The funniest thing about AI benchmarks is that the number everyone screenshots is often not the number that shows up on your cloud bill. DeepSeek V4 Flash 0731 just gave us both numbers, which is inconsiderate to hype merchants but useful for people who deploy things on purpose. Artificial Analysis reports that the model scores 50 on its Intelligence Index, a 10-point jump over the April 2026 DeepSeek V4 Flash, while DeepSeek’s first-party API Cost per Task comes in about 60% lower than GPT-5.6 Luna. Somewhere, a spreadsheet just cleared its throat. ## Artificial Analysis puts Flash 0731 near Luna Artificial Analysis says DeepSeek V4 Flash 0731 scores 50 on its Intelligence Index, up from 40 for the previous DeepSeek V4 Flash released in April 2026. That also puts it 6 points ahead of DeepSeek V4 Pro, according to the same Artificial Analysis report. The interesting bit is not just the jump, because benchmarks are confetti unless tied to deployment reality. It is that Artificial Analysis places the model one Intelligence Index point behind GPT-5.6 Luna at 51, close enough that procurement teams may start behaving like raccoons in a pricing dumpster. Artificial Analysis also frames the model against nearby competitors: GLM-5.2 at 51, Gemini 3.6 Flash at 50, Muse Spark 1.1 at 51, and Kimi K3 at 57 as the open weights frontier. That context matters because a score of 50 is not winning the whole weightlifting meet, but it is suddenly standing on the same platform as models many builders would consider for serious agent, coding, and reasoning workloads. The report says DeepSeek V4 Flash 0731 retains a 1M token context window, and that its size remains unchanged from DeepSeek V4 Flash. In plain English: this looks less like a giant new creature and more like the same creature after a very productive weekend with post-training. ## The price story is really a cache story Artificial Analysis says DeepSeek V4 Flash 0731 shares identical architecture and pricing with the earlier DeepSeek V4 Flash, while landing on its Pareto frontier for Intelligence versus Cost per Task. The same report says that even after OpenAI’s 80% price cut on GPT-5.6 Luna, DeepSeek V4 Flash 0731’s Cost per Task on DeepSeek’s first-party API is about 60% lower than GPT-5.6 Luna. That is the kind of sentence that makes benchmark charts feel less like sports talk and more like an invoice with opinions. If two models are roughly comparable on a task suite, the cheaper one does not need to win every bar chart to win a deployment slot. The secret ingredient, according to Artificial Analysis, is DeepSeek’s roughly 98% cache hit discount on its first-party API, compared with the 90% cache hit discount offered by most of the industry. This is where AI pricing stops being simple menu math and becomes coupon stacking at a grocery store run by linear algebra. Cached input can matter a lot for apps with repeated prompts, system instructions, retrieval scaffolding, long context reuse, or agent loops that keep dragging the same backpack of tokens around. Artificial Analysis provider data lists DeepSeek as the available provider for DeepSeek V4 Flash 0731 and shows a $0.06 blended price per 1M tokens, while noting that provider benchmarks such as output speed and end-to-end response time are not available for that page. ## DeepSeek says the API path stays boring, which is good DeepSeek’s API change log says the official DeepSeek-V4-Flash API entered public beta on 2026-07-31, and that the calling method remains unchanged: set the model name to deepseek-v4-flash to use the latest version. Boring compatibility is underrated. Everyone wants wizardry until their orchestration layer starts screaming because a vendor renamed the spellbook. The same DeepSeek change log reports agent benchmark results for the official release, including Terminal Bench 2.1 at 82.7, NL2Repo at 54.2, Cybergym at 76.7, DeepSWE at 54.4, Toolathlon verified at 70.3, Agent Last Exam at 25.2, Automation Bench Public at 25.1, DSBench-FullStack at 68.7, and DSBench-Hard at 59.6. DeepSeek says the official V4-Flash natively supports the Responses API format and is specifically adapted for Codex. It also says DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained. That last clause is the real nerd candy: not every meaningful model improvement requires making the thing bigger, hotter, and more expensive, like a GPU rack cosplaying as a pizza oven. ## What builders should test next Artificial Analysis says DeepSeek is expected to release the model’s full weights in the coming weeks, so local and self-hosted evaluations may become more relevant if that happens. Until then, the practical move is to test your own workload against both quality and cache behavior. A model that looks cheap in a generic calculator can become expensive if your prompts are all one-off snowflakes, and a pricier model can become tolerable if its quality prevents retries. Token caching is not a garnish here, it is the sauce, the plate, and possibly the waiter. For builders, DeepSeek V4 Flash 0731 is a reminder that frontier-model competition is increasingly about total task cost, not just who can bench-press the tallest benchmark. Watch latency data, real coding and agent performance, cache hit rates, and whether the expected weights release changes deployment options. The scoreboard got tighter, but the invoice is where the plot learned to code. ## Sources - DeepSeek V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index, 10 points above previous DeepSeek V4 Flash

Sources