Fireworks releases DeepSeek-V4.1-Flash: GPT-6 Astra coding accuracy at one-tenth the cost

Fireworks released DeepSeek-V4.1-Flash last week, and the company's full benchmark suite shows the model establishes a new Pareto frontier across the three critical parameters: quality, speed, and cost. On DeepSWE, which measures autonomous coding accuracy, the model matches GPT-6 Astra's accuracy at one-fifteenth the cost per task. The practical upshot: frontier-grade coding agents become viable as continuous background infrastructure, not just expensive point tools.

Split execution architecture tailored for code agents

Under the hood, DeepSeek-V4.1-Flash is the smallest member of DeepSeek's new architecture family, and it brings a fundamental architectural improvement: a 552-billion-parameter MoE with a split encoder-decoder architecture. Instead of a single activation budget for every forward pass, the model allocates 8 billion active parameters to input and 16 billion to output. That asymmetry mirrors the actual workload of coding agents: in a typical trajectory the agent repeatedly reads files, tool outputs, and its own history, then emits a relatively small patch. DeepSWE measured a 174:1 input-to-output token ratio — 36.9 million input tokens versus 211 thousand output tokens per task. The new architecture lets the model devote half its compute to the heavy input side, where most tokens accumulate anyway.

KV cache optimization that slashes real-world cost in long loops

The second improvement concerns KV cache utilization: compared with the previous-generation DeepSeek V4, the new model consumes one-quarter the HBM and one-eighth the SSD storage. That figure matters because in long agent loops, most spend isn't on tokens the model generates but on re-reading the same context over and over. A 4x reduction in cache memory translates directly to a 15x drop in cost per task at a 74% DeepSWE score. The 8x reduction in SSD overhead completes the picture by keeping cache hit rates near 99% across long trajectories.

The DeepSWE numbers: same quality band, massive cost gap

On DeepSWE the four leading models land in the same quality band, all within a 0.7-point range on pass@1, while run-to-run variance spans 1.4 to 3.2 points. The difference is entirely in cost: DeepSeek-V4.1-Flash on Fireworks delivers that accuracy at $0.43 per task — 5.5x cheaper than Gemini 3.8 Flash, 15x cheaper than GPT-6 Astra, and 28x cheaper than Claude Opus 5. Put differently, you can run 15 autonomous coding attempts for the price of a single Astra call.

Terminal-Bench 2.1: one point behind Astra at 1/12 the cost

On a different benchmark, Terminal-Bench 2.1, the picture is similar: DeepSeek-V4.1-Flash trails GPT-6 Astra by a single point but does so at one-twelfth the cost. The combination of split execution architecture and cache optimization lets the smallest model in the family close gaps with significantly more expensive closed models, turning the economics of autonomous coding agents into something engineering teams can run continuously without draining the budget.