You can use DeepSeek V4 Flash with Claude Code through DeepSeek's officially documented Anthropic-compatible endpoint. As of July 31, 2026, the stable DeepSeek API model ID is deepseek-v4-flash, serving the 0731 official release in public beta. The integration is text-only: Anthropic image and document blocks are explicitly unsupported, and Claude Code remains a third-party client whose behavior can change by version.
Claude Code integration snapshot — verified July 31, 2026
Field Current state DeepSeek model deepseek-v4-flashServed version DeepSeek-V4-Flash-0731, public betaCompatibility surface DeepSeek Anthropic-compatible endpoint Input Text only; image/document blocks unsupported Context / max output Up to 1M / up to 384K at the API level Reasoning Thinking on/off; Flash effort low/high/maxFollow the live official Claude Code integration guide for current environment-variable names and mapping values. They are deliberately not duplicated here because client-sensitive setup can go stale.
Independent site — not affiliated with DeepSeek or Anthropic.
Setup path
- Create or use a DeepSeek API credential.
- Open DeepSeek's Claude Code integration guide and apply its current Anthropic base URL, authentication, and model-mapping variables in your shell or secret manager.
- Select the Flash mapping documented there; do not substitute
DeepSeek-V4-Flash-0731as an API model ID. - Start a new Claude Code session and run a small read-only task before authorizing edits.
- Verify usage and model behavior from the DeepSeek side rather than assuming a Claude-style alias proves which backend served the request.
DeepSeek's compatibility layer maps Claude-style model names. Its Anthropic API guide warns that mappings can silently fall back, so a typo may not fail loudly. Keep the official guide open during setup and recheck it after client upgrades.
What works—and what does not
Text prompts, repository context, and Claude Code's tool-driven workflow can pass through the compatibility endpoint. DeepSeek documents the integration and positions Flash for lighter or subagent routing in its coding-agent material.
Images and document blocks do not work through this endpoint. If a task depends on screenshots, diagrams, or scanned PDFs, use a separate vision/OCR step and pass extracted text into Flash. The evidence and safe routing pattern are in Is DeepSeek V4 Flash multimodal?.
Reasoning and tool-use caveats
DeepSeek supports thinking on/off and low/high/max effort for Flash. Its thinking-mode guide notes that some sampling controls are ignored in thinking mode and that reasoning_content must be replayed according to the documented rules around tool calls. A client or proxy that strips the wrong field can make failures look like a model defect.
When tools loop, writes truncate, or the model forgets recent decisions, capture the client version, tool schema, request sequence, reasoning mode, context size, and served model revision. Community reports show the same model succeeding or failing differently across harnesses, but those observations do not isolate whether the cause is model, client, prompt, or provider state.
A safer coding workflow
- Start with a bounded task and explicit acceptance criteria.
- Require tests or checks before the agent declares completion.
- Review the diff in a fresh context for high-risk changes.
- Keep planning, implementation, and review separable so a failure is diagnosable.
- Route screenshot or image work to a vision-capable model rather than pretending Flash can see it.
- Re-run your evaluation after the 0731 switch because post-training changed without changing the stable API ID.
Claude Code is a harness, not a benchmark constant. A community comparison in this corpus found similar diffs but large differences in speed, tokens, and tools among Claude Code, OpenCode, and Pi; another reported a bug solved only after moving the same model into Claude Code. Treat these as workflow signals, not proof that one harness always wins.
Cost and context controls
Large repository sessions can repeatedly send substantial context. Inspect cache-hit, cache-miss, output, reasoning, retries, and accepted diffs rather than quoting the minimum input rate. Current rates and examples are in DeepSeek V4 Flash pricing.
The API advertises up to 1M context, but Claude Code or its configuration may impose a smaller practical limit, and capacity does not guarantee recall. See DeepSeek V4 Flash context window before using the window as a substitute for retrieval or session hygiene. General endpoint and error details live in DeepSeek V4 Flash API.
For the dated model-family overview and release status, see DeepSeek V4 Flash.
FAQ
Can Claude Code use DeepSeek V4 Flash?
Yes. DeepSeek publishes an official integration guide using its Anthropic-compatible endpoint.
What model name should I use?
The DeepSeek API ID is deepseek-v4-flash. Use the current mapping from the official Claude Code guide; do not use the Hugging Face version name as an API ID.
Can Claude Code send screenshots to Flash?
No. DeepSeek's Anthropic compatibility table marks image and document inputs unsupported.
Is this the same as using Anthropic's models?
No. Claude Code is the client; DeepSeek supplies the compatible endpoint and model. Compatibility does not mean identical model behavior or feature coverage.
Why did my cost rise on a long coding task?
Repository context, output reasoning, tool turns, and retries add billable tokens. Track completed-task cost, not just the cache-hit headline.
Return to the complete DeepSeek V4 Flash overview for price, benchmarks, download, context, and local-use guidance.
Independent informational website. Not affiliated with, endorsed by, or sponsored by DeepSeek or Anthropic. DeepSeek, Claude, and related marks belong to their respective owners. Last verified July 31, 2026.