DeepSeek V4 Pro 0813 vs Claude Fable 5: Cascade Routing Cuts Coding Costs

Together AI published a head-to-head comparison of DeepSeek V4 Pro 0813 and Claude Fable 5 on the DeepSWE benchmark, a software engineering test covering 113 tasks across many languages. The run used four trials per task per model, 904 rollouts total, and the results show two models at opposite ends of the price sheet with a surprising amount of disagreement.

First attempt vs retries

Claude Fable 5 wins the single-shot metric: 69.7% pass@1 against DeepSeek V4 Pro 0813's 62.8%, a seven point lead. But that lead disappears with retries. At pass@2 the models are level at 78.5% versus 77.1%, and at pass@4 DeepSeek V4 Pro 0813 pulls ahead, 88.5% versus 84.1%. For a model that costs ninety times as much per rollout, Fable does not hold its first-shot edge under repeated attempts, and it does not own the ceiling either.

The cost gap is the widest Together AI has measured on this board. DeepSeek V4 Pro 0813 costs $0.24 per rollout, Fable costs $21.63. Per $100 spent, Pro solves 260 tasks and Fable solves 3. Median rollout time is roughly even, 35 minutes for Pro and 31 for Fable, because Fable is the most verbose model on the board at 115k output tokens despite taking fewer steps, 79 versus 146.

Where each model wins

Fable's strength is in reasoning-heavy, exact-contract domains. It wins 6 of 8 task categories, led by data modeling and serialization at 88%, 24 points over Pro, and language internals at 78. DeepSeek V4 Pro 0813 takes two categories: stateful reactivity, where it edges Fable 66% to 64%, and concurrency and durability, where it wins 58% to 45%. That concurrency result is Fable's weakest cell, and it is the one domain where the cheaper model is simply the better engineer.

By programming language, Fable wins four of five, with Rust the standout at 85% versus 65%, a 20 point margin and the widest single gap in the matchup. Elsewhere the gap is smaller: Python 70 versus 60, Go 71 versus 67, JavaScript 75 versus 65. DeepSeek V4 Pro 0813 takes TypeScript, 61 versus 57.

Failure profiles differ too. Both models regress the existing test suite in only 11% of failures, well below the GPT family's 20%. But Fable has a larger big-miss share, 18% versus Pro's 10%, meaning when Fable is wrong it is more often far off the mark. Pro fails closer to correct more often, 66% near misses versus Fable's 57%.

The routing case

The most interesting result is the per-task correlation of 0.39, the lowest of any pairing Together AI has measured. The two models both solve 88 tasks, Pro alone gets 12, Fable alone gets 7, and only 6 defeat both. Their union covers 107 of 113 tasks, 94.7%. The disagreements go both ways: Pro sweeps one task four-for-four while Fable never lands it, and Fable four-for-fours four tasks that Pro zeros.

That complementarity makes a cascade attractive. Run DeepSeek V4 Pro 0813 first and escalate to Fable only when the test suite rejects the answer. That solves 82.7% of tasks at $8.28 each, thirteen points above Fable alone at well under half of Fable's per-task price. The cascade also beats a perfect one-shot oracle router, which would reach 78.8%. The order matters: Pro-first costs $8.28, Fable-first costs $21.71 for the same accuracy.

Together AI's conclusion is that Fable is hard to justify as a default, with its premium worth paying only for Rust and serialization-heavy work. DeepSeek V4 Pro 0813 is the value pick with near-Fable first-shot accuracy and a higher ceiling, and the best use of Fable is as a selective escalation behind a low-cost Pro first stage.