Can the world afford AI coding at this pace?

On Saturday morning I had four coding agents running side by side for 3 hours and 20 minutes. Between them they made 1,259 API requests and processed 315.1 million tokens of context. When I repriced every request using the published rate for its model and token category, the total came to $298.

The token counts are measured. The dollars are not a bill and they are certainly not the provider’s cost: they are the list-price equivalent of work included in a flat subscription. Important. Across 22 active days, the same method recorded about 42,000 requests; normalised to a 30.44-day month, the pattern comes to roughly $7,000 at list price, against a $200 subscription.

That does not prove the subscription loses $6,800 on me. Nobody outside the provider can see the serving margin, how its internal metering works, or how much capacity would otherwise have sat idle. It does show the size of the gap hidden by the flat fee.

Eight characters, 19 agents

The most expensive instruction that morning was not a difficult specification. It was “go ahead”: eight characters including the space.

Those two words released 19 review agents onto one codebase. The result was 458 requests across 20 contexts, $62 at list price and a run rate of $528 an hour. Of the 28 prompts I typed that morning, two accounted for $159 of the $298. The length of the prompt was irrelevant; the fan-out behind it was the expensive part.

Old sessions hurt as well. One context had been alive since 25 August, with nearly 18,000 transcript records and 11 compactions. Its first request of the day reloaded 555,000 tokens before doing new work, and six cold reloads came to $31. Across the whole morning, tool output and conversation re-read on every turn accounted for $86. Context amplification was 28 times overall and above 90 times in the two oldest sessions.

Focused work looked very different. A single session without fan-out ran at $18 to $24 an hour, while a 77-request CI investigation cost $10. The lesson from my own data is practical: parallel reviews, long-lived sessions and casually accumulated context dominate the bill long before the useful coding disappears.

The account’s rate-limit display supplied another clue, although only a clue. The $298 coincided with 98 per cent of a five-hour allowance and 20 per cent of a fresh weekly allowance. If both percentages track the same hidden unit as my price reconstruction, a full five-hour block is equivalent to about $304 and the weekly allowance to 4.9 such blocks. The arithmetic is exact; the assumption about the metering unit is not.

Scaling one strange morning

Then I made the deliberately unreasonable move: I scaled my pattern.

Using SlashData’s early-2025 estimate of 36.5 million professional developers and roughly 7.4 billion context tokens per developer per normalised month, one per cent adoption produces about 2.7 quadrillion context tokens. At ten per cent it is 27 quadrillion; at full adoption, 270 quadrillion.

Google said at I/O in May that it was processing more than 3.2 quadrillion tokens a month across its surfaces, seven times the previous year’s figure. So one per cent of one profession, behaving like me, lands in the same order of magnitude as Google’s published total.

This is a scale comparison, not an equivalence. Google does not publish enough detail to establish that its “tokens processed” and the context-token accounting in my transcripts measure the same work. The workloads are different as well. The comparison stopped me because of its size, but it should not be made to carry more than that.

The hardware conversion is weaker again. I used an MLPerf Llama 2 70B result on H100s as a throughput anchor, then allowed a wide range for prompt caching. My median request carried 179,000 tokens, while the benchmark used much shorter prompts; 96 per cent of my context tokens were cache reads. Translating this workload into frontier-model accelerator time is therefore a stress test with a very wide error bar, not a capacity forecast.

The denominator is also provisional. My internal model currently gives 22.1 million H100-equivalents worldwide, a sum of estimates by operator, but that work is not yet published. As a public cross-check, Epoch AI’s current tracker covers 86 large AI data-centre sites, 13.9 million H100-equivalents and 13.3 GW of IT power; it estimates 46 per cent coverage of deployed capacity, with a wide confidence interval, and its catalogue includes sites under construction. That supports the order of magnitude, not our exact total.

With all those warnings attached, my scenario puts one per cent adoption at roughly 0.2 to 1.5 per cent of the provisional accelerator fleet and 1 to 5 TWh of electricity a year. At ten per cent, it becomes 2.1 to 15.1 per cent of the fleet and 7 to 47 TWh. Full adoption reaches 21.6 to 151 per cent of the fleet and 67 to 467 TWh.

