METR's assessment of Anthropic shows AI-assisted research roughly doubling progress speed, an early sign of recursive self-improvement at AI labs.
What did METR find about Anthropic’s research speed? #
METR, a research organization that evaluates AI capabilities, spent 11 weeks examining Anthropic’s internal data and interviewing staff before publishing an assessment on November 5. The team didn’t get direct access to Anthropic’s systems or the chance to test unreleased models on verifiable tasks (problems with checkable right answers). Instead, they relied on internal data and conversations with researchers who work with those models daily.
The conclusion: most of the team judged that Anthropic’s research is now progressing roughly twice as fast as it did between spring 2024 and fall 2025. That’s a rough estimate of two years of research progress happening in about one year. METR attaches real uncertainty to that number, putting the plausible range between 1.3x and 3x. Anthropic, for its part, redacted some of the evidence it gave METR because it wasn’t suitable for public release, which is its own small signal about how sensitive this topic has become.
Why is this happening now? #
METR attributes most of the current acceleration to Anthropic’s own use of AI inside the research process itself. Researchers are using models to write research code, set up experiments, and investigate results. That’s not the same as a model redesigning itself. It’s closer to giving every researcher a very capable junior collaborator that handles the parts of the job with clear, checkable outputs: coding, testing, debugging, running experiment pipelines.
Rapidlybuild enterprise-grade, production-ready software.
Already in:✓ SOC 2✓ HIPAA✓ Audit log
That distinction matters. The tasks where AI helps most are ones with verifiable answers. Code either compiles and passes tests or it doesn’t. An experiment either produces the expected output or it doesn’t. Those are exactly the kinds of tasks current models are strongest at. The harder, fuzzier calls, like which research direction is worth pursuing or how to interpret an ambiguous result, still sit with human researchers.
This pattern isn’t confined to Anthropic. OpenAI has reported that experiments per active experimenter hit a new high in August, in data going back over a year and a half, driven by more coding agent use and more available compute. At Google, CEO Sundar Pichai has said agents are already helping researchers build models, while adding that fully autonomous self-improvement, where models upgrade themselves without humans steering the process, is still a ways off.
TL;DR #
- METR’s assessment found Anthropic’s AI research moving roughly twice as fast as its 2024-2025 pace, with a plausible range of 1.3x to 3x speedup.
- AI-assisted research , not model self-redesign, is the current driver: researchers use models to write code, run experiments, and check results faster.
- Verifiable tasks like coding and testing are where models help most, while research direction and judgment calls still rest with humans.
- OpenAI and Google report similar patterns, with experiment volume rising and agents assisting researchers, though full autonomy remains distant.
- Compute access compounds this effect: labs with the most powerful chips can run more experiments and iterate faster, reinforcing any lead they already have.
- The uncertainty is real. METR couldn’t test unreleased models directly, so these figures are informed estimates from secondhand data, not hard benchmarks.
Is this actually recursive self-improvement? #
The term “recursive self-improvement” usually conjures an image of AI redesigning itself with no human involved, accelerating without limit. What METR describes is a milder version: assisted recursive self-improvement, where humans remain fully in the loop but get meaningfully faster at their jobs because AI handles more of the grunt work.
That’s still a meaningful shift. If researchers finish more useful work per unit of time because AI is doing the coding and experiment-running, the entire research cycle compresses. A lab that used to take a year to go from one model generation to the next might do it in significantly less time, not because the AI is smarter about strategy, but because the mechanical overhead of research shrinks.
Whether that counts as “recursive self-improvement” in the strict sense is partly a semantic argument. The practical reality is what matters: labs are reporting faster iteration, and that faster iteration compounds. A model that helps build a slightly better successor, which then helps build an even better one after that, creates a feedback loop even with humans steering every step.
Why does compute access matter so much here? #
Research acceleration doesn’t happen in a vacuum. It requires running experiments, which requires computing power. A lab with access to more advanced chips can run more experiments in parallel, test more ideas, and iterate through failures faster. That’s one reason the current scramble over advanced Nvidia chips, including disputes over export rules and chip allocation between the US and China, matters beyond immediate sales figures.
