DeepSeek released DeepSeek-V4-Flash-0731 on Hugging Face under the MIT license on July 31, 2026, describing it as the official release with substantially enhanced agentic capabilities. In the nine-benchmark table published in its own model card, the model is ahead of the higher-tier DeepSeek-V4-Pro preview on all nine rows, with an average margin of 20.3 points (calculated by ASAP from the card's table). Against Claude Opus 4.8 in the same table it trails on all nine rows, with an average gap of 5.7 points. API pricing is documented at $0.14 per million input tokens on a cache miss and $0.28 per million output tokens. ASAP works from the model card and the API documentation to separate what this release changes from what it does not.

The lower tier is ahead of the higher tier on all nine rows

The clearest result for DeepSeek-V4-Flash-0731 is that it leads its own company's higher-tier model on every benchmark in the card. Comparing Flash-0731 against the V4-Pro preview row by row gives Terminal Bench 2.1 at 82.7 against 72.1, NL2Repo at 54.2 against 38.5, Cybergym at 76.7 against 52.7, DeepSWE at 54.4 against 12.8, and Toolathlon-Verified at 70.3 against 55.9. The remaining four rows run the same direction, and the mean margin across all nine is 20.3 points. DeepSeek states in the card that this holds despite the model's far smaller activated parameter count.

The generational jump is larger still. Against the Flash preview in the same table, the nine-row mean rises by 25.6 points, with the largest single move on DeepSWE from 7.3 to 54.4, a factor of 7.5. Cybergym nearly doubles from 38.7 to 76.7. The smallest gain is Agents' Last Exam at 9.4 points, a row where the absolute score of 25.2 is among the lowest in the table.

The nature of the comparison deserves a caveat. V4-Pro appears in the table as a preview build, and DeepSeek's API documentation says Responses API support for the production deepseek-v4-pro arrives in early August 2026. The table therefore sets an official Flash against a preview Pro. Confirming the tier inversion requires redrawing the same table once the production Pro ships.

The gap to Opus 4.8 runs from 0.5 points to 15.5 points

The distance between Flash-0731 and Claude Opus 4.8 is a distribution rather than a single number, spanning 0.5 to 15.5 points. Using the Opus 4.8 figures DeepSeek published alongside its own, the closest rows are Agents' Last Exam at 25.2 against 25.7, a 0.5-point gap, AutomationBench Public at 25.1 against 27.2 at 2.1 points, and Terminal Bench 2.1 at 82.7 against 85.0 at 2.3 points. At the far end sit NL2Repo at 54.2 against 69.7, a 15.5-point gap, and DSBench-Hard at 59.6 against 71.7 at 12.1 points.

The shape of that distribution is what matters operationally. The narrow rows are tasks that call tools in sequence and follow a prescribed procedure, while the wide rows are NL2Repo, which builds repository-scale code from natural-language requirements, and DSBench-Hard, a set of difficult coding-agent problems. The model has closed the distance on execution and remains behind on large-unit design and hard problem-solving.

That split routes adoption two ways. For pipelines dominated by structured tool calls and terminal operations, this table supports treating Flash-0731 as a substitute for a frontier model. For work centered on generating a codebase from requirements, a 15.5-point gap is not offset by price. The conclusion the table supports is routing by task type rather than wholesale model replacement.

The GLM-5.2 comparison belongs alongside it. Eight rows allow a comparison between the two, all except Cybergym, and Flash-0731 leads on all eight with a mean margin of 6.4 points. The widest is AutomationBench Public at 25.1 against 12.9, and the narrowest is Agents' Last Exam at 1.4 points. Ranking inside the open-weight field now turns on differences in task character rather than decimal places.

A cache-hit price one-fiftieth of the cache-miss price

DeepSeek's API documentation prices deepseek-v4-flash at $0.0028 per million input tokens on a cache hit, $0.14 on a cache miss, and $0.28 per million output tokens. The same document prices deepseek-v4-pro at $0.003625, $0.435 and $0.87, exactly 3.1 times Flash on both the miss and output lines. The ratio between cache-miss and cache-hit input is 50 to 1, which means a workload that replays the same prefix and one that does not differ by a factor of 50 on input cost.

That structure matters because it interacts with context length. The documented context window is 1M tokens and maximum output is 384K tokens. Pushing a million-token context through as a cache miss on every call costs $0.14 in input alone, while reusing the same context from cache costs $0.0028. For an agent design that pins a long document or a large codebase as fixed context and varies only the query, this changes the order of magnitude of running costs. For a design where the whole context differs on every call, the list price is what gets billed.

One further condition belongs on the price sheet. DeepSeek's documentation announces that the API will soon adopt peak and off-peak pricing, with prices doubling from 9:00 to 12:00 and 14:00 to 18:00 Beijing time. Those windows land squarely inside the business day across East Asia, so any organization running internal agents on this API should budget at the doubled rate rather than the headline rate. Whether batch work can be shifted to off-peak hours determines the real unit cost.

