DeepSeek V4 Pro 0813 Build Explained: What Changed at GA

`deepseek-v4-pro-0813` is the checkpoint DeepSeek shipped on August 13, 2026 — the build that moved V4 Pro from preview to generally available. '0813' is the release date in YYMMDD-style shorthand, matching the Flash-0731 naming convention from July 31^[1][4].

The API keeps the stable alias `deepseek-v4-pro`, which now resolves to the 0813 weights. On OpenRouter the GA build is listed separately as `deepseek/deepseek-v4-pro-0813` (created August 13) alongside the older preview entry `deepseek/deepseek-v4-pro`, so if you use a third-party router, check which entry your config points at[10].

The July 31 Flash update established the pattern: same architecture, retrained weights, new checkpoint name. The 0813 Pro build follows the same logic — a materially improved checkpoint, not a config tweak.

• 0813 = GA checkpoint released Aug 13, 2026
• API alias deepseek-v4-pro resolves to 0813 weights
• OpenRouter lists it as deepseek/deepseek-v4-pro-0813
• Same architecture as preview: 1.6T total / 49B active MoE

The GA release bundled three headline changes plus a pricing overhaul^[1].

The reasoning-effort and Responses API changes are the ones developers actually feel day-to-day. `reasoning_effort` accepts `low` / `high` / `max` and defaults to `high`; the Responses API works with Codex clients through DeepSeek's one-click setup script^[4][5].

Pricing is the other big shift: peak/off-peak rates took effect 16:00 UTC on August 16, three days after GA. Every V4 Pro line costs more than the old flat rate, with output at $1.98 off-peak / $3.96 peak^[2].

The model card reports SWE-bench Verified at 96.4% — the headline coding number — alongside the full agent table: Terminal Bench 2.1 at 87.9, Cybergym 83.3 (first), AutomationBench 31.8 (first)^[4].

Note the SWE-bench number: the preview-era card reported 80.6% (as of August 1), and the 0813 card reports 96.4% — a 15.8-point jump that matches the pattern across the rest of the table^[4].

Independent testing (MindStudio, 8 tasks) scored the 0813 build at 76.25%, versus 24.8% for the preview — same direction, more conservative magnitude^[6]. The full benchmark page has the complete tables and the harness caveat.

The 0813 model card on Hugging Face documents the weights, the encoding folder, and the official vLLM deployment recipe — including the DSpark speculative-decoding flag^[4].

Self-hosting 0813 is a serious lift — third-party estimates put the FP4+FP8 weights at roughly 862GB of VRAM. The hosted API is the practical route for most teams; the model size page has the full hardware math.

• No Jinja chat template on the card — encoding is provided as a Python folder demonstrating OpenAI-compatible format encoding
• vLLM deployment: enable DSpark with a single flag: --speculative-config '{"method":"dspark",...}'
• Reference hardware: 4xGB300 single node for the full model
• Weights are MIT-licensed; self-hosting is officially supported

If you are on the preview, the upgrade is automatic — the stable model ID already points at 0813. The real decision is budget and workload^[2][6].

Upgrade if your work is agentic or coding-heavy: the GA build's agent gains are the largest single-version jump in the family's history, and the Responses API unlocks Codex workflows that preview-era Pro could not do^[1][6].

Hold the line if your workload is high-volume and latency-sensitive chat: independent tests note 0813 over-thinks simple problems, and for everyday tasks V4 Flash often feels snappier at one-third the price^[6]. The peak-hour rates make that calculus sharper — see the pricing page.