On August 13, 2026, DeepSeek released a new model, deepseek-v4-pro (version 0813, the release date). The facts: 1M context, 384K max output, thinking mode on by default; official API pricing is $0.435 per million input tokens and $0.87 per million output tokens, while cache-hit input costs only $0.003625 per million tokens. The official docs also state that overall DeepSeek API pricing will rise significantly in the near future. The model is already available through OpenRouter, and it landed the same day Qwen released its 3.8-series open weights — the two launches collided.
What was released
- The 0813 suffix is a release date, not a minor version: DeepSeek-V4-Pro-0813, following the same naming rule as DeepSeek-V4-Flash-0731 from July 31.
- Minimal release surface: there is no official blog post. The information sources are the official API docs, the WeChat announcement, and the OpenRouter model page. Several users on the HN thread (785 points, 300+ comments) could not find a linkable official page — that is the shape of this release, not an oversight.
- Timeline collision: on August 12, Qwen published the Qwen3.8-2.4T-A95B weights (2.4T total params, 95B active) on Hugging Face and ModelScope. HN commenters read the two dense releases as a head-to-head matchup (source: the HN thread linked at the end); DeepSeek has made no official statement.
Core specs
Official API pricing page (verified 2026-08-13):
| Item | deepseek-v4-pro (0813) | deepseek-v4-flash (0731) |
|---|---|---|
| Context length | 1M | 1M |
| Max output | 384K | 384K |
| Thinking mode | Thinking and non-thinking, thinking by default | Same |
| Input (cache hit) | $0.003625 / 1M tokens | $0.0028 / 1M tokens |
| Input (cache miss) | $0.435 / 1M tokens | $0.14 / 1M tokens |
| Output | $0.87 / 1M tokens | $0.28 / 1M tokens |
| Concurrency limit | 500 | 2500 |
| Features | JSON output, tool calls, Responses API, Anthropic API, chat prefix completion (Beta, non-thinking only), FIM completion (Beta, non-thinking only) | Same |
For reference: V4 Pro output is 3.1x Flash’s output price, and cache-miss input is 3.1x Flash’s. The official spec page lists no vision capability, and the community confirms the model does not accept image input.
The official price hike notice
From the pricing page (verbatim): “We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected.” — pricing will rise overall, the increase is expected to be significant, and the concrete plan will follow official notice.
The notice affects two usage patterns differently:
- Cache-heavy workloads (agent loops re-reading system prompts and history): cache-hit input is $0.003625, and even after a several-fold increase it is likely to remain the cheapest option in its class.
- Long first-pass inputs (cache miss at $0.435): this is the most price-sensitive part and should be re-budgeted after the adjustment.
A community estimate for a typical agentic coding split (750 input / 290 output / 82K cache read, source at the end) puts one request at about $0.000875 versus roughly $0.052 for the equivalent Opus workload — about 60x cheaper. That is a community estimate, not an official claim — DeepSeek has never published cross-model pricing comparisons. Trust your own invoice.
Performance: community-compiled comparison (not officially verified)
There is no verifiable official benchmark page. The numbers below come from the official WeChat benchmark table as transcribed by HN users, plus community aggregations, labeled per source:
Head-to-head vs Fable 5 (community transcription, scrlk’s comment at the end):
| Benchmark | DS-V4-Pro 0813 | Fable 5 (w/ fallback) |
|---|---|---|
| HLE (w/ tools) | 60.0 | 63.0 |
| Terminal Bench 2.1 | 87.9 | 88.0 |
| Cybergym | 83.3 | 83.1 |
| DeepSWE | 62.7 | 70.0 |
| Toolathlon-Verified | 74.1 | 77.9 |
| AutomationBench (Public) | 31.8 | 29.1 |
| DSBench-FullStack | 71.1 | 77.2 |
| DSBench-Hard | 67.2 | 68.3 |
Community geometric-mean ranking (goldenarm’s comment): GPT-5.6 Sol 65.5, Fable 5 64.5, Opus 5 64.0, DS-V4-Pro 0813 62.5, Kimi-K3 62.3, DS-V4-Flash 55.8, GLM-5.2 47.3. By that measure, V4 Pro 0813 sits at the tail of the top tier: it tops no individual chart, but its mean sits within 2-3 points of the top three.
The community’s hands-on reports are split in a way worth noting: one Codex CLI user ran a single same-task comparison (DeepSeek: 12 minutes, $0.12, shipped a bug; Grok 4.6: 3.5 minutes, $1.41, no bug), while another reports “not very good at coding, but quite good at research, evaluation and action — it goes head-to-head with the most expensive models.” Single-run comparisons are not evidence; the HN thread itself makes this point.
Why it matters
- A price-band gap: if the community-transcribed benchmark means hold, V4 Pro 0813 performs within 2-3 points of the top tier at unit prices an order of magnitude lower. DeepSeek has published no comparative pricing, so verify the gap yourself.
