DeepSeek V4 Pro 0813
Flagship GA
App, Web and API now point deepseek-v4-pro at the 0813 build. DeepSeek says production-agent gains are the headline — not another lab-only table.
Updated 2026-09-21
Highlights
HF model size
Total-parameter figures vary by source (HF and broker research disagree); ~49B activated. MIT weights.
QCode catalog price (per million tokens)
Follows the official peak/off-peak pricing: off-peak is roughly half of peak. Cache hits get an additional discount. Concurrency 500.
Terminal Bench 2.1
Official Minimal+max. DeepSWE 62.7 (Preview Pro was 12.8). HLE w/tools 60.0.
Context / max out
low / high / max effort. Responses API + Codex one-click script.
What 0813 changes
0813 supersedes V4-Pro Preview. Reuters covered the formal release as DeepSeek stepping up hiring, compute and fundraising. Simon Willison noted there was no pretty announcement page — this guide is the human changelog.
What people are arguing
HF's own table puts 0813 alongside Fable 5 (w/ fallback), Kimi K3 and Opus 4.8. TB2.1 is essentially tied with Fable (87.9 vs 88.0). The bigger story is pricing: on 08-13 the official peak/off-peak rules were announced, effective 08-17 — peak hours are 9:00-12:00 and 14:00-18:00 Beijing time, off-peak is half price, and peak output is +350% versus pre-change.
Timeline
V4-Pro Preview in API.
0813 GA on App, Web, API. Reuters same day.
Peak/off-peak. Peak out $3.96, off-peak $1.98.
Confirmed vs caution
Confirmed
Official bench table on HF; MIT repo; same model name; three effort levels; Codex Responses path.
Caution
Plenty of "V4 Pro 45 vs Sol 57" charts are still circulating; those describe the pre-0813 Preview build. Artificial Analysis scores the 0813 release at 53.
Pick Flash or Pro
Stay on Flash (V4.1)
Flash is now V4.1 (0731 retired 2026-09-10). Cheaper.
Escalate to 0813
Harder production agents, HLE-with-tools, when Flash stalls. 500 concurrency.
How to call it
model=deepseek-v4-pro on api.deepseek.com (OpenAI or Anthropic base). Local: HF 0813 + DSpark. OpenCode Go lists DeepSeek V4 Pro.
On QCode
QCode currently lists deepseek-v4-pro on /models. This page is a public changelog; the live list on /models is what determines availability.
0813 FAQ
Do I change the model name?
No. Keep deepseek-v4-pro. 0813 is the new default behind the name.
Open weights?
Yes, MIT on Hugging Face.
Better than Fable 5?
Not overall. TB2.1 is a tie; DeepSWE/HLE still favor Fable (w/ fallback). Pro wins on price and self-host.
Better than GPT-5.6 Sol?
They target different jobs. GPT-5.6 Sol is the closed reasoning flagship at 61 on the Artificial Analysis Intelligence Index; V4 Pro 0813 scores 53 there and wins on price and open weights. The "Pro 45" figure still circulating describes the pre-0813 Preview build.
Flash or Pro 0813?
Flash first. Pro when Flash fails the hard agent loop.
When do prices change?
2026-08-16 16:00 UTC, peak hours 01:00–04:00 and 06:00–10:00 UTC.
Sources
DeepSeek changelog 2026-08-13; pricing page; HF Pro-0813; Reuters 2026-08-13.
Route hard jobs to a cheaper flagship
Keep Claude / GPT where they still win; use 0813 when open weights and price matter.
Related
Flash 0731
The default cheap lane.
Pro vs Fable 5
Same-table benches.
DeepSeek Harness
Minimal mode = official scores.
Not affiliated with DeepSeek. Since 2026-08-17 pricing follows the official peak/off-peak scheme (off-peak is roughly half of peak); QCode catalog prices pass the official changes through, and the live /models list prevails.