When Will AI Self-Improve? Anthropic, OpenAI and Meta Disagree
Anthropic says 2x, OpenAI says 3x, Meta uses benchmarks. Here's what the competing claims about AI-accelerated research actually mean.

What does “recursive self-improvement” actually mean?
Recursive self-improvement describes a point where an AI model contributes meaningfully to designing, training, or optimizing its successor, creating a feedback loop that speeds up each new generation. Right now, no major lab claims to have reached that point. What they do claim is something more modest but still significant: that AI tools are already accelerating the humans who build AI. Anthropic, OpenAI, and Meta have each published different numbers for how much faster their research moves with AI assistance, and the gap between those numbers says a lot about how hard this is to measure.
TL;DR
- Anthropic’s IPO filings reportedly cite AI assistance as helping the company achieve roughly two times its historical rate of research progress, described internally as getting two years of work done in one year.
- OpenAI’s internal report, released within the last day of the source video, has researchers reporting they can complete three months of coding work in one month using coding agents, a higher multiplier than Anthropic’s.
- Meta measured actual benchmark trend lines rather than researcher self-reports, tracking progress across roughly 15 benchmarks and subjectively attributing some of the increase to AI, after finding flaws in three of them.
- Self-reported researcher speedup varies wildly by question asked: Anthropic researchers reportedly said AI sped up their personal output by as much as 12x, while OpenAI researchers reported 3x when asked about equivalent work completed, showing how sensitive these numbers are to framing.
- External benchmarks like the Frontier RSI leaderboard (from Abundant Labs) and Valve’s AI comparisons rank Anthropic’s Claude Opus and Sonnet-class models ahead of competitors on AI-research-specific tasks, suggesting Anthropic may be training more heavily on self-improvement-relevant data.
- Anthropic co-founder Tom Brown has publicly described “compute multipliers” as the mechanism behind recursive self-improvement: efficiency discoveries that let a fixed amount of hardware train a smarter model, with the explicit goal of having Claude eventually discover these multipliers itself.
- No one can point to the exact moment on a capability graph when models crossed from “helpful tool” to “research accelerant.” Labs only noticed the slope changing in hindsight, sometime between August and October 2025.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Why do Anthropic, OpenAI, and Meta report different numbers?
The short answer is that they’re measuring different things with different methods. Anthropic’s two-times figure, which had to be disclosed as part of its IPO documentation, is framed around historical rate of progress attributable to AI assistance. OpenAI’s three-times figure comes from asking researchers directly how much coding-agent-assisted work compares to a ChatGPT-only baseline, measured in “work of equal value” completed per unit time. Meta took yet another approach, tracking around 15 benchmarks and trying to separate the normal pace of improvement from the portion caused specifically by AI tools, a judgment call that required discarding three benchmarks they found unreliable.
None of these methodologies produces a clean, comparable output. There’s no single neutral number that tells a researcher “79% of today’s progress came from AI,” as the video’s commentary puts it. Every figure depends on the question asked, the baseline chosen, and which benchmarks or self-reports get included. Even within one lab, numbers diverge depending on whether you ask about personal speedup (how much faster does this feel) versus equivalent output (how much actual work got done). That distinction explains why Anthropic researchers could report being sped up 12x personally while the company’s own IPO documents cite a more conservative 2x figure for net progress.
How much is AI actually speeding up AI research right now?
Based on current disclosures, the honest answer is “meaningfully, but not yet in a way anyone can precisely quantify.” The three data points on the table, roughly 2x from Anthropic, roughly 3x from OpenAI, and a benchmark-driven estimate from Meta, all point in the same direction: AI coding agents are freeing up researchers to test more ideas in parallel, and that is changing the slope of progress. Both Anthropic and OpenAI’s own capability charts show an inflection point sometime between August and October 2025, which both labs attribute to models becoming good enough to automate meaningful chunks of research coding.
What nobody can do yet is predict exactly where that inflection leads or when it accelerates further. The charts that show steady, predictable aggregate progress can’t show you the moment a model crosses a threshold that opens up something qualitatively different, like finding a pattern in enough cybersecurity data to locate zero-day vulnerabilities at scale, or finding enough mathematical structure to crack a previously unsolved problem. Progress on the aggregate chart looks smooth. The downstream effects are not smooth at all.
