DeepSeek-V4-Pro reaches general availability with major agent-benchmark gains

DeepSeek has taken its V4-Pro-0813 model out of preview and into general availability across the app, web, and API, according to the company's official API changelog. The release keeps the same calling convention as before — developers set the model parameter to deepseek-v4-pro to get the new build — but ships a substantial jump in agentic and coding benchmark scores over the preview version.

What's new

The changelog lists a full benchmark table for the GA build, including:

• Terminal Bench 2.1: 87.9
• DeepSWE: 62.7
• CyberGym: 83.3
• NL2Repo: 61.5
• Toolathlon-Verified: 74.1
• HLE (without tools / with tools): 42.7 / 60.0
• Agents' Last Exam: 25.7
• AutomationBench (Public): 31.8
• DSBench-FullStack / DSBench-Hard: 71.1 / 67.2

On the technical side, DeepSeek says the platform "natively supports the OpenAI Responses API format and is specifically adapted for Codex," with one-click configuration scripts provided for developers migrating over. Both V4-Pro and V4-Flash now expose three thinking-effort levels — low, high, and max — so callers can trade latency for reasoning depth on a per-task basis.

The GA release also lands alongside a pricing change: DeepSeek introduced peak/off-peak rates for V4-Pro, with off-peak pricing set at half the peak-hour rate, effective August 16, 2026 at 16:00 UTC.

Context

V4-Pro has been in preview status for several weeks, during which independent developers ran their own comparisons against the checkpoint. The GA changelog frames this release as focused specifically on production agent workloads — the featured benchmarks are almost entirely agentic and tool-use evaluations (terminal use, repository navigation, cybersecurity tasks, full-stack development) rather than general knowledge tests, signaling where DeepSeek is positioning the model competitively.

Why it matters

The jump in scores on agent-centric benchmarks like Terminal Bench and DeepSWE suggests DeepSeek is chasing the same agentic-coding workloads that OpenAI's Codex and Anthropic's Claude Code line have made a central battleground in 2026. Native Responses API compatibility is a pointed move: it lowers the switching cost for developers already building on OpenAI's newer API surface, letting them point existing Codex-style tooling at DeepSeek's models with minimal rework. Combined with DeepSeek's typically aggressive pricing, a GA release with these benchmark claims — even if third-party verification is still pending — keeps pressure on rivals in the price-to-performance segment of the agentic coding market.