- Zero migration cost: the official API offers both OpenAI-format and Anthropic-format endpoints (Base URL
https://api.deepseek.com/anthropic), so Claude Code, Codex CLI, and similar tools connect via environment variables — no toolchain migration. - A cadence signal: Qwen3.8-Max (Aug 3) → V4 Flash 0731 update (late July) → V4 Pro 0813 plus Qwen3.8-2.4T open weights (Aug 12-13). Chinese labs have moved from “chasing SOTA” to shipping a new flagship and open weights on the same day. For API consumers that is a clear downward price signal; for teams doing model selection it means more frequent evaluation cycles.
What still needs verification
- Benchmarks: no official benchmark page exists; the tables above are community transcriptions. Verify via the official WeChat announcement or a future official benchmark page.
- Hike size and timing: the docs say only “near future” and “significant” — no numbers. Current pricing may change at any time; the figures here were verified on 2026-08-13.
- Privacy and data: on OpenRouter the model currently requires enabling “allow paid endpoints that train on request data”; DeepSeek’s privacy policy permits training on prompts and completions. Teams with compliance requirements should evaluate this — it is the main concern in the community discussion.
- Concurrency: 500 is fine for a single team, but large-scale agent fleets should plan isolation.
Set it up and verify yourself
Anthropic-format setup (works for Claude Code / Codex):
export ANTHROPIC_MODEL="deepseek-v4-pro"
export ANTHROPIC_SMALL_FAST_MODEL="deepseek-v4-pro"
export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="你的DeepSeek API Key"
One-line OpenAI-format check:
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-d '{"model":"deepseek-v4-pro","messages":[{"role":"user","content":"用一句话总结 Rust 所有权规则"}],"max_tokens":200}'
Once it works, re-run the one or two benchmarks from the table above that matter most to you (community numbers vs. your own feel can diverge a lot) and record the results — that is exactly the data the “needs verification” section above is asking you to supply.
FAQ
When was DeepSeek V4 Pro 0813 released? August 13, 2026. The 0813 suffix is the release date, following the V4 Flash 0731 naming convention.
How much does DeepSeek V4 Pro 0813 cost? Official pricing (verified 2026-08-13): $0.435/1M input tokens (cache hit $0.003625), $0.87/1M output tokens. The official docs warn of a significant price increase in the near future.
Does DeepSeek V4 Pro 0813 support vision input? No. The official spec page lists no vision capability, and the community confirms text-only input.
Can DeepSeek V4 Pro 0813 plug into Claude Code directly? Yes. DeepSeek offers an Anthropic-format endpoint (Base URL https://api.deepseek.com/anthropic); set ANTHROPIC_BASE_URL, ANTHROPIC_MODEL, and ANTHROPIC_AUTH_TOKEN.
Which camp are you in
The release compresses into three positions. Pick yours:
- A. Team Price: same-tier capability at unit prices an order of magnitude lower — migrate usage now and lock the cost structure before the hike.
- B. Team Execution Layer: planning and design stay on top-tier flagships; V4 Pro handles execution and accepts needing a detailed plan.
- C. Team Wait-and-See: no changes until an official benchmark page and the hike schedule exist — if the current stack works, don’t touch it.
Tell us your pick and why in the comments. We especially want to hear from people who have actually run V4 Pro or V4 Flash in production — one real invoice data point from you (model / task / cost / outcome) is worth ten paragraphs of argument from either side.
If you cannot decide, run the setup commands above tonight, give both models the same task, and post the two results (cost / time / output quality) in the comments. Next week we will compile readers’ results into a follow-up.
If this matters to your team, forward it to the colleague still stuck on “can a Chinese model go into production?”
Related reading
- Qwen3.8-Max Released: 2.4T-Parameter Coding Model, Open Weights Next Week (2026-08-03, Rexai Programming)
- DeepSeek V3.2-Exp sparse-attention breakthrough and its API price cut (Rexai Programming)
- Seven questions for Kimi K3: how far is a Chinese LLM from being truly useful? (Rexai Programming)
Sources
- DeepSeek API official docs “Models & Pricing”, verified 2026-08-13: https://api-docs.deepseek.com/quick_start/pricing/ (specs, pricing, verbatim hike notice)
- DeepSeek API official docs, Anthropic API integration: https://api-docs.deepseek.com/ (Anthropic API section)
- OpenRouter model page deepseek/deepseek-v4-pro-0813, live since 2026-08-12: https://openrouter.ai/deepseek/deepseek-v4-pro-0813
- Hacker News thread “DeepSeek V4 Pro 0813”, submitted 2026-08-12, 785 points: https://news.ycombinator.com/item?id=49274600 (community benchmark tables, price math, hands-on reports, and privacy discussion all come from this thread; commenters are named inline)
- Qwen official Hugging Face weights page Qwen3.8-2.4T-A95B, published 2026-08-12: https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
Price and spec information verified 2026-08-13. The official docs have announced a price increase — re-check the pricing page before publishing.