On August 13, 2026, DeepSeek updated its flagship model to the DeepSeek-V4-Pro-0813 snapshot and the lightweight tier to DeepSeek-V4-Flash-0731, with the calling convention unchanged — you still use the names deepseek-v4-pro / deepseek-v4-flash. But the real headline isn't the version number, it's two cards on the table: 1M context plus 384K output, and official native support for the Anthropic API format — meaning the code you wrote to call Claude, the terminal running Claude Code, can switch to DeepSeek by changing two environment variables. A Chinese model has, for the first time, wedged directly into Claude's ecological niche. This is the news hook for our frontier coding model comparison; the comparison covers how to choose, this piece covers what happened and why it matters.
Boundary first: the facts here come from DeepSeek's official API docs (api-docs.deepseek.com) — the "Models & Pricing", "Using the Anthropic API", and "Integrating with Claude Code" pages, fetched 2026-08-13. Pricing and capabilities are per the official site; this is news analysis, not an independent benchmark. It complements without repeating our earlier DeepSeek-V4-Flash hotspot: that one covered the older Flash snapshot, this one focuses on the Pro-0813 + Anthropic API compatibility storyline.
1. The 0813 Snapshot: Two Cards
DeepSeek's snapshot refresh didn't change the calling names — deepseek-v4-pro now points to 0813, deepseek-v4-flash to 0731. What matters is the spec sheet, where two cards long in the making are now squarely on the table.
| Dimension | deepseek-v4-pro (V4-Pro-0813) | deepseek-v4-flash (V4-Flash-0731) |
|---|---|---|
| Context length | 1M | 1M |
| Max output | 384K | 384K |
| Thinking mode | Non-thinking + thinking (default on) | Non-thinking + thinking (default on) |
| Anthropic API | Supported | Supported |
| Responses API / Tool Calls / Json Output | Supported | Supported |
| Input price (cache miss) /M tokens | ¥3 | ¥1 |
| Output price /M tokens | ¥6 | ¥2 |
| Concurrency limit | 500 | 2500 |
Two cards. First, 1M context + 384K output: a one-million-token input window plus up to 384K tokens of single-pass output means you can stuff an entire mid-size codebase or hundreds of pages of docs in at once and have it produce a long rewrite or report in one go, rather than slicing and stitching. Second, thinking mode on by default: both Pro and Flash support "non-thinking + thinking" modes, and the default is thinking — the model emits a reasoning trace before answering, benchmarked against Claude's Extended Thinking and GPT-5.6's reasoning tier.
The supporting capabilities are complete: Json Output, Tool Calls, Responses API, prefix continuation (Beta), and FIM completion (Beta, non-thinking mode only). FIM is "fill-in-the-middle" — what powers inline completion and block insertion in coding.
2. Anthropic API Compatibility Is the Real Blockbuster
If you only look at parameters, 1M context is strong but not unique — Claude and Gemini are also in the long-context race. The real blockbuster is that DeepSeek officially and natively supports the Anthropic API format, with base_url https://api.deepseek.com/anthropic.
What makes this matter is "swap the engine, keep the pipeline." Apps written with the Anthropic SDK, terminals running Claude Code, open-source coding tools that depend on the Anthropic API — they all call Anthropic's messages interface under the hood. DeepSeek turned its own interface into an Anthropic-format compatibility layer: point your base_url here, swap the key for a DeepSeek one, change the model name, and nothing else moves. The official team even built in automatic model-name mapping: pass a claude-opus-prefixed model name and the backend routes to deepseek-v4-pro; pass an unsupported name and it falls back to deepseek-v4-flash. It's a backdoor left open for every tool that hard-codes calls to Claude.
The most direct landing is Claude Code. DeepSeek's docs include a dedicated "Integrating with Claude Code" guide — from scratch install to migration is a few steps (see this batch's DeepSeek + Claude Code SOP): install Claude Code, set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic, put your DeepSeek key in ANTHROPIC_AUTH_TOKEN, set ANTHROPIC_MODEL=deepseek-v4-pro[1m], and claude runs on DeepSeek. This isn't a community hack — it's the officially endorsed path.
3. Pricing: Absurdly Cheap — But a Price-Hike Warning Is Already Posted
DeepSeek's pricing has always been its killer feature, and the 0813 snapshot keeps that position (per million tokens).
| Billing item | deepseek-v4-pro | deepseek-v4-flash |
|---|---|---|
| Input (cache hit) /M | ¥0.025 | ¥0.02 |
| Input (cache miss) /M | ¥3 | ¥1 |
| Output /M | ¥6 | ¥2 |
Two points. First, cache-hit price is negligible: Pro cache-hit input is ¥0.025/M, Flash even lower at ¥0.02/M. Scenarios that reuse the same system prompt and code context (coding agents are the textbook case) have high cache-hit rates, so actual input cost is driven to near zero. Second, Pro is 3x Flash: output ¥6 vs ¥2, input (miss) ¥3 vs ¥1 — Pro costs more for reasoning depth, Flash is cheap for throughput. Pro concurrency is 500, Flash 2500, neatly matching the split of "a few heavy jobs on Pro, massive light jobs on Flash."
