DeepSeek-V4-Pro-0813: Pro Leaves Preview

DeepSeek did not ship a splashy 0813 launch post. It pinned DeepSeek-V4-Pro-0813 on the official pricing table as the model version behind deepseek-v4-pro, and OpenRouter dated the listing August 12, 2026. That is the GA. Same 1M / 384K window as Flash-0731.

DeepSeek now uses time-of-day pricing. Peak rates are $1.32 input / $0.044 cached input / $3.96 output per 1M tokens; off-peak rates are half. This is Pro leaving preview, not a new architecture drop.

At a Glance

• Catalog ID: deepseek-v4-pro-0813
• API ID: deepseek-v4-pro (version DeepSeek-V4-Pro-0813)
• Release signal: ~August 12, 2026 (pricing-table pin + OpenRouter dated listing)
• Context / max out: 1M / 384K
• Thinking: on by default, default effort high
• Features: JSON output, tool calls, Responses API, Anthropic API, Chat Prefix Completion (Beta), FIM (Beta, non-thinking only)
• Concurrency: 500 (Flash is 2500)
• Base URLs: api.deepseek.com and api.deepseek.com/anthropic
• OpenRouter: deepseek/deepseek-v4-pro-0813

What's New

Pro leaving preview, not a new family. Flash graduated July 31 as DeepSeek-V4-Flash-0731. 0813 is the promised Pro follow-through: a quiet GA via version pin and a dated gateway listing, not a new architecture card.

The April V4 preview note still frames the family; see DeepSeek's preview release note. Treat 0813 as the dated sibling that puts Pro on the same GA cadence Flash already hit.

Pricing

Official table vs Flash, from api-docs.deepseek.com/quick_start/pricing:

Cache-miss input is 30× the cache-hit rate in both periods. Multi-turn loops that miss cache pay for it. Set up caching before you scale Pro traffic, and account for the peak schedule in cost estimates.

Architecture

April 2026 preview materials describe DeepSeek-V4-Pro as a 1.6T total / 49B active MoE with hybrid long-context attention. Flash: 284B total / 13B active. These are preview-era published specs until DeepSeek ships a distinct 0813 card. Do not treat the April numbers as a confirmed description of the 0813 API build.

Thinking / Effort

Thinking is enabled by default with effort high. Requested effort maps to a smaller actual set: low stays low; medium, high, and xhigh land on high; max stays max. Chain-of-thought returns in reasoning_content. Tool-call turns must pass that field back or the API returns 400.

Same mapping applies to Flash and Pro. Re-measure token-per-task at the default high effort before you assume preview-era cost models still hold.

Weights

Hugging Face still hosts the April preview at deepseek-ai/DeepSeek-V4-Pro (MIT). There is no confirmed separate 0813 open-weight repo. API dated builds have been shipping first (Flash-0731, now Pro-0813). Self-host needs a confirmed weight drop; do not assume April HF artifacts equal the 0813 API.

When to Use It

Good fit

• Long-horizon Pro loops that need the 1M context contract
• OpenAI Chat Completions / Responses or Anthropic-compatible clients
• Cache-heavy multi-turn work where hits cost $0.022–$0.044 per 1M

Prefer Flash

When throughput and cost matter more than Pro headroom. Flash has 2500 concurrency and lower per-token rates at the same 1M / 384K window.

Migrating

If you already call deepseek-v4-pro, you are on 0813 after the pin. Five checks before you scale:

• API id stays deepseek-v4-pro; the version behind it is now DeepSeek-V4-Pro-0813.
• Re-measure token-per-task at default effort high.
• Pass reasoning_content on tool-call turns or expect 400s.
• Do not assume April HF weights equal the 0813 API build.
• Account for weekday peak and off-peak pricing.

Outlook

0813 is a version pin that ends Pro preview under the same 1M contract Flash already graduated into. There is no official 0813 bench card, so this post does not invent scores. DeepSeek's current pricing now varies by weekday UTC schedule.

Sources: pricing, thinking mode, V4 preview news, OpenRouter listing, and the HF preview repo.