DeepSeek Now Speaks Both: Responses API Joins the Anthropic Format

DeepSeek’s Responses API support landed on July 31, 2026 with the V4-Flash-0731 release — and with it, the same model now speaks both of the agent-API dialects that matter: OpenAI’s Responses format, adapted specifically for Codex, and Anthropic’s Claude-compatible message format, which DeepSeek has quietly supported since August 2025. No translation proxy, no forwarding layer — a base URL and a few environment variables.

The nuance most same-day coverage will miss: this is not two simultaneous launches. The Anthropic format shipped with DeepSeek-V3.1 on August 21, 2025 and already covers both current model tiers. What is genuinely new today is the Responses API — and only on deepseek-v4-flash, with V4-Pro support a DeepSeek-stated “early August 2026” promise, not a delivered feature. Read as a whole, July 31 is the day the convergence became complete rather than the day it began.

This guide covers what actually changed, the eleven-month history behind the Anthropic endpoint, the community-proxy era this release retires, a feature-by-feature compatibility matrix across both dialects, drop-in setup for Codex CLI and Claude Code, and what the two-dialect standard means for multi-provider routing. Every pricing and capability figure below comes from DeepSeek’s primary documentation or independently from Artificial Analysis, each labeled as such.

Per DeepSeek’s own change log, July 31 promoted V4-Flash from Preview to an official public-beta release as a re-post-trained checkpoint named DeepSeek-V4-Flash-0731 — same architecture and parameter count as the Preview build, retrained. The API-surface news rides on that checkpoint: the change log states the model natively supports the Responses API format and is “specifically adapted for Codex.” The Responses API guide opens just as plainly: “To meet the demand for Codex, our API now supports the Responses API format.” That is vendor documentation copy, not an executive announcement — no named DeepSeek spokesperson has commented on this release.

Just as important is what did not ship. V4-Pro was not updated on July 31 — it remains on Preview status, with DeepSeek’s docs saying its official release “will follow soon.” On the live pricing table, Responses API support is marked with a check for deepseek-v4-flash and an explicit cross for deepseek-v4-pro, footnoted with a DeepSeek-stated plan to add Pro support in early August 2026. The consumer app and web models are unchanged. We covered the V4-Flash-0731 release itself and its agent benchmarks separately — this post is about the API-surface story.

Because the July 31 change-log entry mentions both dialects in one breath, it is easy to read “DeepSeek now speaks the Anthropic format” as news. It is not. DeepSeek’s V3.1 release notes, dated August 21, 2025, list Anthropic API format support alongside that generation’s 128K context and strict function calling. By the time the Responses API arrived, DeepSeek had been serving the Anthropic dialect for roughly eleven months — a maturity signal, not a rushed bolt-on.

The Anthropic-format endpoint lives at https://api.deepseek.com/anthropic — a distinct base URL from the OpenAI-format https://api.deepseek.com — and it covers both current tiers. Claude model names map automatically: names starting claude-opus-* route to deepseek-v4-pro, names starting claude-sonnet-* or claude-haiku-* route to deepseek-v4-flash, and unrecognized names default to deepseek-v4-flash. That mapping is what lets Claude Code run against DeepSeek with nothing but environment variables — the tool keeps asking for Claude models by name and DeepSeek answers with its own.

The pain this release removes is concretely documented. A developer bug report filed on May 3, 2026 describes the exact failure: pointing Codex CLI at DeepSeek’s official API with wire_api = "responses" returned HTTP 404, because DeepSeek’s endpoint spoke only OpenAI-compatible Chat Completions while Codex rejects the chat wire protocol and requires the Responses format. The two tools were speaking different dialects of nominally the same vendor’s API shape.

The community answer was a translation layer. DeepSeek’s own awesome-deepseek-agent repository documented the open-source Moon Bridge forwarding proxy — a local server on 127.0.0.1:38440 that translated between Codex’s Responses wire protocol and DeepSeek’s Chat Completions API. Workable, but one more process to babysit, one more failure surface in every agent run, and one more thing to explain in an incident review. As of July 31, that entire layer is obsolete for V4-Flash workloads: the official endpoint speaks Responses natively, including the apply_patch custom tool that Codex-style coding agents use to write file diffs — described in-doc as built for Codex compatibility.

