OpenAI-compatible API
The published base URL is https://api.deepseek.com. The model list documents deepseek-v4-flash as an available model ID.
Independent technical reference · reviewed August 2, 2026
DeepSeek
V4 Flash 0731
DeepSeek V4 Flash 0731 is the July 31, 2026 API update to DeepSeek-V4-Flash. This independent reference separates confirmed release details from interpretation and links material claims to their closest source.
01 / Release record
A production API update, not a new chat-app model.
DeepSeek’s release note identifies DeepSeek-V4-Flash-0731 as the official V4-Flash API release in public beta. It says the model retains the preview version’s architecture and size and was re-post-trained. The stated scope is the V4-Flash API; the V4-Pro API and the APP/WEB models were not part of this update.
The 0731 suffix is therefore best read as a release-date identifier for the API update, not a separate consumer product name.
- Released
- July 31, 2026
- Model ID
deepseek-v4-flash- Scope
- V4-Flash API only
- Status
- Public beta at release
02 / Confirmed capabilities
What the published API documentation supports.
Thinking and standard modes
The model-and-pricing documentation lists both thinking and non-thinking modes, with thinking enabled by default at the time of review.
Tool calls and JSON output
The same documentation lists support for tool calls and JSON output. Integration behavior still depends on the calling framework and current API documentation.
Responses API format
The release note says V4-Flash natively supports the Responses API format and is adapted for Codex-oriented workflows.
Context limits, output limits, pricing, concurrency, and compatibility settings can change. Use the live official documentation—not this static record—for implementation decisions.
03 / Integration directory
Eighteen documented paths for using the Flash API.
DeepSeek’s API documentation publishes setup guides for a broad set of coding agents, editors, and terminal tools. Every entry below links to the relevant documentation page—not to a reseller or an unverified proxy.
Most routes use your own DeepSeek API key and a model ID such as deepseek-v4-flash. Compatibility details can matter in thinking/tool-call workflows, so follow each guide rather than assuming generic OpenAI compatibility is enough.
Official integration guideCodex Responses-format integrationOfficial integration guideClaude Code Anthropic-compatible endpointOfficial integration guideCline Editor and CLI agent routeOfficial integration guideCursor IDE provider configurationOfficial integration guideOpenCode Terminal coding workflowOfficial integration guideRoo Code VS Code agent integrationOfficial integration guideAider Terminal pair-programmingOfficial integration guideContinue IDE assistant configurationOfficial integration guideWindsurf Editor model connectionOfficial integration guideGitHub Copilot Copilot Chat model integrationOfficial integration guideGitHub Copilot CLI BYOK agent workflowOfficial integration guideKilo Code CLI and editor extensionOfficial integration guideCrush OpenAI-compatible providerOfficial integration guideOh My Pi Model configuration with compat fieldsOfficial integration guidePi Terminal agent model selectorOfficial integration guideDeep Code DeepSeek-oriented terminal agentOfficial integration guideReasonix DeepSeek-native coding agentOfficial integration guideWorkBuddy / CodeBuddy Desktop model configuration
04 / Reading benchmarks responsibly
Vendor-reported results are useful context, not a universal ranking.
DeepSeek’s release note reports stronger agent-task results than V4-Pro-Preview across the listed benchmarks, including Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon verified, Agent Last Exam, Automation Bench, DSBench-FullStack, and DSBench-Hard.
The same note qualifies its setup: public code-agent benchmarks used DeepSeek Harness minimal mode with max effort, top_p=0.95, and temperature 1.0. Two named DSBench sets are identified as internal. Treat these as release-note measurements under a disclosed setup, then test against your own workload.
05 / Deep reading
The meaningful change is post-training for agent work—not a larger model.
DeepSeek says 0731 retained the preview model’s architecture and size while changing post-training. That places the release in a useful category for evaluators: it is evidence about behavior under a revised training recipe, not evidence that a larger parameter budget suddenly explains the result.