Is there evidence any single company is ahead on self-improvement specifically?
There are external signals pointing that way, though they remain proxies rather than direct proof. The Frontier RSI leaderboard from Abundant Labs, built specifically to test models on AI-research-relevant tasks, currently ranks Anthropic’s top model ahead of competitors, with OpenAI’s flagship and Google’s Gemini models trailing by varying margins. Valve has also run direct comparisons pitting OpenAI and Anthropic models against each other on related tasks. Two independent measurements pointing the same direction is suggestive, but it’s still indirect evidence.
The more telling signal might be qualitative. Anthropic co-founder Tom Brown has talked publicly about the company’s strategy around “compute multipliers,” efficiency gains that let the same hardware train a more capable model. He’s described these multipliers as compounding (each new one boosts the value of previous ones) and said Anthropic’s explicit goal is for Claude to eventually discover these multipliers autonomously, which is about as direct a description of recursive self-improvement as any lab has given publicly. Whether that capability exists today, or is still aspirational, is not something outside observers can verify independently. It depends entirely on what each lab chooses to disclose.
What’s the risk in racing toward a self-improving AI researcher?
The core tension is that acceleration in one domain tends to accelerate everywhere the model has relevant training data, including domains with safety implications. The same dynamic that lets a coding agent speed up research also appears to be what’s let newer models get dramatically better at finding security exploits. One frontier lab has reportedly built an internal cyber-focused model scoring highly on exploit-finding benchmarks, restricted to trusted testers rather than public release. Recent disruptions affecting banks, airlines, and other infrastructure have been linked by some observers to defensive patching done in response to vulnerabilities surfaced by AI systems.
That raises an incentive question worth sitting with: if a lab can sell cyber-defense services built on the same models capable of finding those vulnerabilities offensively, does that create pressure to push capability forward faster than a purely safety-focused calculus would recommend? Anthropic’s public positioning, sometimes summarized internally as “you can’t do safety from second place,” argues the opposite: that staying at the frontier is necessary to set safety norms other labs will follow, citing choices like declining to build advertising into their consumer products as evidence of responsible practice. Whether leading the race and practicing safety are actually compatible goals, or just a convenient narrative, remains an open and contested question.
Frequently Asked Questions
What is the difference between AI acceleration and recursive self-improvement?
AI acceleration means AI tools make human researchers faster or more productive, which labs are already measuring (with inconsistent methods). Recursive self-improvement means AI systems contribute directly to designing or optimizing their own successors with reduced human involvement. Current evidence supports the former happening now; the latter remains a stated goal rather than a confirmed capability.
Why do Anthropic and OpenAI report different acceleration multipliers?
They use different baselines and different survey questions. Anthropic’s figure comes from IPO disclosures comparing current progress to historical trend rates. OpenAI’s figure comes from researchers directly comparing coding-agent-assisted output to a chatbot-only baseline. Different methodologies produce different, non-comparable numbers.
Is there a specific date predicted for a superhuman AI researcher?
No credible, verifiable date exists in current disclosures. Labs have pointed to an inflection in their own progress charts around August to October 2025, but that marks a change in slope, not an endpoint. Public speculation about 2027 or 2028 timelines exists but is not something any lab has confirmed with hard evidence.
Can benchmarks reliably tell us which company is closest to self-improving AI?
Only approximately. Benchmarks like the Frontier RSI leaderboard are designed to proxy AI-research capability, and they currently favor some models over others, but they remain indirect measures. Actual self-improvement claims depend heavily on what each lab discloses about its internal processes, which isn’t independently verifiable.
Does faster AI research also mean faster emergence of AI risks?
The available evidence suggests capability gains aren’t isolated to one domain. Models trained with more research-relevant data also show jumps in other areas, including cybersecurity exploit-finding, which has prompted real-world defensive patching. This suggests that acceleration in research productivity and acceleration in riskier capabilities may not be separable.



