By ByteBot
DeepSeek’s V4 Pro left preview on August 13 with improved agent benchmarks, native OpenAI Responses API support, and a price increase that could hit 11x for some users when new tiers kick in on August 16. The 1.6-trillion-parameter model is now generally available across the DeepSeek app, web, and API under an MIT license on Hugging Face. The numbers are impressive. The caveat is that every agent benchmark cited at launch is vendor-reported, with no independent replications on record yet.
What Changed in the GA Build
The 0813 designation marks the shift from preview to general availability — and the diff is meaningful. DeepSeek reworked the model around agentic workloads: tool use, code execution, and multi-step workflows that run without human intervention. Three reasoning effort levels (low, high, max) are now exposed via API, matching the pattern set by Claude and OpenAI for exposing chain-of-thought depth. The default is high.
The context window is 1,048,576 tokens (1M) with up to 384,000 output tokens — the largest available output ceiling of any widely-accessible model right now. Native support for the OpenAI Responses API (not just Chat Completions) means teams using Codex can switch with a one-click configuration script from DeepSeek’s docs. DSpark, DeepSeek’s speculative decoding technique, handles inference efficiency on the backend.
The Benchmark Numbers — and the Asterisk
DeepSeek reports large gains over its April 2026 preview. On Terminal Bench 2.1, the score jumped from 72.1 to 87.9. DeepSWE went from 12.8 to 62.7. DSBench-FullStack moved from 41.8 to 71.1 and DSBench-Hard from 31.1 to 67.2.
Those are big numbers. Here is the problem: DeepSeek has not released the benchmark harness it used for any of the agentic tests. At GA time, independent benchmark trackers showed zero third-party evaluations for V4 Pro 0813. The open weights (MIT license) mean any team with enough GPUs can run their own evaluation — and that’s genuinely useful — but nobody has published results yet. Take the vendor numbers as hypotheses to test, not settled rankings.
The comparison to frontier alternatives looks good on paper: V4 Pro claims roughly 85% on SWE-bench Verified versus Claude Fable 5’s 95%, at approximately 1/57 the output token cost. V4 Pro Max tops LiveCodeBench at 93.5%, while Claude Opus 4.8 still holds an edge on repo-scale SWE-bench Pro. For a breakdown, BenchLM has a current comparison.
The Price Increase
Anyone running V4 Pro on the API in production needs to act before August 16. Before that date, pricing is $0.435 per million input tokens and $0.87 per million output tokens. Starting August 16, DeepSeek introduces peak/off-peak tiers:
- Off-peak: $0.66 input / $1.98 output per million tokens
- Peak: $1.32 input / $3.96 output per million tokens
Peak output pricing is 4.5x higher than today’s rate. Versus the earliest preview prices, some routes are up to 11x more expensive. InfoWorld reports the increase is driven by demand straining DeepSeek’s infrastructure capacity. The off-peak option cuts the bill in half — teams running batch inference or async pipelines should restructure scheduling now.
The relative cost position still holds: off-peak V4 Pro output is roughly 1/25 the cost of Claude Fable 5 ($50/M) and far cheaper than most Western frontier models. For high-volume workloads that can tolerate scheduling flexibility, the math remains compelling even after the hike.
Migrating to V4 Pro
If you’re already on the OpenAI SDK, migration is two changes: swap the base URL and API key. Everything else runs unchanged.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_DEEPSEEK_KEY",
base_url="https://api.deepseek.com"
)
response = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[{"role": "user", "content": "Your prompt here"}],
extra_body={
"thinking": {"type": "enabled"},
"reasoning_effort": "high" # low | high | max
}
)A few things to know before you flip the switch: Max reasoning mode produces long chains of thought — set your context window to at least 384K and be generous with max_tokens. Thinking mode disables temperature, top_p, presence_penalty, and frequency_penalty. If your production code sets those, you’ll need to handle that before migrating. For Codex specifically, the GA release notes include a one-click configuration script.
When to Use V4 Pro
V4 Pro makes most sense for three scenarios: long-context jobs pushing past 128K tokens where you want deep reasoning on large codebases or document sets; high-volume agentic pipelines where cost is the primary constraint and you can tolerate off-peak scheduling; and research or fine-tuning use cases where the MIT license matters.
For everyday coding tasks, V4 Flash is faster and cheaper. For repo-scale software engineering where Anthropic’s reliability and track record matter, Opus 4.8 still holds an edge. V4 Pro is the option for teams who want frontier-adjacent capability at a fraction of the price and can live within the peak/off-peak structure.
The benchmark claims are worth watching — once independent evaluations land, the picture will sharpen. For now, the open weights mean you can run your own tests on your actual workloads. That tells you more than any leaderboard number.