Its reported agent gains should be read beside the disclosed test harness, maximum effort setting, top_p=0.95, and temperature 1.0. Those settings are part of the result. They are not portable defaults for every coding assistant.
Independent tracker and community discussion provide another signal: the update is often described as near the frontier on intelligence-per-cost. Treat that as a hypothesis for a controlled pilot. In production, measure completed tasks, retries, tool-call repair, latency, and cache behavior—not only tokens or an aggregate index.
Official signal
The release note reports agent benchmarks and names the test conditions.
Read the release note ↗Independent signal
Artificial Analysis tracks models under its own index and provider measurements. Results can change as routes update.
Inspect the independent tracker ↗Community signal
Practitioner reports are useful for failure modes and workflow fit, but are not controlled evaluations.
Read practitioner discussion ↗06 / Source ledger
Primary sources used for this page.
Release noteDeepSeek-V4-Flash Update — release scope, re-post-training statement, reported evaluations.Official ↗Model docsModels & Pricing — modes, context/output information, features, and live commercial terms.Official ↗API referenceList Models — model identifier and current availability endpoint.Official ↗Independent trackerArtificial Analysis — independent context only; values and routes can change.Independent ↗Community discussionr/DeepSeek — used to identify practitioner questions and failure modes, never as controlled evaluation evidence.Community ↗
Last reviewed August 2, 2026. External links are marked nofollow. DeepSeek and associated product names are used descriptively; this independent reference is not affiliated with, endorsed by, or operated by DeepSeek. Read the methodology and editorial policy.
07 / Common questions
A few precise answers.
Is DeepSeek V4 Flash 0731 a separate model name?
The official API identifier listed in the model endpoint is deepseek-v4-flash. “0731” identifies the July 31 release update described in DeepSeek’s release note.
Did this update change the DeepSeek web or app model?
The release note says the update only upgraded the V4-Flash API. It specifically states that the V4-Pro API and APP/WEB models were unchanged at release.
Where should developers verify a current integration?
Start with the official release note, model list, and current model/pricing documentation linked above. They are the appropriate source for live availability, limits, and implementation requirements.
What API model ID should an integration use?
The published model list identifies deepseek-v4-flash. A client may expose its own provider field or compatibility setting, so confirm the exact configuration in that client’s current DeepSeek integration guide.
Does the release note guarantee a benchmark result for my team?
No. The reported results belong to the release note’s disclosed harness, effort, sampling, tools, and benchmark setup. Use them to form a test hypothesis, then evaluate completed tasks, retries, latency, and cost on your own workflow.
Are price and context details fixed by this page?
No. Provider terms can change. This reference records what was reviewed, but the current DeepSeek model and pricing documentation is the appropriate source for live limits, cache conditions, and commercial rates.
How should I compare Flash with GLM-5.2 or GPT-5.6 Luna?
Begin with the comparison pages’ source notes, then run a matched test: use representative prompts, tools, retries, token volumes, and success criteria. Do not treat a context-window figure or a single index score as a complete buying decision.
Is this site an official DeepSeek support channel?
No. This is an independent technical reference. For account access, production incidents, billing, and live API behavior, use DeepSeek’s official documentation and support channels.
Is DeepSeek V4 Flash 0731 API compatible with OpenAI-style clients?
The published documentation lists an OpenAI-format base URL at https://api.deepseek.com. Compatibility still depends on the client, endpoint, and feature being used, so validate tool calls and structured output in your intended workflow.
Does the DeepSeek V4 Flash 0731 API support thinking mode?
The current model documentation lists both thinking and non-thinking modes, with thinking enabled by default at the reviewed time. Follow the live documentation for the current switching method and client-specific compatibility requirements.
How does API cache status affect usage costs?
The official price table distinguishes cache-hit input from uncached input and output. Those rates and rules can change, so use the live price table for a current estimate and test your actual prompt-reuse pattern before forecasting cost.