DeepSeek V4 Pro for Coding Agents: What to Know
Published August 13, 2026 · by the AQ team
DeepSeek V4 Pro is the flagship of DeepSeek's V4 family: an open-weights mixture-of-experts model (about 1.6T total parameters, 49B active per token) with a 1M-token context window and a thinking mode that streams its reasoning before it answers. It reached general availability on August 13, 2026 as the "0813" build, after an April preview, and the weights are on Hugging Face under the MIT license. Two things make it the most interesting open model of the moment for coding-agent work: the 0813 update posted enormous gains on agentic benchmarks (Terminal Bench 2.1 up from 72.1 to 87.9 on DeepSeek's own card), and three days after GA, on August 16, DeepSeek's first-party API prices rise steeply, which reshuffles where it makes sense to run it.
What the 0813 release changed
The base architecture is unchanged from the April preview; DeepSeek attributes the gains to post-training, plus a new speculative-decoding module (DSpark) that inference stacks like vLLM and SGLang can enable with a flag. The published deltas are concentrated exactly where agent users live. On DeepSeek's own 0813 model card: Terminal Bench 2.1 rose from 72.1 to 87.9, DeepSWE from 12.8 to 62.7, CyberGym from 52.7 to 83.3, and Humanity's Last Exam with tools reached 60.0. The April preview card already claimed the highest open-weights SWE-bench Verified score at 80.6 and a 93.5 on LiveCodeBench, both in max-effort mode.
Read those numbers with the right skepticism. They are DeepSeek's own, several benchmarks (the DSBench pair) are internal with no public leaderboard, and the GA-week coverage noted that independent verification is still catching up. The model card's own comparison columns are also honest enough to show it losing rows: Kimi K3 edges it on Terminal Bench 2.1 (88.3 vs 87.9), and Claude Opus 4.8 leads on NL2Repo (69.7 vs 61.5). Independently, Artificial Analysis moved its Intelligence Index for V4 Pro from 45 to 53 after 0813: second among open-weights models it tracks, tied with GLM 5.2, against 63 for Claude Opus 5. One widely shared framing from GA week: Claude Fable 5 averaged about 5% better across the nine agentic benchmarks at roughly 46 times the blended per-token price. Whether that trade is good depends entirely on your work.
It thinks out loud, and your harness has to handle that
V4 Pro is a reasoning model. Through the API, chain-of-thought arrives in a dedicated reasoning_content field alongside content, with thinking on by default at "high" effort (the levels are low, high, and max; the GA build added "low"). Two mechanical consequences matter for agent workflows. First, in tool-calling loops the reasoning_content from prior turns must be passed back in continuation requests or the API rejects the call. Second, sampling parameters like temperature and top_p have no effect while thinking mode is on. Self-hosted checkpoints emit <think> tags instead of the API field.
This is why the loudest theme in the GA-week Hacker News discussion was harness sensitivity: the same model swings ten or more points across different agent harnesses, because how a harness manages the reasoning stream, the tool loop, and the cache changes what the model can actually do. The practical advice follows directly: evaluate V4 Pro inside the harness you would actually use, not from a leaderboard row somebody else produced with a different one.
Pricing, and the August 16 increase
Until August 16, 2026, DeepSeek's first-party API charges $0.435 per million input tokens (cache miss) and $0.87 per million output for V4 Pro, rates that made it almost free by frontier standards. From August 16 at 16:00 UTC, first-party pricing moves to peak and off-peak windows: $1.32 input / $3.96 output at peak, half that off-peak, with the cache-hit discount also shrinking substantially. Off-peak hours align with nights in China, which for many US teams conveniently overlaps the workday; even so, some line items rise several-fold, and DeepSeek's stated reason is to "allocate resources more reasonably."
Because the weights are MIT-licensed, first-party is not the only meter. US-hosted serverless providers list V4 Pro at around $1.74 per million input and $3.48 per million output as of August 13, 2026. After the increase, first-party peak pricing and US-hosted pricing nearly converge, which quietly removes the cost penalty that used to come with keeping traffic under a US provider's terms instead of DeepSeek's own API. If data jurisdiction matters to your team, the sovereignty choice and the economical choice are now roughly the same choice. Self-hosting under MIT remains the full-control option, with community GGUF and FP4 conversions already published, though serving a 1.6T-parameter MoE well is its own project.