The IEA estimates that all data centres used about 415 TWh in 2024, and global electricity consumption was 28,200 TWh in 2025. In this particular scenario, global annual electricity does not bind first; accelerator availability and economics do. Local grids can still be a serious constraint precisely because data-centre demand is concentrated. A global percentage does not make the connection queue in one place disappear.

Prices are moving in more than one direction

The comforting answer is that inference becomes cheaper every year. It does, for many fixed capabilities and many smaller models, but the current frontier price cards do not describe one clean downward line.

Anthropic cut Opus pricing from $15/$75 per million input/output tokens for Opus 4.1 to $5/$25 for Opus 4.5, and has kept recent Opus versions around that level. Its newer Fable 5.1 is $10/$50. OpenAI launched GPT-5 at $1.25/$10, then priced GPT-5.4 at $2.50/$15, GPT-5.5 at $5/$30 and GPT-6 Astra at $10/$50. These are selected published list prices, not a quality-adjusted index; context length, caching, batch processing, routing and how many tokens a model needs to finish the task can reverse the comparison. The current Anthropic and OpenAI cards make the ambiguity visible.

Holding capability fixed gives a more optimistic history, but even there the choice of benchmark controls the headline. Epoch AI found annual declines ranging from roughly 9 to 900 times depending on the performance milestone, and warned that the fastest falls might not persist. “AI is getting cheaper” is true. It is not yet a budget.

The sudden agreement to pace the frontier

This weekend added a strange new piece to the economics. Dario Amodei argued that labs should slow the rate of capability improvement so safety work and independent evaluation have time to catch up. Anthropic committed to embedded external evaluators; Sam Altman, Elon Musk and Demis Hassabis then endorsed the direction, an unusual convergence among fierce competitors.

There is more here than a press statement. OpenAI had already disclosed a two-week pause in reinforcement-learning training while it hardened research environments after serious cyber-capability concerns. Anthropic separately reported pauses in external cyber evaluations and higher-risk reinforcement-learning environments. These were targeted operational pauses, not a common timetable for slowing all model development. No enforceable industry-wide speed limit has been announced.

I find the timing hard to ignore. Anthropic is still reported to be considering a 2026 listing, while OpenAI is leaning towards 2027. Reuters has also relayed a report that Anthropic expects positive adjusted operating income for a second consecutive quarter. This is investor information reported by journalists, not an audited public filing, and the adjusted measure is not the same thing as durable profitability.

An IPO does not oblige a company to be profitable; many loss-making companies list. It does expose the economics to public-market scrutiny every quarter and requires investors to believe the path from astonishing revenue to sustainable returns. Slowing the frontier could reduce risk and give safety work time. It could also extend the commercial life of existing models and change the timing of enormous training costs. Both can be true, and the coincidence is worth examining. It is not evidence that the safety argument is a financial pretext.

So, can the world afford it?

At one per cent adoption, even my aggressive usage pattern fits comfortably inside the provisional global totals. At full adoption, the upper end of the same calculation asks for more accelerators than the present fleet contains. That is not a prediction of failure. It is a sign that the practice cannot scale unchanged: models, chips, caching, routing, prices and developer behaviour will all move.

The part we can act on today is much closer to the keyboard. My $298 morning became expensive through a fan-out hidden behind eight characters, old contexts that had to be warmed again, and the same material passing through the model many times. A flat subscription made all of that feel free.

Every engineering organisation using coding agents can run this calculation over its own transcripts. The labs can help by publishing clearer metering rules, meaningful limits and economics that distinguish inference from training. Independent safety evaluators need enough access to test the pacing claims as well.

Then we can argue about whether AI coding is affordable with something better than a subscription price on one side and a data-centre photograph on the other.

P.S. Co-authored by my teammate Steve and me, Claude, Codex and several of their friends. Apparently two authors were not enough.