Fireworks AI has released a set of benchmark numbers for DeepSeek-V4.1-Flash, and the takeaway comes down to one line: on agentic coding tasks, the model lands in the same accuracy band as GPT-6 Astra, while costing 15 times less per task.
Start with the scores. On DeepSWE's pass@1 metric, DeepSeek-V4.1-Flash posts 74.34%, GPT-6 Astra 74.12%, Gemini 3.8 Flash 73.83%, and Claude Opus 5 73.65%. All four models are squeezed into a 0.69-percentage-point band, which leaves little room to argue over who's actually first. On Terminal-Bench 2.1, DeepSeek scores 86.5% against Astra's 87.5% — Astra edges ahead here.
The math behind the numbers
Fireworks' per-task cost figures: DeepSeek-V4.1-Flash at $0.430, Gemini 3.8 Flash at $2.362, GPT-6 Astra at $6.524, and Claude Opus 5 at $11.838. Using DeepSeek as the baseline, the other three cost 5.5x, 15x, and 28x as much respectively.
Scale that multiple to a real budget. Take a ten-person engineering team, each running 20 agentic coding tasks a day, 22 working days a month — 4,400 tasks in total. Running on DeepSeek costs roughly $1,892 a month; running on Astra costs roughly $28,706. That's about a $26,800 gap per month, or somewhere around $320,000 a year. This is a rough estimate — real-world token distributions won't match the benchmark exactly — but the order of magnitude holds.
A gap on HLE runs the other way
The same dataset shows a reverse gap elsewhere. On Humanity's Last Exam, DeepSeek-V4.1-Flash scores 34.52% against GPT-6 Astra's 50.40%, a 15.88-percentage-point gap. The parity on coding tasks doesn't carry over to this kind of high-difficulty, general-purpose reasoning.
Fireworks also offers an Oracle Router figure: combine the two models and the theoretical score reaches 54.80%. That's a ceiling, and it assumes you can tell in advance which model should handle which question. No real routing system has that information — the number says more about the two models' blind spots not overlapping than it does about a deployable setup.
On the architecture side
The model itself is a 552B-parameter MoE with input and output activation split apart: 8B activated on the input side, 16B on the output side. The KV cache changes are more direct — HBM usage drops to a quarter of the previous generation, and SSD usage drops to an eighth.
That's where the cost advantage lines up. Agentic coding tasks have long contexts and many turns, so KV cache overhead eats up a much bigger share of cost than a one-off Q&A does. Squeezing cache memory and disk usage down to a quarter and an eighth respectively lets a provider pack more concurrent requests onto the same hardware, which is what makes the $0.43 figure possible in the first place — it isn't just a pricing decision.
Where these numbers come from
Worth stating plainly: this benchmark was published by Fireworks AI itself, and DeepSeek-V4.1-Flash is listed for sale on its own platform — the party running the benchmark and the party selling the model are the same company. DeepSeek has not confirmed these figures, and no independent third party has reproduced them yet.
The choice of benchmark also shapes the conclusion. DeepSWE and Terminal-Bench both lean toward engineering-style coding tasks that overlap reasonably well with day-to-day development work, but swap in a different benchmark or task distribution and there's no guarantee the four models would stay within that 0.69-point band. The cost ratio is worth taking at face value; treating this as a final ranking of model capability is not.
Sources: Fireworks AI technical blog, CocoLoop; DeepSWE, Terminal-Bench 2.1, and HLE scores, per-task costs, and architecture parameters were checked line by line against the original blog post; the ten-person team's monthly spend is a rough estimate.