Wait Which Model?
← Back to directory
DeepSeek

DeepSeek-V4-Pro-0813

Released Aug 13, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
China
Modality
Text
Context window
1M
Max output
384K
Speed (max effort)
78 tok/s · 26s to first answer token
Price ($/MTok in / out)
$1.32 / $3.96
Cost per task
—
Open weights
Yes
Benchmarks6 of 10 reported

GPQA Diamond

93%

SWE-bench Verified

96.4%

Terminal-Bench 2.1

87.9%

Humanity's Last Exam

42.7%

LMArena Elo

1463

ARC-AGI-2

61.3%

MMLU-Proretired

—

SWE-bench Pro

—

AIME (latest)retired

—

GDPval-AA v2

—
Strengths & weaknesses

Strengths

  • Same 1.6T/49B MoE structure as April's preview with a DSpark speculative-decoding module bolted on — the agentic gains arrive without a new architecture to re-integrate
  • Cyber-defense and terminal work are where the re-post-training shows most; independent evaluators single out CTF-style tasks as its clearest jump over the preview
  • Thinking and non-thinking modes with three effort levels behind one endpoint, so a quick lookup and a long agent run share the same model id
  • MIT-licensed weights mirrored on Fireworks, DeepInfra and OpenRouter within days — no dependence on DeepSeek's capacity-constrained first-party API

Weaknesses

  • Headline agent scores come from DeepSeek's own unreleased minimal-mode harness — on the Terminal-Bench reference harness it lands below its cheaper Flash sibling, and nobody has reproduced the card's numbers
  • Multi-turn thinking mode requires replaying the previous turn's reasoning verbatim — the most common cause of 400 errors when migrating existing agent loops
  • Landed as a stealth update with no launch post (the brief notice on DeepSeek's site was pulled within a day), so documentation is thin and behavior is still being characterised

The production release of DeepSeek-V4-Pro, superseding April's preview per DeepSeek's own model card: same 1.6T/49B MoE with a DSpark speculative-decoding module, all gains from re-post-training. Pricing is the peak-hour rate that took effect 2026-08-16 (peak = 01:00–04:00 and 06:00–10:00 UTC); off-peak is 50% lower at $0.66/$1.98 and cache-hit input is cheaper still — roughly a 3–4x rise from the preview's $0.435/$0.87. Terminal-Bench 2.1 87.9 and HLE 42.7% (60.0% with tools) are DeepSeek's own figures at max effort in its minimal-mode harness; independent runs on the Terminus 2 reference harness report 54.68 (V4-Flash-0731: 67.04). GPQA Diamond 93 is Artificial Analysis' measurement as displayed on its charts, not lab-reported — DeepSeek's own 0813 card figure could not be confirmed directly (one unofficial site cites 92.8; the 90.1 many aggregators list is April's preview). SWE-bench Verified 96.4 is third-party (Vals AI, #2 on its board behind Claude Opus 5) — DeepSeek published no SWE-bench figure and the 80.6 some trackers carry is the preview's. Also reported by DeepSeek: DeepSWE 62.7, CyberGym 83.3, NL2Repo 61.5, Toolathlon 74.1, Agents' Last Exam 25.7, AutomationBench 31.8. Artificial Analysis scores it 53 on its Intelligence Index — 8 points above April's preview but only 1 above V4-Flash-0731 on the current index version — measuring 78.1 tok/s and 25.67 s to first answer token on DeepSeek's endpoint at max effort; its per-task cost figure could not be retrieved. MMLU-Pro and AIME are unpublished; the arena.ai listing and ARC Prize result arrived after launch (below). Max output is the 384K DeepSeek recommends for high/max effort. The API went live 2026-08-12 under the unchanged `deepseek-v4-pro` id; the Hugging Face card and DeepSeek's release notice followed on 2026-08-13. Updated 2026-09-15: DeepSeek's V4.1-Flash launch post (2026-09-10) said all `deepseek-v4-pro` requests would route to V4.1-Flash from 2026-09-14 04:00 UTC until V4.1-Pro ships, but the API change log and pricing page then reversed this "in response to user demand" — V4 Pro service continues after 2026-09-14 with billing unchanged, so no retirement is recorded. ARC-AGI-2 61.3% is ARC Prize-verified at max effort ($0.60/task; high 59.7%, low 56.3%), published 2026-08-13 at arcprize.org/results/deepseek-v4-pro-0813. LMArena Elo 1463 is the "deepseek-v4-pro-high-20260813" text-arena listing (arena.ai, 2026-09-13 update); the undated "deepseek-v4-pro" listing sits at 1457.

For developers

API model strings

  • deepseekdeepseek-v4-pro

Licence

MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

Replaces DeepSeek-V4

News