Together AI: cascading cheap DeepSeek-V4 Flash into GPT-5.6 Luna beats either model alone on cost and accuracy

Together AI published a head-to-head benchmark of DeepSeek-V4 Flash 0731 against OpenAI's GPT-5.6 Luna on DeepSWE, a coding-agent benchmark built from 113 real open-source tasks run across 900 total rollouts — and found that routing tasks through the cheap model first, escalating to the expensive one only on failure, beats running either model alone.

What's new

On raw accuracy, Luna wins outright: Together's data shows "GPT-5.6 Luna leads DeepSWE pass@1 decisively at 67.2% vs 53.3%, a 14 point gap, and holds the lead at every equal attempt count." But DeepSeek's advantage is cost. Together found DeepSeek-V4 Flash is "the cheapest model on the DeepSWE board: $0.10 per rollout vs $0.61, delivering 532 solves per $100 against Luna's 110" — roughly a six-times price gap per attempt.

The more interesting result is what happens when the two are combined in a cascade: run the cheap model first, and only fall back to the expensive one when it fails. Together reports that "running DeepSeek-V4 Flash first and escalating to Luna only on failure solves 78.9% of tasks at $0.385 each: more accurate than Luna alone and 37% cheaper." A cheaper variant of the same idea — giving DeepSeek two attempts before escalating — also beats a single Luna call: "DeepSeek's pass@2 (70.1%) already edges Luna's single shot (67.2%), and two DeepSeek attempts cost about $0.20 to Luna's $0.61."

Context

DeepSeek-V4 Flash 0731 is DeepSeek's low-cost, high-throughput coding model, opened to public API beta in recent weeks with strong agentic benchmark results. GPT-5.6 Luna is OpenAI's higher-end coding-focused variant in the GPT-5.6 family. Together AI hosts both models on its inference platform and has an incentive to demonstrate cost-efficient routing strategies across providers — DeepSWE is a benchmark it has used before to compare hosted models on real coding tasks rather than synthetic ones.

Why it matters

As frontier labs push out increasingly capable but increasingly expensive flagship models, cascading — routing most requests to a cheap model and reserving the expensive one for hard cases — is emerging as a practical cost lever for teams running coding agents at scale. Together's numbers suggest the payoff is not marginal: a two-tier DeepSeek/Luna cascade both beats Luna's solo accuracy and undercuts its cost by more than a third. For any team building an AI coding agent on a budget, that reframes the model-selection question away from "which single model is best" toward "which cascade of models is cheapest for a given accuracy bar."