DeepSeek's new coding model costs 99% less than Claude and scores within two points of it

A Chinese laboratory has released a coding model that performs within striking distance of the best available system and costs roughly one per cent as much to run.

The Release

DeepSeek officially released DeepSeek-V4-Flash-0731 on July 31, moving the model from preview into production and publishing its weights on Hugging Face the same day.

The benchmark results are close to the frontier. On SWE-bench Verified — a test measuring whether a model can resolve real software issues drawn from open-source repositories — V4 Flash scored 79.0 per cent, within 1.8 percentage points of Claude Opus 4.7's 80.8 per cent. On LiveCodeBench it reached 91.6 per cent, one-tenth of a point behind Gemini 3.1 Pro.

The Price Gap Is The Story

V4 Flash costs roughly $0.09 to $0.14 per million input tokens and $0.18 to $0.28 per million output tokens.

Anthropic's Claude Opus 4.8 costs $5 per million input tokens and $25 per million output. That is approximately 30 times more expensive on input and close to 90 times more on output — a difference of roughly 99 per cent for comparable output on coding tasks.

The architecture explains part of it. V4 Flash has 284 billion total parameters but activates only about 13 billion per token, using a mixture-of-experts design that routes each token to the subset of the model relevant to it. Serving costs scale with the parameters actually engaged, not the parameters stored — which is how a large model can be cheap to run.

What It Does To Everyone Else's Pricing

This lands in the middle of the sharpest price collapse the AI industry has seen.

In July alone: OpenAI's GPT-5.6 family launched from $1 per million tokens, then cut its lower tiers by up to 80 per cent on July 30, taking GPT-5.6 Luna to 20 cents per million input tokens. Meta entered the paid market for the first time with Muse Spark 1.1 at $1.25 and $4.25. Moonshot AI released Kimi K3's full weights for free download on July 27. Anthropic retreated on Fable 5 pricing after three deadline extensions under competitive pressure.

The going rate for frontier-class capability has fallen by roughly an order of magnitude in a month, and V4 Flash pushes the floor lower again.

The Caveats Worth Keeping

Two qualifications belong on any account of these numbers.

Benchmark scores are not the same as production reliability. SWE-bench Verified measures resolution of curated software issues; it does not capture how a model behaves across a long, messy engineering task with ambiguous requirements. Independent testing has found V4 Flash strong but not uniformly interchangeable with more expensive models, and at least one detailed comparison concluded the cheaper model was not the right choice despite the cost gap.

And a near-miss on a benchmark is not a tie. Within two points at the top of a difficult evaluation still means losing, and for organisations where the failure cost of a wrong code change is high, the gap may be worth ninety times the price.

What It Establishes Regardless

The strategic point holds even with those caveats. The assumption underpinning frontier laboratory economics has been that the best models command premium prices because nothing else comes close.

V4 Flash, alongside Kimi K3's free weight release and Qwen3.8-Max's autonomy claims, is evidence that something else now comes reasonably close — and at a price that makes the comparison unavoidable for any company running these systems at volume. That is a harder problem for the American labs than any single benchmark result, because it is not about capability. It is about what capability is worth once it is no longer scarce.