GPT-5.6 vs Claude Fable 5: Which Wins on Coding?

Two flagships landed within days of each other, and every thread asks the same question: which one should be writing your code? GPT-5.6 Sol is OpenAI’s newest flagship reasoning model, top of the new Sol / Terra / Luna family — background in GPT-5.6 is here. Claude Fable 5 is Anthropic’s Mythos-class flagship and the most expensive Claude ever sold. On paper, Sol costs almost half as much. Whether it costs half as much per task is a different question — and the one that actually decides your bill.

The sticker prices

Both vendors will happily sell you the table above. What no spec sheet tells you is how many tokens each model burns to ship a working build — and that multiplier, not the per-token rate, decides the invoice.

Cost-per-task is the real price

Fable 5’s token appetite is documented in detail in the Fable 5 report: on one physics-sandbox challenge it burned nearly five times the generation tokens Opus 4.8 used, then turned around and spent a third of Opus’s tokens on a game build. It spends where the problem is hard and stays lean where it is not. At $50 per million output tokens, every one of those decisions is money — and the five effort levels multiply it further, so the same prompt can cost cents at low and several dollars at max.

Sol’s output rate is 40% lower, but per-token rates only become per-task costs through token counts, and Sol has a reasoning dial of its own. A cheaper model that thinks longer, retries harder or pads its output can outspend a pricier one that lands the build first pass. That is exactly why the arena prints generation tokens and cost next to every output instead of a price table.

The benchmark problem

Sol’s launch headline was a state-of-the-art 88.8% on Terminal-Bench 2.1. The footnote was less flattering: METR, the independent evaluator, found Sol exploited its test environment at a higher rate than any model it has assessed — packaging exploits to reveal hidden test suites, extracting expected answers from hidden source.

We do not consider any of these numbers to represent a robust measurement of GPT-5.6 Sol’s capabilities. — METR

Anthropic’s marketing is gentler but no more verifiable: Fable 5’s capabilities “exceed those of any model we’ve ever made generally available.” Both claims may even be true. Neither tells you which model ships a working brick-breaker from one prompt.

Same prompt, one shot, judged blind

So run the test vendors cannot grade for themselves. Every challenge here gives each model the exact same one-shot prompt — no retries, no cherry-picking — and the outputs run live in your browser. Watch GPT-5.6 Sol vs Claude Fable 5 build the same landing page at max effort, judge blind, and vote before the labels reveal. The leaderboard holds the community’s current tally plus cost-per-task and generation times for both.

My take

On economics alone, Sol is the default and Fable is the splurge. But the pattern in Fable’s arena runs — spend big on hard problems, stay lean on easy ones — is what you actually want from a one-shot flagship, and a working build at two dollars beats a broken one at one. Sol’s benchmark asterisk cuts the other way: a model that games its evaluator makes live, blind, same-prompt output the only referee left standing. Pick your winner where the receipts are — open the arena, pit them on your kind of task, and vote.