MANY MINDEDDeepSeek's cheap model scores within a point of OpenAI's budget tier
DeepSeek has released V4 Flash 0731, an update to its budget model. On the Artificial Analysis Intelligence Index the new version scores 50, ten points above the V4 Flash that launched in April 2026. That leaves it a single point behind OpenAI's budget model GPT-5.6 Luna while costing roughly 60% less per task — a comparison that already accounts for OpenAI's own 80% price cut.
Much of the cost gap is not raw model efficiency but caching. DeepSeek discounts cached input by 98%, against the 90% that is standard across the industry, and the model also spends 12% fewer tokens than its predecessor to do the same work, which compounds the saving. Scores improved in every tested category, with the largest gains on agentic tasks: on GDPval, a benchmark built around complex real-world office work, it climbs from 1,189 to 1,559 Elo. It also hallucinates less often. The architecture is unchanged — 284 billion total parameters with 13 billion active, and a one-million-token context window.
The weights are on Hugging Face under an MIT license, and that is the part that does the most to cost. Near-frontier reasoning anyone can download and run puts a ceiling on what a closed provider can charge for the same tier of work — the same pressure sitting behind OpenAI's own 80% cut.
The caveat is that these are benchmark figures from a single index published within days of release, and real price-per-task depends on how much of a given workload actually hits the cache.
Source: The Decoder