🔍 Anthropic: Claude is generating it's own successors

Anthropic recently announced in a paper that its AI is generating its own successors. It’s already 80% there.

Anthropic published a research paper on when AI builds itself and the title is not a metaphor. The first data point that the paper presents is this: over 80 percent of code merged into Anthropic's internal systems was written by Claude in May 2026.

That number was in the low single digits two years ago.

What the numbers actually show

Going from a few percent to 80% in the course of two years is by design. Anthropic began routing much of its own development work through AI when Claude Code launched in early 2025. Even those carrying out that handover have been surprised by the speed of it. Anthropic engineers now commit about eight times more code per day on average than in 2024. The humans are still in the loop. But they are typing a lot less.

Anthropic runs its own internal coding benchmark, and the results tell a similarly stark story. One test feeds Claude some code and asks it to re-write the code so that it runs more quickly. May 2025, Claude Opus 4 had an average speedup of around 3x. Mythos Preview clocked in at 52x by April, 2026, a qualified human engineer could generally get about 4x in four to eight hours, so that makes AI more than a factor of 13 ahead of the state-of-the-art human effort, and in less than a year.

The research tasks are going on a similar curve. Performance on CORE-Bench, which tests if an AI can replicate published scientific results, went from roughly 20% in 2024 to almost perfect 15 months later. Mythos Preview chose the more effective next step in a research session 64% of the time in one internal study, compared with 51% for Opus 4.5 six months prior.

What Anthropic is so terrified of

This paper is not a product announcement. It is a warning.

The theory is, at the point when an AI system could do to contribute in a way that provides utility in creating the next AI system, it can draw on itself. In the process, it can be maddening because each generation of AI is better at building the next one. Each iteration comes faster. At some point, the humans in the loop are no longer ‘steering’.

The paper says, "Recursive self-improvement is not inevitable.” However the company also says that the direction of travel is sufficiently clear that governments, regulators and international institutions need to start preparing rather than attempting to react at the moment itself.

Jack Clark, one of the papers author told Axios: "Anthropic hopes to engage lawmakers directly on this over the next few months. At least the paper itself makes a specific proposal though: A coordinated, transparent alternative for the top AI labs in the world (including Chinese ones) to all pause frontier development at once, as needed, if and when risk is particularly high. It is not just a pause from one company. An international regime, similar to arms control, for AI.

The timing is worth noting

This paper was published by Anthropic one week after the company confidentially filed for an IPO which could value it at over $1 trillion. The combination of touting 8x engineer productivity to investors while also arguing for a possible world slowdown in development has raised eyebrows.

The honest read is likely that both are true at the same time. Anthropic actually thinks it is doing something important enough to justify global governance regimes. It is also a soon-to-be publicly traded corporation that wants you to know its engineers are now eight times more productive than two years ago.

The paper is real. The concern is real. This is also not an accident in timing

Pioneer AI — Unlimited inference across every major AI model for $20 a month.

The pitch is simple. Pay $20, get unlimited access to GPT, Claude, Gemini, DeepSeek, and other frontier models through a single API endpoint. No per-token billing. No watching usage meters. Just use what you need.

The catch is small but worth knowing. It does not plug into desktop apps directly. You need to configure it through a CLI, Cursor, OpenCode, or any other coding interface that accepts a custom API endpoint. If you are not sure how to set it up, ask the model itself once you are in. It will walk you through it in about two minutes.

At current model pricing, a developer running Claude Opus 4.8 or GPT-5.5 heavily could easily rack up $200 to $500 in API costs in a month. Pioneer caps that at $20. The deal runs until August 1, so there are roughly two months left at this price before terms may change.

For builders, researchers, and anyone running agents or coding workflows at volume, this is worth trying before the window closes.

Strategy is all red.

Join the Conversation!

We'd love to hear your thoughts and comments. Join our community and stay updated with the latest trends and discussions in crypto.

Twitter | Instagram | Telegram Channel | Linkedin | Facebook