- Published on
A three-line prompt cost $15,000, and the prompt was the cheapest part
- Authors
- Name
- AbnAsia.org
- @steven_n_t
Listen to this article · A three-line prompt cost $15,000, and the prompt was the cheapest part
0:00 / 0:00
Watch this article · A three-line prompt cost $15,000, and the prompt was the cheapest part
On 9 November, AI researcher and YouTuber Matthew Berman demonstrated a third-person shooter built in Unreal Engine 5 by Claude Fable 5.2, Anthropic's model released the same day. The starting instruction was three lines long. He then switched on the IDE's autonomous mode and let it run.
The result is a Gears of War: E-Day pastiche: a soldier who sprints, takes cover, leans out to fire, and performs the series' active-reload timing check. Enemies shoot back with working combat behaviour, a downed player can be revived by a squadmate, and the textures and lighting are credited to the model.
The bill is the interesting part.
The prompt is not the work
Three lines sounds like the headline. It is the least interesting figure here, because a prompt is a seed and this was a search. Forty-four unattended hours of the model writing, compiling, failing, reading its own errors and trying again is the work.
A team that reads "three-line prompt" and budgets for three lines of instruction will be wrong by the size of the project. What was bought was 44 hours of a machine that never got bored.
Where the money actually went
Divide the reported cost by the reported tokens and a number falls out that nobody is quoting: about $1.70 per million tokens, blended across input and output.
The same 8.63 billion tokens at the input rate alone would be roughly $69,000. Coming in at a fifth of that is only possible if the overwhelming majority of tokens were cache reads rather than fresh input.
So the cost story of 2026 is not a model-price story. It is a context-caching story. An agent that re-reads the same repository thousands of times is affordable or unaffordable depending on whether that re-reading hits cache — and most of that difference has nothing to do with which model you chose.
A loop re-reads everything
Every iteration reloads project context: engine source, assets, earlier diffs. That is where billions of tokens come from.
Cache hits set the price
Most of that context is unchanged between iterations, so it can be billed far below list. The $1.70 blend against an $8 list rate implies exactly that.
The first level carries the seam
This artifact includes all the setup cost. Later work in the same project reuses that context — but no one has published that curve yet.
At $336 an hour, the agent bills like a contractor
The hourly figure is the one to sit with. 336 per hour** of autonomous work — inside the range a senior freelance developer charges in most Western markets.
What came back is one playable level with visible animation glitches (Berman flags the vault-over-cover movement) and no independent testing of performance, stability or maintainability.
The suggestion is not that AI replaces developers, nor that it is too expensive. It is that the agent now sits between the two: a compute-metered contractor with no hourly floor and no judgement, which is a new line item in a budget rather than a cheaper version of an old one.
One level is not a game
The extrapolation trap is to multiply $15,000 by the number of levels in a AAA release. That sum is wrong at both ends: the first artifact carries all the setup, and a shipped game carries far more than geometry.
What the demo establishes is narrower and still significant: one Unreal Engine level with cover combat, working enemy AI and detailed lighting can come out of a short brief from a model left alone. Two years ago that was not a question you could ask; now the answer arrives with an invoice and a glitch.
The rule it appears to sit against
An awkward footnote: in late October Claude Code received instructions discouraging the recreation of existing games, and Anthropic confirmed the change specifically involved Opus 5.5. This run is a close recreation of a Microsoft franchise's signature mechanics and setting.
Whether that policy covers Fable 5.2, and whether an early-access configuration sits outside it, is not established by anything published. A capability demo is not a licence; the gap between "impressive" and "shippable" is where legal review enters.
What we would watch
The artefacts that would make this more than an anecdote: a cost curve across a second and third level, a token-per-feature metric that survives an audit, and a clear statement of which policy applies to which model.
We build data and inference systems, so the practical lesson is about budgeting rather than spectacle: an agent's cost is a property of your context strategy, not of its price list. Teams that treat cached context, retrieval and task granularity as engineering get the 69,000 one and conclude the model was expensive.
Source: Wccftech, reporting Matthew Berman's 9 November 2026 demonstration and Decrypt's figures — 43 hours 48 minutes, 8.63 billion tokens, approximately 8 per million input tokens and $40 per million output tokens. The blended rate and hourly rate are our arithmetic on those reported numbers, not claims made by Berman, Decrypt or Anthropic. The cost is an estimate of API pricing, not a statement of money spent.
Contact: [email protected]