Fireworks says DeepSeek-V4.1-Flash matches Astra on coding tests for far less

What Fireworks is claiming

Fireworks, a company that hosts AI models for developers, put out a post on 14 September 2026 about DeepSeek-V4.1-Flash, a model it released the week before. As fireworks.ai reported, the model reaches the coding accuracy of GPT-6 Astra on a benchmark called DeepSWE while costing 1/15th as much per task.

DeepSWE is a test of software engineering work, where an AI agent has to fix real coding problems. Fireworks compared four models and measured pass@1, meaning the share of tasks solved on the first attempt. All four finished within 0.7 points of each other, while repeat runs of the same model varied by 1.4 to 3.2 points. In other words, the quality differences are smaller than the noise.

That is why the post puts the cost gap front and center. DeepSeek-V4.1-Flash ran at $0.43 per task on Fireworks. By the post's math, that is 5.5x cheaper than Gemini 3.8 Flash, 15x cheaper than GPT-6 Astra and 28x cheaper than Claude Opus 5.

Why the design is cheap to run

The model is a mixture-of-experts system with 552B parameters. That means only a slice of its internal parts switches on for each step, which saves computing power. DeepSeek's twist, according to Fireworks, is to give input and output separate budgets: 8B active parameters while reading, 16B while writing.

This suits coding agents because they read far more than they write. In the DeepSWE run, the model took in 36.9M input tokens per task and produced only 211K output tokens, a ratio of 174:1. Tokens are the small chunks of text that AI models are billed on. An agent keeps re-reading files, its own notes and tool results, then writes a fairly small patch.

Spending less compute on the reading side therefore targets where most of the work happens.

The memory trick behind the bill

The second change is to the KV cache, a store of what the model has already processed so it does not redo that work on every step. Fireworks says the new version needs only a quarter of the HBM (the fast memory attached to AI chips) and an eighth of the SSD storage used by the previous DeepSeek V4.

This matters because of how agent bills actually break down. In the DeepSWE run, 99.6% of input tokens were cache hits, and cache-hit charges made up 60% of the bill. Many people assume they mostly pay for what a model writes. In a long agent loop, though, the larger cost is reading the same context again and again.

So a leaner cache goes after the biggest line item. Fireworks credits it, along with the other design choices, for the 15x drop in cost at a similar score.

Where it falls short

On Terminal-Bench 2.1, a test of command-line tasks, the model finished one point behind Astra. It wrote about 4x more output tokens than Astra, so it works through more text, yet Fireworks puts its total cost per task at 12x lower once uncached input, cache hits and output are all counted.

The picture changes on Humanity's Last Exam (HLE), a multimodal test of graduate-level academic questions. Run alone, DeepSeek-V4.1-Flash scored 34.52% while Astra scored 50.40%. Fireworks itself says that if your workload is only hard scientific reasoning, this is probably the wrong choice.

There is a twist. The two models miss different questions. An oracle router, an idealized setup that always sends each question to the cheapest model that gets it right, beat Astra alone by 4.4 points. Fireworks admits a real router would capture less of that gain. It only shows that mixing models can beat one model on its own.

How much to trust it

Fireworks sells access to this model, and every figure here comes from its own testing. The post also leaves out how each rival was configured, so the cost ratios are best read as a strong lead, not a settled fact. The test is per task, which is the right unit for agents, but your tasks may look different from DeepSWE's.

Availability is straightforward: the model runs on Fireworks' serverless service or on dedicated capacity for production use. The company says US-hosted endpoints for regulated industries are coming soon, and it is inviting people to ask about training support.