AI
“Felix Rieseberg used Claude Fable 5.2 to create a full playable FPS video game from a single prompt, with every model, sound, and texture made from scratch, taking 8 hours and about $550 in API costs.”
This post lists 5 claims. We checked item 1.
Plain restatementAn individual named Felix Rieseberg used the Claude Fable 5.2 model to produce a playable first-person shooter from one initial prompt, in which the 3D models, audio and textures were generated by the model rather than sourced, with the run taking approximately 8 hours and costing approximately $550 at API list pricing.
This one is real but oversold. Claude Fable 5.2 is a genuine Anthropic model released on 9 November 2026, and Felix Rieseberg really did post a browser-playable first-person shooter called Saltline made with it. The 8 hour and roughly $550 figures match his own post word for word, and he says the models, sounds and textures were all generated by the model as code rather than taken from asset files. Three things the post leaves out matter. Saltline is a compact single level with one objective, not a full game, and the prompt itself asked for a tiny level. Rieseberg works at Anthropic and published this on launch day, so it is the maker's own demonstration rather than an independent user's finding, and the time and cost figures are his unverified self-report that nobody has reproduced. The version most people click and play is also one he edited by hand afterwards, with the untouched single-prompt build on a separate link.
Felix Rieseberg used Claude Fable 5.2 to [drifted from the evidence:] create a [drifted from the evidence:] full playable [drifted from the evidence:] FPS video game from [drifted from the evidence:] a single prompt, [drifted from the evidence:] with every model, sound, and [drifted from the evidence:] texture made from scratch, taking 8 hours and [drifted from the evidence:] about $550 [drifted from the evidence:] in API [drifted from the evidence:] costs.
[added by the neutral restatement:] An individual named Felix Rieseberg used [added by the neutral restatement:] the Claude Fable 5.2 [added by the neutral restatement:] model to [added by the neutral restatement:] produce a playable [added by the neutral restatement:] first-person shooter from [added by the neutral restatement:] one initial prompt, [added by the neutral restatement:] in which the 3D models, audio and [added by the neutral restatement:] textures were generated by the model rather than sourced, with the run taking [added by the neutral restatement:] approximately 8 hours and [added by the neutral restatement:] costing approximately $550 [added by the neutral restatement:] at API [added by the neutral restatement:] list pricing.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Claude Fable 5.2 exists and was released by Anthropic on 9 November 2026. It is not a fabricated model name.
- Felix Rieseberg is real, works at Anthropic, and did post a browser-playable first-person shooter called Saltline built with Fable 5.2 on 9 November 2026.
- The figures in the claim match his own published figures exactly: approximately 8 hours and approximately $550 at API pricing, alongside about 2.7M output and 445M input tokens with heavy caching.
- A single initial prompt was used and he states he gave no feedback while the model worked. He published the prompt text.
- "Every model, sound, and texture made from scratch" reflects his own description. By his later account the mechanism is stronger than the phrase suggests in one sense and weaker in another: there are no image or 3D asset files at all, because the textures and geometry are generated by shader and function code written by the model.
- The game is genuinely playable in a browser rather than being a recorded demo video.
- The claim says "a full playable FPS video game". The underlying artifact is a compact single-level browser shooter with one objective, cutting three relays, and the prompt itself asked for the scope to be kept small, "like a tiny level". "Full video game" invites a reader to picture a finished commercial product rather than a one-level technical demonstration.
- The post frames this inside "here's what people have already built" and the caption calls him "one engineer", which presents the result as an outside user's independent finding. The demonstration was produced and published by an employee of the company that makes the model, on the day it launched, and the cost and duration figures are that employee's unaudited self-report. That is an interested party's promotional material, not independent evidence.
- The build most readers will click through and play is a version Rieseberg modified by hand after the run, to improve some models and gameplay. The unmodified one-shot build is a separate, less prominent link. The claim's "from a single prompt" is attached to a result that, as commonly encountered, includes human edits.
- The claim reports $550 without the token counts or the caching that produce it. Published arithmetic against list pricing puts the same workload between roughly $197 fully cached and roughly $3,560 fully fresh, so the headline figure is a point inside a wide band that depends on cache behaviour, and it cannot be reconstructed from the public numbers.
