How Much Is AI Speeding Up AI Research? 3x vs 12x Explained
OpenAI says coding agents give researchers 3x output, while Anthropic researchers say 12x. Here is why the numbers differ and how to read any AI speedup claim.
For Executives, Managers and Engineering

How much is AI speeding up AI research?
Nobody agrees, and the gap is large. An OpenAI report says its researchers get about three months of work done in one month with coding agents, a 3x gain in output of equal value. Anthropic researchers, asked how much models speed them up, said 12x. Anthropic's own threshold for a major milestone is progress at twice the historical rate because of AI. The numbers differ mostly because each lab asks a different question and measures it differently.
TL;DR
- OpenAI's figure is 3x: its researchers said coding agents let them do three months of work in one month, compared with using ChatGPT alone.
- Anthropic researchers reported 12x when asked how much models speed up their personal work, which is a different question from how much faster the lab's overall progress is.
- Anthropic's milestone is 2x: the historical rate of progress, attributable to AI, doubled, meaning two years of work done in one year. Meta used a similar benchmark-trend approach.
- None of these numbers has a true baseline. No one can run the same lab twice, once with AI and once without.
- Outside tests are proxies. The Frontier RSI leaderboard ranks models on AI research tasks, but it cannot replace what labs see inside their own work.
- Anthropic's pitch is "compute multipliers": efficiency gains that stack, which a model might one day find on its own.
- The same lesson applies to you. Speedup claims depend on the task, the question, and who is answering.
What does "AI automating AI research" actually mean?
It means AI systems doing part of the work that AI researchers do: writing code, running experiments, reviewing results, and trying ideas in parallel. The stronger version is recursive self-improvement, where a model gets good enough at AI research that it helps build its own successor, which then helps build the next one, and so on.
We are not at the strong version. We are at the stage where researchers use coding agents heavily. One analysis of the public record points to a shift between August and October 2025. Around then, models became good enough to automate much of researchers' coding, which freed people to test many ideas at once. Both OpenAI and Anthropic have charts showing model progress hitting a steeper slope at about that time. OpenAI's chart starts in August 2025 for that reason.
Why do OpenAI and Anthropic report such different numbers?
Because the questions are different. Here is what each measure actually asked.
| Measure | What it asked | Result |
|---|---|---|
| OpenAI report | How much work of equal value can you finish with coding agents versus just ChatGPT? | 3x (three months of work in one month) |
| Anthropic researcher survey | How much are you personally sped up by models? | 12x |
| Anthropic IPO threshold | Is the rate of progress, attributable to AI, double the historical trend? | 2x (two years of work in one year) |
| Meta and Anthropic benchmark method | Is progress on about 15 benchmarks faster than the historical rate? | Compared with the historical trend |
Three things drive the gaps.
Output versus speed. Feeling faster at your own tasks is not the same as producing more of equal value. A researcher can write code 12x faster and still be limited by waiting for experiments, reviewing results, and deciding what to try next.
Personal speedup versus lab progress. A person's speedup is one input. Whether the lab's models improve twice as fast depends on compute, data, and many other bottlenecks.
Different baselines. OpenAI compared coding agents against a ChatGPT-only workflow, which is already a faster baseline than working with no AI at all. Meta and Anthropic compared against the historical trend line of progress.
On its face, OpenAI's 3x looks like a higher acceleration than the 2x threshold Anthropic and Meta used. But these are different measurements, so putting them on one scale misleads.
Why is it so hard to measure the speedup?
There is no counterfactual. No researcher gets a note at the end of the day saying 79.2% of today's work was thanks to AI. To estimate AI's effect on a benchmark trend, Meta had to make subjective judgments. They also found flaws in three of the benchmarks they used.
Then there is a second problem: separating a researcher's personal speedup from a real increase in what the AI can do. A person who learns to use an agent well gets faster even if the model has not changed.
Task type matters too. A time-horizon graph from the research group METR suggests models can do coding and machine-learning research tasks that take humans dozens of hours (extrapolated to around 40 hours) at 80% reliability. But OpenAI's own report, using data from as recently as October 18, shows 80% reliability only on tasks under 15 minutes in this area. The difference comes down to the task mix and the data models have seen. For reviewing code, a model like GPT 6.2 Astra succeeds on its own roughly 70% of the time. For writing code for inference and serving stacks, it is closer to 25%. Any single "speedup" number averages over wildly uneven ground.
