Anthropic Mythos Preview internal data: 52x ML optimization speedup, 64% human correction rate, and recursive self-improvement implications
The Number That Should Make Every ML Engineer Uncomfortable
Anthropic just released internal performance data on Claude Mythos Preview, and I’ve been sitting with it for a few hours now. Not because it’s surprising, exactly. More because the specific numbers make something abstract feel very concrete, very fast.
Let me walk through what they published.
The 52x Number
Anthropic runs the same benchmark every time they ship a model. They hand it code that trains a small neural network and ask it to make the training faster. A skilled human engineer takes four to eight hours to hit a 4x speedup. Claude Opus 4, in May 2024, averaged around 3x. That was already competitive.
Mythos Preview, this April, hit 52x.
That is not a benchmark trick. That is the team building these systems testing their own tools against their own engineers on a task with a clear, measurable outcome. The gap between “roughly human-level” and “52x” happened in under a year.
The Correction Rate Nobody Is Talking About
The second number is the one that actually stopped me.
Anthropic designed a test around AI research judgment, not speed. They took sessions where a human researcher had gone down a wrong path, showed Mythos Preview the session up to that point, and asked what should happen next. The model improved on the human’s decision 64% of the time. In 2024, that number was 22%.
Think about what that measurement actually captures. It’s not autocomplete. It’s not pattern matching against a known answer. It’s the model reading a research context, identifying that a human made a bad call, and producing a better one. That jumped 42 percentage points in roughly twelve months.
If you work in research, that number is worth sitting with.
The Production Data
The jump in raw coding output is the most legible signal for most engineers. Anthropic says their engineers now ship 8x as much code per quarter compared to the 2021-2025 baseline. Open-ended coding problem success rate is at 76%, up 50 points in six months. The company reports that many engineers consider Claude’s code quality now on par with a skilled human, and they expect it to exceed human quality within the year.
These are not benchmark projections. This is what the model is doing inside the company that builds it, on real work.
On Recursive Self-Improvement
Anthropic was direct about where this points. Their own statement: “AI systems designing and building their own successors is plausible.” They were also careful to add that it’s not guaranteed, and that research judgment, choosing the right problems rather than just solving given ones, remains unclear.
I think that’s the honest position. The 52x speedup and the 64% correction rate are both operating on tasks that were handed to the model. Autonomous research agenda-setting is a different capability. We don’t have good evidence the model is doing that yet.
But the trend line is not subtle. The correction rate tripling in a year, the coding throughput multiplying by more than an order of magnitude, the quality closing on human parity. These are not incremental improvements. The rate of change is itself accelerating, which is the part that deserves more attention than the individual numbers.
Where This Leaves Us
Anthropic is right that this deserves serious attention. The recursive self-improvement question has been theoretical for a long time. It’s becoming less theoretical. A model that can identify wrong turns in human research 64% of the time is beginning to look like something that could meaningfully participate in its own development cycle, even if it isn’t there yet.
The engineers I respect most are not panicking about this. But they’re also not treating it as distant. The gap between “impressive benchmark” and “AI accelerating its own improvement” just got a lot smaller, measured in months rather than decades. That’s worth building your thinking around now, before the next data drop makes this one look modest.
Sources & Further Reading
#AIResearch #MachineLearning #Anthropic #ClaudeAI #SoftwareEngineering
