, two different truths.
These numbers are not audited
Anthropic's own line in that piece is "we should do everything possible to minimize the gap between what frontier labs know and what the public knows" — it treats the gap between what labs hold and what the public sees as something to shrink. OpenAI's chief scientist grounds his judgment the same way: "Based on internal results". Every percentage and multiple here was measured and published by the lab itself; as of 2026-09-29 none of it has been audited by a third party.
On QCode
This page makes no availability promise about any self-improving system — no lab has published such a product, with a model ID and a price to go with it. What is directly relevant here is that the labs themselves are heavy users of coding agents, and as of 2026-09-29 the frontier models driving them are callable on QCode: claude-opus-5-5 and gpt-6-sol.
Recursive self-improvement FAQ
What does recursive self-improvement mean?
The strictest official definition is Anthropic's: a model fully autonomously building its successor. The looser reading (per the AP report of 2026-09-19) counts any feedback from AI on model improvement — progress under the two is not the same thing, so check the definition first.
Has RSI already been achieved?
No lab has announced that. Anthropic writes "Claude is not operating fully autonomously for any measured subset of AI R&D work" for August 2026; per CNBC on 2026-09-11, OpenAI and Anthropic have both said in recent months that autonomous model improvement is happening faster than they expected — but faster than expected is not already done.
What does Anthropic's 26% refer to?
It is the share of AI R&D work at AL4, the leads rung, whose wording on that page is "In AL4, AI “leads”: it can complete most of the task end-to-end from a high-level prompt, while the human supervises." The figure is Anthropic's own measurement, self-reported and, as of 2026-09-29, unaudited.
How far along does OpenAI say it is?
Its official articles of 2026-09-06 say the automated research intern goal is reached: a system that carries out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. For the March 2028 automated researcher it claims only strong progress. Chief scientist Pachocki expects RSI to be at the very core of future scientific discovery — all from OpenAI's own articles, unaudited.
When was Sakana's RSI Lab founded?
The official page announces "the formal establishment of the Sakana AI RSI Lab" and prints no founding date, so this page does not guess one. What is confirmable is the 2026-09-24 announcement: Schmidhuber joining as Chief Scientific Advisor, alongside his current positions, and involved in that lab.
What does this mean for developers today?
The labs became the heavy users first: OpenAI says that as of mid-August 2026 its research organization uses 3.1 agent-workdays of effort for every workday of human labor, and Anthropic's figures are shares of R&D work instead — 26% at the leads rung, above 90% at collaborates or higher. Both are self-measured. The same setup is reproducible in your coding agent today.
Sources and check date
Anthropic's site: the measurement piece on its AI R&D (data through August 2026) and the companion piece on building itself on the same site (source of the 8x figure); OpenAI's two official blog articles of 2026-09-06 (Wayback capture); Sakana AI's official RSI Lab page and its 2026-09-24 announcement; per CNBC on 2026-09-11 and the AP story carried by Fortune and ABC on 2026-09-19; prices and default models checked against Anthropic's price list, OpenAI's model docs and the Claude Code changelog, all on 2026-09-29.
Wire the models the labs use into your coding agent
As of 2026-09-29: claude-opus-5-5 ($4 / $20) and gpt-6-sol ($2 / $10) are both callable on QCode with one key, and drop straight into Claude Code and Codex.
Related reading
Claude Opus 5.5 guide
Released 2026-09-22 at $4 / $20, the default Opus in Claude Code since 2.1.280.
GPT-6 Sol status tracker
Released 2026-09-22, the same day as Luna, at an official $2 / $10 and callable on QCode.
AI model radar 2026
Dated cards: released, cancelled, still unpublished.
This page is written in three tiers: official wording (dated), press reporting (attributed) and unverified material (listed, never adopted). 26%, 90%, 8x and 3.1 are the vendors' own tallies, unaudited by any third party; vendor-claimed speed and cost do not describe your workload. Callable on QCode covers only current model IDs such as claude-opus-5-5 and gpt-6-sol, and promises nothing about self-improvement.