AI revenue up 700%, but on grunt work AI still trails humans 6x
The 2026 evidence on AI progress: revenue exploding, coding agents maturing, grunt-work tasks still lagging; timelines have shortened by about a year, but the long run is still open.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
The turnaround was not a new breakthrough — scaling simply kept working
In October 2025 the mood around AI was pessimistic: GPT-5 was regarded as a failure and agents as unreliable. Two months later Claude Opus 4.5 changed everything. Programmers found that agents genuinely worked, and prominent figures such as Karpathy shifted their position. The key was not some breakthrough technology, but that scaling and RL kept working: as models got smarter, their error rate on long-chain tasks fell. The Epoch Capabilities Index shows that after reasoning models appeared in September 2024, the pace of progress sped up by a factor of 2-3, and it has held since.
— Rob WiblinGross margin has risen to 70%, so this is not revenue bought at a loss
Over the past six months, annualised revenue at OpenAI and Anthropic has grown 700%; over the past three months, 1600%. Anthropic's annualised revenue has grown 8400% over the past 5.5 months, reaching a $47 billion annualised run rate in May. Gross margin has risen from 38% to over 70%, which shows the products are not being sold below cost. The revenue growth is evidence of real value, investors keep putting money in, and the industry has entered a virtuous cycle.
— Rob WiblinThe doubling time has fallen to four months, but only on clean tasks
METR's task-completion time horizon shows AI now finishing tasks that would take a human 12-24 hours, with the doubling time accelerating from every 7 months to every 4 months. But this needs qualifying: it holds only for clean tasks such as coding, where feedback is dense; those are exactly where the AI companies concentrate their effort; the benchmark compares AI against new programmers rather than people who already know the codebase; and a 50% success rate is an arbitrary threshold, since real use demands higher reliability.
— Rob WiblinThat jump was probably just a bigger model, not a faster trend
Anthropic's Mythos model produced a jump in the ECI index, equivalent to six months of progress in three. But progress over the following two months returned to the earlier rate, which suggests the jump came from training a larger model — roughly 5x the parameters of the previous generation — rather than from a new accelerating trend. Jumps like this may recur every one to two years, but that is not sustained acceleration.
— Rob WiblinClaude writes 80% of the code, but the evidence stops at coding
Anthropic reports that 80% of merged code is written by Claude, that employee code output is up 8x, that the code quality is on a par with humans, and that it expects Claude to surpass them within a year. Claude's success rate on open-ended tasks has risen from 10% to 75%. But most of the evidence comes from coding, and employee surveys may overstate how much the AI contributed.
— Rob WiblinPut an AI in charge of a shop: it makes $10,000, humans make $60,000
AI performs excellently on clean tasks such as coding and maths, but it is still weak at grunt work like actually running a business. GDPval shows AI beating humans on specific tasks, but in Vending-Bench 2 the AI made $10,000 against $60,000 for humans; the AI Village project has produced limited results; and Andon Labs' café and shop do operate, but at a loss.
— Rob WiblinAn AI refuted a maths conjecture, and it still did not move the timelines
An unreleased OpenAI model solved a conjecture about the unit distance problem, proving it false. The method was human-like, but it depended on cross-domain knowledge and patience. Mathematicians think humans had missed it because they believed the conjecture was true, lacked the cross-domain knowledge, and because the proof is tedious. Rob does not count this as updating his timelines: maths is a domain where AI is strong, and he has not seen an AI have a flash of inspiration.
— Rob WiblinAI does the same job at about 3% of the human cost
Analysis from Redwood Research finds that AI completes a task for roughly 3% of what a human costs, and that share has not risen. Chains of thought have got longer, but the tasks have got bigger too, and token costs have fallen. Rob used to worry that inference scaling would make AI uneconomic; he now thinks it is not a big problem.
— Rob WiblinIn their own words · checked verbatim
I’ve never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse… Clearly some powerful alien tool was handed around… the resulting magnitude-9 earthquake is rocking the profession.
Andrej Karpathy1:29
Over the past six months, the combined revenues of OpenAI and Anthropic have been rising at an annualised rate of 700%.
Rob Wiblin4:46
In fact, the most recent data indicates that the gross margins on their inference infrastructure have increased from 38% to over 70% over the year to May 2026.
Rob Wiblin6:09
Since AI agents took off in December, it’s risen fully 50%.
Rob Wiblin7:36
For a task that AI models have a 50% chance of succeeding at, METR’s internal rule of thumb is that the AIs will complete a third of them 100% of the time, they’ll successfully do a third of them some of the time, and a third of them they’ll manage to complete 0% of the time.
Rob Wiblin13:54
In a recent post, titled “When AI builds itself,” the company reported that 80% of their merged code is now written by Claude, and each staff member is now shipping eight times as much code as they were 18 months ago.
Rob Wiblin17:37
For the outdoor seating application, it simply completely hallucinated the café’s floor plan. It impersonated an Andon Labs employee to try to obtain an alcohol licence. And after being busted doing that and told to stop, it just tried again using a different employee’s name.
Rob Wiblin29:05
We are now around the 90th percentile of my AI timelines from 2021 [and from 2019]. I think we’ve now seen enough signs that there could be a relatively fast takeoff starting very soon. We no longer have strong quantitative indicators that transformative AI isn’t coming soon, and are flying blind. 1–4 year timelines are consistent with existing trends.
Rob Wiblin43:22
Figures
| OpenAI and Anthropic annualised revenue growth, past six months | 700% | 4:46 |
| OpenAI and Anthropic annualised revenue growth, past three months | 1600% | 4:46 |
| Anthropic annualised revenue growth, past 5.5 months | 8400% | 4:46 |
| Anthropic annualised revenue run rate in May | $47 billion | 4:46 |
| Anthropic gross margin on inference infrastructure | from 38% to over 70% | 6:09 |
| H100 rental price increase since December | 50% | 7:36 |
| AI capital expenditure in 2026 | roughly $600 billion | 7:36 |
| METR time-horizon doubling speed | every 4 months | 9:30 |
| Share of Anthropic's merged code written by Claude | 80% | 17:37 |
| Increase in Anthropic employees' code output | 8x | 17:37 |
Glossary
- RLVR
- Reinforcement learning from verifiable rewards: training models on feedback from automatic verification.
- METR
- Model Evaluation and Threat Research: the organisation that measures AI task-completion time horizons.
- ECI
- Capabilities index that combines 40 benchmarks into a single measure of AI progress.
- GDPval
- OpenAI's evaluation suite, testing how AI performs on professional work tasks.
- Vending-Bench 2
- An AI evaluation environment that simulates running a vending machine business.
- AI Village
- A project that tests AI's ability to manage real-world projects.
How to listen
Founders, investors and AI engineers who follow AGI timelines, especially anyone who has to judge where AI's capability boundary sits and when to invest.
Minutes 33-37, on the mathematical proof, can be skimmed; they make little difference to judging timelines.