DeepSeek V4 Pro 0813: Frontier Agentic Coding at Commodity Price

On August 12, 2026, DeepSeek shipped the GA release of V4 Pro under the same model ID and the same price as the preview: DeepSeek-V4-Pro-0813. If the preview was a prototype, this is the production build — and it changes the default model choice for most developers.

The release in one line

Same API call, same price, dramatically better model. DeepSeek's own documentation notes the calling method is unchanged, so any integration already using deepseek-v4-pro receives the new build with zero code changes. OpenRouter's model page confirms it is the GA release, live the same day, and OpenCode Go announced availability within hours.

What the GA numbers show

DeepSeek's published GA benchmark chart (which the community flagged as initially leaked through its WeChat group, and as of this writing remains unverified by independent evaluators) shows a large jump over the preview build:

Every lane improved; the largest gains are in long-horizon agentic benchmarks — exactly the workloads where the preview had been weak.

How it stacks against the closed frontier

On Terminal-Bench 2.1 — the agentic terminal benchmark most relevant to real coding loops — Artificial Analysis independently scores GPT-5.6 Sol at 89.5 and Claude Opus 5 at 89.1. DeepSeek's reported 87.9 for 0813 sits within 1.6 points of the top, ahead of Claude Opus 4.8's 85.0. On CyberGym, 83.3 beats Opus 4.8's 78.3.

The tradeoffs are real: on pure knowledge reasoning without tools (HLE), the GA build trails Claude Opus 5 by roughly twelve points, and the model is text-only with no vision input. The GA figures are DeepSeek-reported and not yet independently verified — an X analysis echoed the community's "seismic if accurate" caution.

The price argument

This is where the comparison stops being close. At $0.435 per 1M input and $0.87 per 1M output tokens, DeepSeek V4 Pro 0813 costs roughly 1/12th of Claude Opus 5 ($5/$25) and about 1/30th of GPT-5.6 Sol ($5/$30) for a model within a few points of both on agentic coding.

Against the open-weight field it is similarly disruptive: Kimi K3 ($3/$15) and GLM-5.2 ($1.4/$4.4) both trail it on several agentic lanes at 3-7x the price. The model carries a 1M-token context window with 384K max output, supports reasoning-effort control, and keeps the cache-hit price near zero.

Flash as the subagent, Pro as the main

The same release cycle produced V4 Flash 0731 — a 284B-parameter MoE with 13B active at $0.14/$0.28. For high-volume loop work, Flash is cheap enough to retry without thinking, and its own GA chart numbers improved substantially over its preview. The practical configuration is two-tier: Pro (0813) as the long-horizon thinker — it beats Flash on every published benchmark, with the largest margins in the agentic lanes — and Flash as the high-volume worker that keeps per-loop cost near zero.

The caveats worth keeping

The GA numbers are DeepSeek-reported, initially surfaced through a WeChat group leak, and not yet verified by Artificial Analysis or other independent evaluators. The model remains text-only. For knowledge-heavy reasoning without tools, the closed frontier still leads. And DeepSeek's pricing page now warns of a planned price increase, so the current rates may not hold indefinitely.

For agentic coding — the workload that defines most of modern development — the value equation is unambiguous. Frontier-adjacent performance at a commodity price, with the same model ID that upgrades silently. That combination is the reason the developer-tool ecosystem adopted it within hours, and it is a strong candidate for the default model on most machines.

Benchmark figures for DeepSeek V4 Pro 0813 are as reported by DeepSeek and, as of publication, have not been independently verified by third-party evaluators. Frontier-model scores for GPT-5.6 Sol, Claude Opus 5, and pricing comparisons are from Artificial Analysis.