DeepSeek V4 Flash 0731 reaches the current frontier at a fraction of the cost

The update was published on 31 July 2026 and is currently being delivered through the existing deepseek-v4-flash API identifier. DeepSeek describes the release as a public beta rather than a completely new architecture. The company says that V4 Flash 0731 retains the same architecture and model size as the April preview, but has been improved through additional post-training.

That distinction matters. Customers do not need to change their API integration, but they may see different responses and benchmark performance from the same model name. According to DeepSeek’s official change log, the new version is now the model served when developers select deepseek-v4-flash.

A ten-point jump on Artificial Analysis

Artificial Analysis currently gives DeepSeek V4 Flash 0731 an Intelligence Index score of 50, compared with 40 for the April DeepSeek V4 Flash. The index combines a range of evaluations covering reasoning, coding, scientific problem-solving, knowledge, agentic tasks and factual reliability.

In practical terms, the update moves DeepSeek from a strong value-oriented open-weight model into the group of systems that compete directly with leading proprietary offerings. It is listed one point below GPT-5.6 Luna in maximum reasoning mode and level with Gemini 3.6 Flash. It also sits six points above the current DeepSeek V4 Pro entry in the same comparison.

These results should not be interpreted as a universal ranking for every use case. Composite indexes depend on test selection, prompting, effort settings and the amount of reasoning a model is allowed to use. Artificial Analysis itself presents model scores as a way to compare broad capability, not as a guarantee that one system will win every workload. Its earlier listing for the April model recorded a score of 40 and approximately 230 million generated output tokens during evaluation. The updated result indicates both higher capability and better token efficiency.

The most visible improvement is in agentic work: tasks in which the model must plan, use tools, inspect files or complete multi-step coding assignments. DeepSeek reports a Terminal-Bench 2.1 result of 82.7, while its newer agent evaluations also show strong results on NL2Repo, Cybergym, DeepSWE and Toolathlon. Some of these datasets are internal or were tested with DeepSeek’s own harness, so they should be treated as useful indicators rather than directly interchangeable league-table scores.

The same 284-billion-parameter model, but more efficient

DeepSeek V4 Flash 0731 remains a Mixture-of-Experts model with 284 billion total parameters and 13 billion active parameters during inference. It supports a context window of up to one million tokens, which is roughly enough for very large codebases, long technical documents or extensive project histories in a single working context.

The architecture is unchanged from the preview version. DeepSeek’s technical material describes the V4 Flash family as using compressed and sparse attention techniques designed to reduce the memory and compute cost of long-context inference. The model is text-only: it accepts text and produces text, with no native image input or image output listed in the current technical specifications.

DeepSeek also says that the new evaluation required approximately 206 million output tokens, down from about 234 million for the April model. That represents a reduction of roughly 12 per cent. The result is important for two reasons. Fewer reasoning tokens can reduce API spending, and shorter internal deliberation may also improve latency for agentic applications.

Pricing remains the main competitive advantage

The official DeepSeek API price has not changed for V4 Flash. It is currently:

• $0.14 per million input tokens when the prompt is not already cached;
• $0.28 per million output tokens;
• $0.0028 per million input tokens for a cache hit.

Using the latest available ECB reference rate of €1 = $1.1392, these prices are approximately €0.12 per million uncached input tokens, €0.25 per million output tokens and €0.0025 per million cached input tokens. The conversion is indicative and excludes VAT, payment charges and any intermediary provider markup. The official pricing page also warns that DeepSeek plans to introduce peak and off-peak pricing, with further details to be announced.

For European developers, the cache price is particularly significant. Long-running agents often resend the same system instructions, repository context or tool definitions. If those repeated sections qualify for caching, the input cost can fall dramatically. However, cache hits are not automatic savings for every request: developers need to structure workloads so that repeated prefixes remain identical, and they should confirm how caching interacts with user isolation and sensitive information.

Available in Europe, but not automatically EU-hosted

DeepSeek V4 Flash 0731 is available to European users through DeepSeek’s official API and the company’s existing web services. The API supports OpenAI-compatible calls, an Anthropic-compatible interface and, for Flash, the Responses API. Developers can therefore test the model without rebuilding their applications from scratch.

Availability in the EU should not be confused with EU data residency. DeepSeek’s public documentation does not establish that API traffic is processed exclusively inside the European Economic Area. Businesses sending personal data from Europe must therefore assess the provider’s terms, retention practices, security controls and international-transfer mechanisms before using production data.

The GDPR requires appropriate safeguards when personal data leaves the EU. The European Data Protection Board identifies Standard Contractual Clauses as one possible mechanism, but their use does not remove the need for a transfer-impact assessment and suitable technical measures. The practical rule is simple: use synthetic or anonymised data for initial testing, and do not place confidential customer records into the API until the legal and security review is complete.

The EU AI Act adds another layer for companies distributing or integrating general-purpose AI. The European Commission’s guidance says relevant provider obligations include technical documentation, copyright policies and training-data summaries, while additional duties may apply to models classified as presenting systemic risk. Those obligations primarily concern providers and deployers according to their role, so an organisation using DeepSeek as an API does not automatically inherit every duty imposed on the model developer.

What developers should take from the release

DeepSeek V4 Flash 0731 is most compelling for applications that need a combination of high reasoning ability, tool use and low operating cost. Coding agents, internal research assistants, document-processing pipelines and long-context analysis are obvious candidates.

It is less suitable when multimodal input is essential, when a company requires guaranteed EU-only processing, or when contractual support and compliance documentation are more important than raw token economics. Developers should also remember that the model is in public beta and that the same API name can point to future revisions.

For teams already using the April V4 Flash, the sensible approach is controlled A/B testing. Compare completion quality, tool-call reliability, latency, token consumption and failure recovery on representative European-language workloads. The headline score of 50 is notable, but production performance will depend on the prompts, tools and data that surround the model.