- A single curated launch-day run, with no published count of prior or discarded attempts, is presented as a straightforward description of what the model does. Players reported freezes, stuttering, mouse-capture failure and geometry rendering faults, which the claim does not mention.
- Whether the 8 hour duration and the $550 cost are accurate. No billing record, log, or independent rerun has been published, and no third party has reproduced the result.
- Whether the run began from a genuinely empty context. The published prompt opens "you should also build me a FPS", which indicates prior conversation in the same session, and the 445 million input tokens are consistent with a long multi-step agent run rather than a single turn. How much of that preceding context shaped the result is not established.
- What harness, effort setting, tool access, or retry behaviour was used. These settings are not published.
- Whether any attempts were discarded before the published run. Not stated either way.
- How the one-shot build compares in quality and stability to the tweaked build, since the specific changes Rieseberg made were not itemised.
Claude Fable 5.2 is a real model. Anthropic announced it on 9 November 2026 alongside Claude Mythos 5.2, generally available on paid plans and major cloud platforms, priced at $8 per million input tokens and $40 per million output tokens. On launch day, Felix Rieseberg, who works at Anthropic and co-maintains Electron, posted a browser-playable first-person shooter called Saltline. His own wording is that he wanted to test the model "with one-shot prompts without human feedback", that "Everything in this game is from scratch: models, sounds, textures", and that the run was "~$550 at API pricing, ~2.7M output and 445M input tokens, heavily cached. Took 8 hours!" He separately published the prompt, reported as three sentences, which asked for a FPS with "insanely detailed quality", visuals that "can keep up with AAA", and a small scope, "like a tiny level", with photorealistic graphics. The prompt as quoted opens "You know, you should also build me a FPS", the word "also" indicating it sat inside a wider session in which he was generating several games. He also stated that the game contains no image files and no 3D model assets: textures and models are mathematical functions and shaders written in code, inspectable in the page source. Saltline is a compact single-level game set in a rainy container port, with a mission to cut three relays, running as a Claude artifact in the browser. Reaction was divided. Several players praised the visuals; others reported the game freezing or stuttering, mouse capture problems, and rendering faults such as missing undersides on stacked containers. Several commenters objected to the "one prompt" framing on the grounds that 445 million input tokens describes a long autonomous agent run rather than a single conversational turn. TweakTown and Hacker News commenters both record that Rieseberg's post contains two playable links: the more prominent one is a version he subsequently tweaked to improve some models and gameplay, and the untouched one-shot build is a separate link. Third-party arithmetic against Fable 5.2 list pricing finds the same token counts would cost roughly $3,560 to $3,668 if all input were billed fresh and roughly $197 if all input were billed at cache-read rates, so the reported $550 sits between those bounds and cannot be reconstructed exactly from the public numbers.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Is Claude Fable 5.2 a real model and did Felix Rieseberg actually build this game with it?
Yes. Anthropic released Claude Fable 5.2 on 9 November 2026, and Felix Rieseberg, who works at Anthropic, posted a browser-playable first-person shooter called Saltline built with it that same day.
Do the 8 hours and $550 cost figures check out?
They match Rieseberg's own post exactly, but no billing record, log, or independent rerun has been published to confirm them. Third-party arithmetic against list pricing suggests the real cost could range from about $197 to about $3,560 depending on caching, so the $550 figure cannot be verified from public numbers.
Is Saltline really a full video game made from a single prompt?
It is a compact single-level game with one objective, cutting three relays, and the prompt itself asked for a small scope like a tiny level. The prompt also opens with the word 'also,' suggesting it came after prior conversation in the same session rather than starting from nothing.
Is the version people play the actual untouched one-shot output?
Not necessarily. The more prominent link is a version Rieseberg hand-edited afterward to improve models and gameplay, while the untouched one-shot build is a separate, less visible link.
Should this be treated as independent proof of what the model can do?
No. It was produced and published by an Anthropic employee on the model's launch day, framed in a way that can look like an outside user's finding, and the figures are his own unaudited self-report with no count of discarded attempts or details on tool settings.