__>__add rate limiting to the API
● Reading src/server.ts
● Editing 3 files
+ import{ limit } from'rate'
- app.use(cors()) ✓ Tests passed · ready for PR
● Live
Approved
Approved
Approved
The chip generation in use changes how much research a lab can actually do. Older processors can still train and run large models, but newer chip generations offer substantially more capacity for the kind of demanding, iterative work that research acceleration depends on. A lab with a head start on both compute and AI-assisted research tooling doesn’t just move faster today. It compounds that advantage, because faster research cycles mean faster access to the next generation of models, which can then accelerate research further.
This is the mechanism behind why “getting the strongest model first” matters so much to AI labs right now. It’s not just about selling a better product. It’s about using that model internally to build the next one faster than competitors can.
What does “spiky intelligence” mean for how labs use these models? #
Current frontier models show what’s sometimes called spiky intelligence: extraordinary performance in narrow areas alongside much weaker or inconsistent performance elsewhere. A model can be exceptional at mathematics or code generation while still making poor judgment calls on open-ended business decisions.
Inside a research lab, that spikiness is actually useful. Researchers can hand off the specific tasks where models excel, writing code, running tests, checking verifiable results, while keeping the judgment calls, like which experiments are worth running at all, in human hands. This is a practical division of labor: let the model do what it’s good at, keep humans responsible for the parts that require broader context or strategic judgment.
That’s also why the acceleration shows up unevenly across companies and even within them. A lab that restructures its workflow to maximize the model’s strengths in coding and experimentation will see bigger gains than one that tries to use AI for tasks outside its current strengths.
Is the acceleration trend reliable or just anecdotal? #
It’s somewhere in between. METR is a credible, independent evaluator, and its methodology involved real internal data and staff interviews rather than guesswork. But the assessment comes with built-in limits: no direct system access, no ability to run unreleased models through verifiable benchmark tasks, and reliance on secondhand accounts from people close to the work. That’s why METR’s own range spans from a 1.3x to a 3x speedup rather than a single confident number.
The corroborating signals from OpenAI and Google add weight. OpenAI’s experiment volume data is a more concrete, if narrower, metric, tracking actual experiment counts over time rather than researcher impressions. Google’s public comments from Pichai align directionally without providing hard numbers. None of this proves a precise acceleration figure, but the consistent direction across three major labs makes the underlying trend harder to dismiss as a one-off.
Frequently Asked Questions #
What is METR and why does its assessment matter?
METR is a research organization focused on evaluating AI capabilities. Its assessment of Anthropic matters because it’s an independent, outside look at how AI tools are changing the pace of AI research itself, based on internal data and staff interviews rather than company self-reporting alone.
Does this mean AI is already improving itself without humans?
No. What’s happening now is assisted acceleration: humans still decide what to research and how to interpret results, but AI tools write code, run experiments, and handle other verifiable tasks faster than humans could alone. Full autonomous self-improvement, without humans in the loop, hasn’t arrived yet according to figures like Google’s Sundar Pichai.
Why is the speedup estimate such a wide range?
mission-criticalenterprise applications.
✓ HIPAA
✓ SOC 2
✓ Self-hosted
METR didn’t have direct access to Anthropic’s unreleased systems or the ability to test them on verifiable benchmark tasks. Its 1.3x to 3x range reflects honest uncertainty from working with secondhand data and interviews rather than hands-on testing.
How does chip access connect to research speed?
Faster research cycles depend on running many experiments, and running experiments requires computing power. Labs with access to more advanced chips can iterate through more experiments in less time, which compounds any existing lead in model capability.
Is this trend limited to Anthropic?
No. OpenAI has reported rising experiment volume tied to coding agent use and compute access, and Google’s leadership has described agents already assisting researchers in building models, suggesting this is a broader pattern across major AI labs rather than one company’s experience.