DeepSeek V4 Flash: Luna-Class Speed at 60% Lower Cost

DeepSeek V4 Flash: Luna-Class Speed at 60% Lower Cost

Categories:AI news

DeepSeek shipped a major upgrade to its budget model on July 31, 2026, and the numbers are worth paying attention to. The new V4 Flash build, tagged "0731," scores 50 on the Artificial Analysis Intelligence Index, ten points above the April version and just one point behind OpenAI's budget tier GPT-5.6 Luna, while costing roughly 60 percent less per task. That gap exists even after OpenAI cut Luna prices by 80%, which makes the release the clearest evidence yet that the budget tier, not the frontier, is where the price war is being decided.

How DeepSeek gets the price gap

The cost difference comes from two places. First, DeepSeek offers a 98% cache discount, above the industry-standard 90%, which makes repeated-context workloads dramatically cheaper, including the long multi-turn sessions that agent workflows generate. Second, the model uses 12% fewer tokens than its predecessor for the same output, which compounds the savings on every request. On a price-to-performance basis, Artificial Analysis now lists V4 Flash as the top model it tracks.

Where it improved

The 0731 build improves across every category the index tracks, with the largest gains in agentic tasks. On GDPval, a benchmark for complex real-world office work, it climbs from 1,189 to 1,559 Elo points, and it hallucinates less than the previous version. The architecture is unchanged: 284 billion total parameters, 13 billion active, with a 1M-token context window. Weights are on Hugging Face under an MIT license, so the model is fully self-hostable.

The budget-tier race

V4 Flash sits at the center of a two-front pricing war. On the API front, OpenAI cut Luna prices by 80% and Anthropic introduced Sonnet 5 at introductory $2/$10 pricing, moves that look directly aimed at the price point DeepSeek defined in April. On the open-weights front, DeepSeek keeps its weights MIT-licensed while larger Chinese labs move to bespoke licenses, which makes it the easiest serious model to self-host. The result is that the cheapest serious AI stack now looks like this: DeepSeek V4 Flash for high-volume work, a frontier model for hard tasks, and a cache-aware routing layer between them. DeepSeek's 98% cache discount is the reason routing matters, since repeated context across agent turns is exactly where that discount compounds.

Who should switch

Teams running large-scale agent workflows, batch summarization, or any workload where token volume dominates should benchmark V4 Flash 0731 against their current budget tier this week. The same teams should verify three things first: that their prompts benefit from the token reduction, that cache hit rates justify the discount structure, and that their reliability requirements can tolerate DeepSeek's operational track record, which includes the widely reported API outage on July 24. For everyone else, the release is a reminder that the budget tier is no longer a quality compromise, it is a deliberate engineering choice.

How the architecture delivers the efficiency

The V4 Flash architecture is a mixture-of-experts design with 284 billion total parameters but only 13 billion active per token. That is the efficiency secret: the full model has breadth across domains, but inference cost tracks the active parameter count, not the total. The 0731 build pairs this with an improved attention mechanism that keeps context-cost low across the full 1M-token window, and the 12% token reduction compared to the April build comes from server-side prompt formatting improvements that work without any client-side changes. On GDPval, the benchmark for complex real-world office work, the improvement from 1,189 to 1,559 Elo points is the largest single-category gain and reflects better agentic task handling specifically. The model also hallucinates less across all tested domains, which matters for production deployments where reliability is the bottleneck, not capability.

The budget-tier race is a three-front war

DeepSeek is not the only player resetting the budget tier. OpenAI cut Luna prices by 80% within days of the GPT-5.6 launch, and Anthropic introduced Sonnet 5 at introductory $2/$10 per million pricing through August 31, a price point that directly competes with Flash for high-volume agentic workloads. The three-way dynamic means the cheapest model this quarter may not be the cheapest next quarter, and each player is optimizing a different variable: DeepSeek on cache discount depth, OpenAI on throughput speed, Anthropic on effort-to-cost configurability. Teams that build model-agnostic routing layers will capture whichever discount wins in a given month; teams that hard-wire to a single budget provider will miss the price volatility that is now a feature of the market.

What the follow-up means

This is the second act of a story this site covered in June, when DeepSeek V4 made AI coding dramatically cheaper. The Flash line is the budget tier: good enough for high-volume agent work, and now close enough to Luna on intelligence that the remaining argument is price and speed, not capability. For teams running large batches of agent tasks, the math is the headline: Luna-class scores at 60% lower cost per task, with an MIT license on top. The main caveats are operational rather than technical: DeepSeek's API had a widely reported outage on July 24, and teams that need five-nines reliability should plan multi-provider fallbacks regardless of price. For teams running long multi-turn agent sessions, the 98% cache discount is the real headline: repeated context across dozens of agent turns means the effective per-token cost can drop well below the sticker price, and routing middleware that optimizes for cache hit rate will capture more savings than model selection alone.