Will OpenAI or Anthropic announce recursive self-improvement?

Prediction markets and leading AI labs define recursive self-improvement similarly. Recursive self-improvement (RSI) is defined by prediction markets as an AI autonomously conducting research and development to materially improve its most capable systems, which then drive further enhancements ^[1][2][3]. This definition aligns with the understanding attributed to both OpenAI and Anthropic, describing RSI as an AI autonomously designing and training its successor ^[1][4][7]. The market's resolution criteria further specify that an announced RSI must represent an achieved capability where the company's AI autonomously performs R&D to substantially enhance its most capable or near-frontier systems, or creates a more capable successor, with the improved system subsequently driving further progress ^[1][2][3]. While humans may offer high-level goals and oversight, they must not provide substantive research choices or technical solutions, and trivial, narrow, or benchmark-only improvements are explicitly excluded from these criteria ^[1][8][7][9].

Official records do not list an RSI announcement from either company. As of November 14, 2026, official market records do not indicate an announcement by either OpenAI or Anthropic of an achieved RSI that meets the specified market criteria, which includes a strict requirement for an affirmative official announcement ^[1][4][7]. Anthropic's September 2026 update reported that Claude authored over 80% of merged code and managed 26% of core research and development tasks autonomously, yet explicitly stated that no measured workflow runs were fully closed-loop, thereby falling short of the market's RSI benchmark ^[4][5][6]. Similarly, OpenAI reached its 'automated research intern' milestone in September 2026, but maintains a March 2028 target for a fully automated AI researcher, which has not yet been announced as an RSI achievement ^[4][5][6].

For an official Recursive Self-Improvement (RSI) announcement from OpenAI or Anthropic before 2029, a specific sequence of research breakthroughs is anticipated, coupled with an executive statement confirming this capability ^[4][2][1]. The initial breakthrough involves sustained R&D acceleration of approximately 2x, a phenomenon observed in Anthropic's AECI benchmark scores as of November 5, 2026, primarily attributed to AI-assisted R&D ^[1].

Research agents achieving full autonomy and closed-loop improvements are crucial next steps. Following R&D acceleration, the next critical step is for research agents to attain full-researcher autonomy, with OpenAI setting a March 2028 target for a fully automated AI researcher ^[2]. Anthropic reported that Claude autonomously managed 26% of core R&D tasks as of September 2026 ^[2][3]. The most crucial breakthrough required is a closed-loop demonstration where the improved model itself generates further improvements ^[4][2][1]. Anthropic's update from September 2026 indicated that no measured workflow runs had achieved a fully closed-loop state, identifying this as the primary obstacle to an RSI claim ^[2][3]. This capability implies that the AI would autonomously conduct R&D to materially enhance its frontier system or create a more capable successor, with human involvement limited to setting goals and approving, rather than supplying substantive research choices ^[4][5].

Lab statements and market rules suggest potential recursive self-improvement (RSI) announcement delays. This challenges the high market odds for an announcement before January 1, 2029. While the market prices the probability for an RSI announcement before January 1, 2029, at 84%, the probability for such an announcement before January 1, 2027, is only 15.5% ^{[1][2][3][4][5][6]}. This indicates a meaningful probability that no qualifying announcement will occur by 2029.

Anthropic and OpenAI statements indicate RSI has not yet been achieved. Anthropic's September 2026 update explicitly stated no measured workflow runs were fully closed-loop, indicating that RSI, defined as autonomous design and training of a successor, has not been achieved ^[7][8][9]. OpenAI, having reached an 'automated research intern' milestone in September 2026, maintains a March 2028 target for a fully automated AI researcher, implying its internal timeline places full automation close to the 2029 deadline ^[7][8][9]. Even an OpenAI co-founder's personal forecast, which shifted his RSI timeline forward to 2027, remains a personal view and not a company claim, while Polymarket's December 31, 2026 RSI market traded near 8% 'Yes' ^[10][7][8].

Strict market rules and safety concerns incentivize labs to delay announcements. Market resolution rules mandate an official company communication or executive statement describing RSI as currently achieved, defined as an AI autonomously conducting R&D to materially improve its frontier system, with the improved system driving further improvements; AI-assistance statements, future plans, speculation, leaks, and third-party claims do not qualify ^[2][3][4]. Both labs emphasize safety limits and human oversight amidst calls for regulation, creating institutional and political incentives for labs to avoid a premature or loosely defined RSI announcement ^[10][7][8]. Additionally, Anthropic redacted 'a majority of the evidence we found most informative' in a METR report, with a cofounder citing risks of 'handing out a roadmap for others to go faster,' which complicates verification of any near-term RSI claim ^[10].

OpenAI projects a fully automated AI researcher by March 2028. This ambitious goal is supported by a September 2026 milestone where an 'automated research intern' achieved a productivity equivalent to 3.1 agent-workdays per human workday ^[1][2]. The rapid progress has led OpenAI co-founder Jack Clark to advance his personal timeline for recursive self-improvement to 2027, influenced by considerable improvements in research-taste benchmarks ^[1][2].