But the cold water: the pricing page posts an explicit warning — DeepSeek plans to raise API pricing across the board soon, with a large expected increase, asking users to plan usage accordingly. The current low-price window may narrow. If you'll depend heavily, now is the time to bank usage and push cache-hit rates, rather than betting it stays cheap. The exact plan is subject to official notice.
4. Cold Water and the Landscape
Three buckets of cold water. First, thinking mode is on by default — count the cost. Thinking mode emits reasoning tokens before the answer, and those tokens are billed. Under Pro's thinking mode, actual token consumption is higher than non-thinking, and long runs can cost more than the headline unit price implies. For cost-sensitive scenarios, explicitly switch to non-thinking. Second, Pro concurrency is only 500. Against Flash's 2500, Pro's cap is low; high-concurrency batch jobs (running dozens of agents at once) may hit the limit and have to queue or split batches. Third, Anthropic API compatibility ≠ 100% equivalence. The compatibility layer covers the interface format; the underlying model capabilities, tool-call details, and prompt-caching behavior still differ from native Claude. Complex agent workflows migrated over need regression testing — don't assume "runs" means "performs identically."
On the landscape, this is a frontal position-grab. DeepSeek didn't compete on "my parameters are bigger"; it used Anthropic API compatibility to bore straight into Claude's toolchain — for developers the switching cost drops to "change two env vars," and that's the real threat. On selection: if you want out-of-the-box, don't mind paying, and depend on native Claude behavior, stay with Claude; if you want long context + cheap + able to reuse the Claude toolchain and are willing to regression-test, DeepSeek-V4-Pro-0813 is currently the smoothest substitute; for massive throughput go Flash. The three coexist — none replaces another. Our frontier coding model comparison lines DeepSeek up against Claude and GPT-5.6 item by item for reference.
Common Questions
Q1: What's the relationship between deepseek-v4-pro and DeepSeek-V4-Pro-0813?
A1: The former is the calling name, the latter is the version snapshot. On 2026-08-13 DeepSeek updated the version that deepseek-v4-pro points to DeepSeek-V4-Pro-0813, and deepseek-v4-flash to DeepSeek-V4-Flash-0731. The calling method is unchanged — keep using deepseek-v4-pro / deepseek-v4-flash in your code and you automatically get the latest snapshot.
Q2: What does it mean that DeepSeek supports the Anthropic API, and what's it good for?
A2: DeepSeek officially provides an Anthropic-format compatibility interface, base_url https://api.deepseek.com/anthropic. Apps written with the Anthropic SDK, terminals running Claude Code, and open-source tools depending on the Anthropic API can switch by pointing base_url here and swapping in a DeepSeek key. There's also model-name mapping: claude-opus-prefixed names auto-route to deepseek-v4-pro; unsupported names fall back to deepseek-v4-flash.
Q3: What can 1M context and 384K output do? A3: 1M context lets you load an entire mid-size codebase or hundreds of pages of docs in one shot; 384K max output means a single pass can produce a long rewrite or report without slicing and stitching. Suited for large-codebase comprehension, long-document processing, batch rewrites. Note thinking mode is on by default and thinking tokens are billed — for long tasks, count the total cost.
Q4: Will DeepSeek's prices stay this way? A4: Not necessarily. The pricing page explicitly warns that API pricing is planned to rise across the board soon with a large expected increase, advising users to plan usage. The current low-price window may narrow. For heavy-dependence scenarios, push cache-hit rates now (cache-hit price is minimal — Pro ¥0.025/M, Flash ¥0.02/M). The exact change is subject to official notice.
Q5: Can DeepSeek-V4-Pro directly replace Claude? A5: Not a simple equivalence. Anthropic API compatibility covers the interface format; the underlying model capabilities, tool-call details, and prompt-caching behavior still differ from native Claude. Complex agent workflows migrated over need regression testing. But on the combination of "reuse the Claude toolchain + long context + low cost," V4-Pro-0813 is currently the lowest-switching-cost substitute. For a detailed comparison see our frontier coding model comparison.
References
- DeepSeek official API docs · Models & Pricing (version, context, output, thinking mode, Anthropic API, pricing, concurrency, price-hike warning): https://api-docs.deepseek.com/zh-cn/quick_start/pricing
- DeepSeek official API docs · Using the Anthropic API (base_url, model-name mapping, SDK calls): https://api-docs.deepseek.com/zh-cn/guides/anthropic_api
- DeepSeek official API docs · Integrating with Claude Code (env vars, install steps): https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/claude_code
- Related on this site: Frontier Coding Model Comparison ｜ DeepSeek + Claude Code SOP ｜ Cline + DeepSeek Open-Source Analysis ｜ DeepSeek-V4-Flash Hotspot ｜ Claude Code vs Cursor vs Codex