Baba Is Solved by Fable 5 and GPT-5.6 Sol, but at what cost?

A Hacker News thread titled “Baba Is Solved by Fable 5 and GPT-5.6 Sol, but at what cost?” points to a Quesma post at quesma.com/blog/baba-is-bench/. The claim in the title is direct: systems labeled Fable 5 and GPT-5.6 Sol are presented as having solved Baba, framed not only as a capability win but as a cost question. At the time of the listing, the thread showed 1 point and 0 comments, so the public discussion had barely started.

The linked write-up is framed as a bench piece on Baba rather than a product launch note. That implies the interesting material is how these systems approach a puzzle-style benchmark: what they try, how far they get, and what resources they burn doing it. The title pairs Fable 5 with GPT-5.6 Sol, so the comparison is multi-system, not a single-model victory lap.

For engineers and builders, the useful signal is the cost framing. A solve on a hard puzzle benchmark only becomes an engineering input when someone also reports compute, time, retries, tooling overhead, or dollars. Without that, “solved” is a leaderboard line; with it, teams can decide whether the approach is worth copying, renting, or ignoring for their own agent stacks.

The thread tags Claude, Codex, and Fable in the source line, which puts this in the same arena as coding agents and puzzle/agent evals rather than pure chat demos. Competitive context here is less “who has the biggest model” and more “who can close a known hard game/bench with an agent pipeline that others can inspect or reproduce.”

Practical takeaway: treat the Quesma post as the primary artifact and read it for method and cost, not just the headline solve. Watch whether follow-on discussion adds independent replications, cost breakdowns, or failed attempts under different budgets—those details matter more than the initial 1-point, 0-comment HN listing.