What the card establishes about the architecture, and what it leaves out

The published configuration establishes 43 layers, a hidden size of 4,096, 256 routed experts plus one shared expert, and 6 active experts per token. Maximum position embeddings are 1,048,576, matching the 1M-token context in the API documentation, weights ship in FP8 E4M3 format, and the vocabulary holds 129,280 entries. The model card states that this release shares its structure with DeepSeek-V4-Flash-DSpark and carries an attached speculative decoding module.

What the card does not establish is the exact total and active parameter count. DeepSeek writes only that the activated parameter count is far smaller than V4-Pro's, without giving a number. The third-party evaluator Artificial Analysis puts the model at 284 billion total and 13 billion active, which is not a figure from the official card and should be attributed separately when cited. The same organization scored the model at 50 on its intelligence index, up 10 points from the previous build.

Operating conditions are specified in detail. Reasoning effort has three levels, low, high and max, and the card recommends a maximum output length of 384K tokens at the high and max levels. Under vLLM, dspark speculative decoding is enabled with a single flag with 7 speculative tokens given as the example value, and the serving example in the card is a single node of 4 GB300 accelerators. The release ships no Jinja chat template and requires a separate encoding script to turn messages into input strings.

Three conditions that limit how far this table can be cited

The first condition is self-reporting. All nine rows are numbers DeepSeek published in its own card, including the scores for the comparison models GLM-5.2 and Opus 4.8. When the party publishing a rival's score is also the party measuring it, outside readers cannot verify how much the configuration choices moved the result.

The second condition is the composition of the benchmark set. Two of the nine, DSBench-FullStack and DSBench-Hard, are internal DeepSeek test sets by the card's own footnote. Scores from evaluations outsiders cannot access make up 22% of the table, and both rows sit among the wider gaps against Opus 4.8. Redrawing the table with public benchmarks alone would change the strength of the conclusion.

The third condition is the evaluation setup. The card states that code-agent tasks were evaluated with the minimal mode of the DeepSeek Harness, and notes that this harness is still to be released. No third party can reproduce those conditions while the tooling remains unpublished. The evaluation also ran at the max reasoning effort level, which increases output tokens, so the table reports scores from the most expensive setting and says nothing about performance at cheaper ones. Multiplying the $0.28 output price directly against these scores produces a cost-performance figure more optimistic than reality.

What the MIT license and the 1M context open up

The materially useful condition in this release is the MIT license. Weights and repository ship under MIT, so commercial use, derivative models and redistribution require no separate negotiation. It is a simpler instrument than Apache 2.0, with less patent and notice review to work through. For finance, public sector and healthcare environments that cannot send data outside their perimeter, licensing is not the obstacle to a self-hosted agent.

The obstacle is hardware rather than licensing. The serving example in the card is a single node of 4 GB300 accelerators, and holding FP8 weights plus a million-token KV cache requires deliberate memory planning. For most organizations the practical options remain the API or a smaller derivative model. That returns the earlier pricing condition to the table: the API brings doubled peak-hour rates and cross-border data transfer, while self-hosting removes both and leaves the hardware bill.

The 1M-token context is the variable that unsettles this choice again. Loading an entire policy corpus, a contract set or a full repository as context reduces the cost of designing and maintaining a retrieval pipeline. Combined with a cache-hit price one-fiftieth of the cache-miss price, pinning a fixed context and varying only the query becomes economically viable. The first thing worth testing here is not the benchmark ranking but whether that configuration holds up at the document scale a given organization actually has.

What to watch next

Three indicators are what settle the verdict on this release: independent reproduction of the benchmark table, the production V4-Pro build, and the rollout of DeepSeek's announced peak pricing. The first is third-party reproduction. As an MIT open-weight model, outside researchers can redraw the seven public benchmarks of the nine, and can match the remaining conditions once the DeepSeek Harness is published. How far self-reported and independent numbers diverge determines the credibility of the whole table.

The second is the production V4-Pro. DeepSeek's API documentation announces Responses API support for V4-Pro in early August 2026. Whether the current inversion, where the cheaper tier outperforms the expensive one, survives that release or gets tidied up will define the structure of DeepSeek's product line. Two tiers priced 3.1 times apart currently sit in reversed performance order.

The third is the actual rollout of peak pricing. The announced doubling windows overlap the East Asian business day, so the timing and scope of enforcement set the real cost of adoption. Comparison tables need to treat $0.14 and $0.28 as off-peak figures to avoid distortion.

Source: ASAP analysis based on the official model card deepseek-ai/DeepSeek-V4-Flash-0731 (Hugging Face, MIT license, July 31, 2026), DeepSeek's official API pricing documentation, and the Artificial Analysis intelligence index