PyroDash Cuts a $49 Reasoning Run to $1.78

PyroDash cuts a $49.36 reasoning run down to $1.78 by teaching a small model to recognize when it's stuck, per a paper posted to arXiv (2607.20327).

That's a steep drop. For teams paying per token to run reasoning workloads at scale, a cut like that changes what's worth running in production.

The system exposes a tunable knob for how much of that savings you keep. At one setting (λ=0.05), it hits 64.04% average accuracy across five math benchmarks, 6.36 percentage points above an LLM-only baseline, while cutting cost 20.4%. Push the knob further (λ=0.6) and accuracy falls to 54.55%, but the frozen large model ends up handling just 1.90% of the total tokens.

The mechanism skips the usual router model entirely. Instead of a separate classifier deciding upfront which queries need a big model, the small model emits a control token mid-generation. It fires the moment it hits a wall.

A Collaborate Engine then hands the query and the partial reasoning trace to a frozen LLM, which finishes the job in a single completion. No retraining of the large model, no access to its logits, just one handoff.

Training happens in three stages, ending with cost-aware GRPO. That's reinforcement learning that optimizes the small model's escalation decision against the actual dollar cost of making it.

The bet here is that a model judging its own confidence beats a router guessing from outside it. If these numbers hold beyond math, that makes standalone router models optional, not required. Whether it holds on code or open-ended QA, the paper doesn't say.