If your team re-points integrations at DeepSeek regularly, note this is the second API-surface change in a month — the legacy deepseek-chat and deepseek-reasoner alias retirement landed a week earlier, and our V3.2-to-V4 migration playbook covers the model-level move for teams still on the older stack.

Most coverage repeats DeepSeek’s bullet-point summary — “supports OpenAI and Anthropic formats” — without listing what each endpoint explicitly does not support. That unsupported list is exactly what an engineering team needs before committing an agent harness. The matrix below is compiled feature-by-feature from DeepSeek’s two dialect guides and the live pricing table, all retrieved July 31, 2026.

Source: DeepSeek’s Responses API guide, Anthropic API guide, and Models and Pricing table, all retrieved July 31, 2026. Two rows deserve a second read. If your Claude Code workflows lean on image or document content blocks, or on MCP tool-use blocks, the DeepSeek Anthropic endpoint will silently drop capability you currently rely on — those are real functional gaps, not edge cases. And on the Responses side, the absence of previous_response_id means any harness that leans on OpenAI-hosted conversation state must resend full context on every turn.

DeepSeek publishes official integration guides for both tools — the setups below are theirs, not community folklore. Neither requires a proxy, a wrapper SDK, or a code change to the agent itself.

The asymmetry in the two tags is the practical takeaway. DeepSeek’s Codex guide states plainly: “Currently, only deepseek-v4-flash supports integrating with Codex.” A Claude Code harness, by contrast, can reach both tiers today — including routing its opus-class requests to deepseek-v4-pro. Of the two dialects covered here, the Anthropic-format one is the only route to DeepSeek’s stronger tier this week.

Step back from the config files and a larger pattern resolves. Until recently the industry treated “OpenAI-compatible” as the one interoperability standard worth shipping — the thesis we traced in our analysis of the OpenAI-compatible API standard. What July 31 confirms is that the standard has quietly become plural. Agent tooling split the wire-format world in two — Codex and its descendants require OpenAI’s Responses shape, Claude Code and its ecosystem speak Anthropic’s messages shape — and a serious model provider now has to answer both. DeepSeek is the clearest signal yet: a third frontier lab, after the two Western incumbents whose formats these are, speaking both dialects natively on a single model.

The direction of travel among Chinese labs makes the pattern hard to dismiss as one vendor’s quirk. Secondary reporting through 2026 has tracked several — Moonshot, Alibaba’s Qwen team, Zhipu among them — converging on OpenAI-SDK-compatible endpoints on day one, with a growing subset also exposing Claude-compatible endpoints usable directly inside Claude Code. We hedge that as industry-trend framing rather than per-lab verified fact; DeepSeek’s specifics are the primary-verified core here. But the strategic logic was stated openly months before this release, by Zhipu’s chief executive:

Project that logic forward and the convergence looks less like courtesy and more like distribution strategy. If the API is the business model, then the agent harnesses developers already run — Codex, Claude Code, and their forks — are the distribution channels, and speaking their native dialects without a proxy is shelf placement. The plausible next step is that dual-dialect support stops being news at all: it becomes a checklist item every serious model launch carries, the way OpenAI-SDK compatibility became table stakes during 2025. The cost of that world falls on harness builders, who can no longer assume one wire format — and the benefit accrues to teams that treat models as swappable capacity behind stable tooling.

Compatibility only matters if the economics justify the re-point. The primary pricing table (retrieved July 31, 2026) prices deepseek-v4-flash at $0.14 per 1M input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per 1M output tokens; deepseek-v4-pro runs $0.435 cache-miss input, $0.003625 cache-hit input, and $0.87 output. In ratio terms, Flash output costs 32% of Pro’s $0.87 output rate, and Flash cache-miss input is likewise 32% of Pro’s $0.435 — a consistent roughly one-third relationship across the two metered dimensions.

How should a routing program absorb this? The framework we laid out in our guide to multi-provider model routing treats each provider-dialect pair as a route with a capability profile and a price; DeepSeek just collapsed two routes into one provider. Paired with a cost-optimization routing discipline, the decision tree looks like this:

One operational note for teams doing this evaluation seriously: treat the swap as a controlled experiment, not a migration. Same prompts, same repos, measured completion quality and token spend on each route. If you want a partner for exactly that comparative eval — across DeepSeek, the Western frontier, and your actual workloads — our AI transformation engagements start there.