Where to run it
As of August 13, 2026: DeepSeek's own API speaks the OpenAI ChatCompletions format, the Anthropic API format, and now the OpenAI Responses API, and DeepSeek publishes official integration guides for coding agents (including a documented OpenCode setup and a one-click Codex configuration). The open weights run on US serverless hosts and on your own hardware. And inside AQ, V4 Pro is one of the models on AQ Models, covered below, so a team can try it without creating any provider account at all.
When to pick it, when not to
Pick V4 Pro when the work is agentic and terminal-shaped (its strongest published results and its strongest word of mouth: executing against a detailed plan, long tool-use runs), when cost per unit of work dominates (it is dramatically cheaper than closed frontier models even at post-increase rates), when you need the 1M context, or when open weights are a requirement: audit, fine-tuning, or jurisdiction. Early users consistently rank it at or near the top of the open-weights class for coding.
Look elsewhere when you need vision (V4 Pro is text-only, the most common complaint at GA), when unguided planning quality matters more than execution (GA-week reports have closed frontier models ahead there), or when your task is latency-critical interactive editing and the thinking stream gets in the way. And if most of your work fits a smaller model, DeepSeek's own V4 Flash at a fraction of the price was many early users' actual pick; several argued the Pro premium only pays off on the hardest tasks.
Where AQ fits
AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. DeepSeek V4 Pro is available in AQ through AQ Models: pick it from the model menu when launching OpenCode or Pi in a workspace and it runs with no DeepSeek account, no provider key, and no separate billing setup. That inference consumes AQ credits, priced at provider cost during early access, with $5 in credits included when you sign up, and AQ adds no markup on your model usage.
Because harness and model pair so tightly for this model, the honest way to evaluate it is the side-by-side: open two workspaces on the same repository (each is an isolated git worktree), give V4 Pro and your incumbent the same backlog item, and let the team watch both live sessions from the browser. Our guide to reviewing an AI coding session covers what to compare beyond the final diff. Sessions run in persistent tmux on your team's VM, so a long V4 Pro run keeps executing after your laptop closes.
Frequently asked questions
Is DeepSeek V4 Pro open source, and what is the license?
The weights are open under the MIT license, published on Hugging Face (deepseek-ai/DeepSeek-V4-Pro, with the GA build at DeepSeek-V4-Pro-0813). MIT means self-hosting, fine-tuning, and commercial use without licensing fees. The training data and pipeline are not open, so it is open-weights rather than open-source in the strict sense.
What does DeepSeek V4 Pro cost?
Three meters, as of August 13, 2026. DeepSeek's first-party API: $0.435 per million input tokens and $0.87 per million output until August 16, then peak rates of $1.32 and $3.96 (half that off-peak). US-hosted serverless providers: around $1.74 input and $3.48 output. In AQ, it runs on AQ Models with credits priced at provider cost during early access, with $5 included at sign-up.
How good is DeepSeek V4 Pro at coding compared to Claude or GPT?
On DeepSeek's published numbers it is the strongest open-weights coding model as of August 2026: 80.6 on SWE-bench Verified (preview card) and 87.9 on Terminal Bench 2.1 (0813 card). Independent measurement puts it second among open models on Artificial Analysis's index, still well behind Claude Opus 5. GA-week comparisons had Claude Fable 5 modestly ahead on agentic benchmarks at a much higher price. Its reputation: excellent at executing detailed plans cheaply, weaker at unguided planning, no vision.
What is thinking mode, and does it affect tool use?
V4 Pro reasons before answering and returns that reasoning in a dedicated reasoning_content field (self-hosted checkpoints use think tags). Thinking is on by default at high effort, with low and max also available. It affects tool use directly: in multi-turn tool loops the prior reasoning_content must be sent back with continuation requests, and temperature and top_p are ignored while thinking is on. Agent harnesses that handle this well get visibly better results, which is why benchmark scores vary so much by harness.
How do I try DeepSeek V4 Pro with my team without setting up API keys?
In AQ, launch OpenCode or Pi in a workspace and pick DeepSeek V4 Pro from the model menu: it runs on AQ Models with no DeepSeek account or provider key, metered as AQ credits priced at provider cost during early access ($5 included at sign-up). Teammates can open the same workspace and watch the session live, which makes side-by-side comparisons against your current default a same-day exercise.