What is the Frontier RSI leaderboard?
It is an outside attempt to measure how good models are at self-improvement. Abundant Labs built it. In the version discussed, Opus ranks ahead of Fable 5.2, which is narrowly ahead of Gemini 4 Carbon, which is well ahead of Astra.
The value is that it does not come from a lab marketing its own model. The limit is that any benchmark is a proxy for real AI research, which is messier. For the real thing, we depend on what labs choose to disclose. It is still a useful signal: this ranking and other results both suggest Anthropic's models may have been trained on more data related to AI research itself.
What are compute multipliers, and why is Anthropic pitching them?
Anthropic co-founder Tom Brown described "total effective compute" as physical compute capacity multiplied by the compute multipliers a lab discovers. A multiplier is an efficiency gain. His example: if you need 100 units of compute to train a model, a 4x multiplier means you can train the same model with 25 units. Or you can use all 100 units to train a model that would previously have needed 400.
His claim is that these multipliers stack, and each new one benefits the ones before it. The pitch is that this becomes a hard-to-copy advantage, and that it could accelerate further near recursive self-improvement, when Claude could discover multipliers by itself. This is a company's argument made around its IPO, so treat it as a claim about strategy, not a measured result.
Why does the data behind a model matter?
Progress is uneven because models learn patterns from the data they are trained on. If raw intelligence were all that mattered, every model would score about the same everywhere. Instead, one model leads on a coding benchmark, another on a science exam, another on a protein task. Different aggregate indexes even crown different leaders.
It also means breakthroughs are hard to predict. Smooth benchmark charts do not show when a model will suddenly become useful at, say, finding software vulnerabilities. The same may hold for AI research: strength there depends on how much relevant data a lab trained on.
What does this mean for you?
You will probably never see a lab's internal numbers, but the same traps show up in your own workplace when someone claims "AI makes us 5x faster."
- Ask what was measured. Time to first draft? Finished work that passed review? Those give different answers.
- Ask for the baseline. Faster than what? No AI, or a basic chatbot?
- Test on your own task. Models differ by domain, so try the main model families on the work you do. A leaderboard lead elsewhere may not carry over.
- Track your own numbers. For a few weeks, log tasks, rough time spent, and whether the output needed rework. Your own data beats any lab's headline.
- Expect uneven gains. Reviewing and drafting may speed up a lot while some steps barely change.
If you manage a team, run a small pilot with a clear question and the same task done both ways before you commit budget based on a vendor's multiplier.
Can you build a tool to track your own AI speedup?
If you want those logs in something better than a shared spreadsheet, a small internal app does the job. With Remy, you can describe it in plain English: a form where people log a task, the tool used, estimated hours with and without AI, and whether the result needed rework, plus a dashboard that averages these by task type. Remy builds and deploys the full-stack app, with sign-in and a database included, so your team can see where AI helps and where it does not. You own the code, and nothing here needs a new subscription for each person.
Frequently Asked Questions
Is OpenAI's 3x higher than Anthropic's 2x?
Not in a way you can compare directly. OpenAI's 3x came from researcher answers about work of equal value done with coding agents versus ChatGPT alone. Anthropic's 2x is a threshold for the AI-attributable rate of progress against a historical trend. The source analysis noted that, on its face, OpenAI's figure is higher, but they measure different things.
What is recursive self-improvement?
It is when an AI system becomes good enough at AI research to help improve itself or build better successors, creating a loop. Labs describe it as a future possibility. Today's evidence is researchers using coding agents to speed up their work, with humans still directing it.
Why did Anthropic researchers say 12x?
They were asked how much models sped them up personally. That is a question about individual speed, not overall lab progress or output of equal value, so it naturally produces a bigger number than a lab-wide rate.
Can I trust a single benchmark to tell me which model is best?
No. Different benchmarks and aggregate indexes crown different leaders, and performance depends on the domain and the data a model saw in training. Test the leading models on your actual tasks.
How we source, write, and correct our articles. Editorial standards
Related articles

How Much Should You Spend on AI? Hormozi's Budget Lesson

Which AI Model Is Best? Why Benchmarks Disagree by Domain

Do AI Models Get Nerfed After Launch? What Labs Admitted