Anthropic demonstrates significant autonomous capabilities but no fully closed-loop systems. In a September 2026 update, Anthropic's models were responsible for authoring over 80% of merged code and independently managed 26% of core R&D tasks ^[1][2]. Despite these advancements, the company explicitly noted that no measured workflow runs have yet achieved a fully closed-loop state ^[1][2]. The available information does not mention other specific public timelines or roadmaps from Anthropic through 2028 regarding autonomous AI research.

Expert forecasts for recursive self-improvement diverge from market expectations. A Kalshi market tracking an announcement of recursive self-improvement (RSI) before January 1, 2029, was priced at 96% as of November 12, 2026, indicating high market confidence in a long-term announcement ^[1]. This contrasts with much lower probabilities in shorter-dated Polymarket RSI markets, which showed approximately 8-10% for an announcement by December 31, 2026, and less than 1% by October 31, 2026 ^[2][4][3]. Amidst this, Anthropic co-founder Jack Clark has personally advanced his RSI timeline to 2027, driven by rapid gains in research benchmarks, suggesting an earlier occurrence than the market's 2029 resolution ^[2][3].

Traders anticipate significant AI capability jumps by 2029, which likely contributes to the high probabilities observed in the long-term market. These expectations are influenced by developments such as Anthropic's September 2026 update, where its Claude model reportedly authored over 80% of merged code, and OpenAI's stated March 2028 target for a fully automated AI researcher as an 'automated research intern' milestone ^[2][3]. Separately, the AI Futures Project, led by Daniel Kokotajlo, has expressed criticism regarding Anthropic's progress and is in the process of developing updated timeline forecasts ^[4][5]. However, the provided sources do not include an explicit numerical expert probability from this project for a 2029 RSI announcement, which prevents a direct numerical comparison with the market's implied probability based solely on these facts ^[4][5].

Market expectations for Recursive Self-Improvement (RSI) announcements from OpenAI or Anthropic have recently shown volatility. Kalshi's 'Before Jan 1, 2029' market, KXRSIANNOUNCE-29, reached 96% on Nov 12, 2026, but has since pulled back to 84% (No 20¢) ^[1][2]. The 'Before Jan 1, 2028' contract currently stands at 71% (Yes 87¢), while 'Before Jan 1, 2027' is at 5% (Yes 53¢/No 95¢) ^[2]. Polymarket's Dec 31, 2026 contract for an announcement shows a 9% Yes probability ^[3]. Bullish catalysts include OpenAI's announced Sep 6, 2026, 'automated research intern' goal, targeting a fully automated AI researcher by March 2028 ^[4]. Anthropic's Frontier Safety Roadmap identifies automated AI R&D as plausible as soon as early 2027 ^[5]. An amended S-1/A filed Nov 2, 2026, states Anthropic 'may imminently meet our Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI' ^[4]. An internal Aug 2026 staff poll indicated median expectations for AI performing 80% of frontier-model-building work by end-2027 ^[4]. Anthropic co-founder Jack Clark moved his full-RSI forecast to 2027 on Oct 12, 2026, citing rapid benchmark gains ^[6], with Vals AI's RSI Index predicting full MLE-engineering automation at Anthropic around July–August 2027 ^[6]. Sam Altman stated in July 2026 that the next six months would see more AI capability progress than the prior two years ^[7], and Dario Amodei said in Sept 2026 that RSI 'is starting to happen across the industry' ^[4].

However, significant bearish risks and practical constraints persist. Anthropic's own institute analysis, 'When AI builds itself,' posits that RSI is not yet achieved and not inevitable, suggesting a scenario where capability gains flatten due to compute, energy, or chip-fab supply constraints ^[8]. A Seeking Alpha piece from Sep 15, 2026, notes that current evidence shows agents optimizing human-defined problems rather than choosing them, with human discovery driving all capability steps to date ^[9]. Physical limitations, including HBM sold out through 2026 with relief expected late 2027, grid interconnection lead times of 18–36 months, and the need for hundreds of gigawatts of new power by 2030, could impede the compute buildout essential for RSI ^[10][11]. A LessWrong scenario from Aug 2026 places 'prosaic RSI' as late as 2028–2032 ^[12]. Furthermore, the Kalshi market's resolution rules require an official announcement of RSI by OpenAI or Anthropic to trigger a Yes outcome, explicitly excluding mere assistance, acceleration, or narrow/toy improvements ^[3][13]. Cross-market indicators show Kalshi's OpenAI AGI announcement markets before 2027/2028/2030 at 3.5%/45%/71% ^[4], while Metaculus medians for weakly-general AI announcement are Aug 2027 and general AI Sep 2030 ^[4]. An LLM-ensemble forecast suggests a ~40% chance of a 'fast year' by 2030, with ~70% attribution to AI-automated R&D if it occurs ^[4].