Quick answer. DeepSeek-V4-Pro reached general availability on August 13, 2026 with the 0813 checkpoint, ending a preview that began in April. It ranks #2 on SWE-bench Verified at 96.40% — behind only Claude Opus 5 — at roughly 1/36th the cost per test. But it ranks last of seven frontier models on agentic coding, and prices rise on August 16.
DeepSeek-V4-Pro is now generally available. The April 24 launch was explicitly a preview; the 0813 checkpoint released on August 13, 2026 is the production release.
It is one of the more interesting models to evaluate right now, because it is simultaneously excellent and disappointing depending entirely on which benchmark you look at — and the split is not random.
What shipped in the GA release?
From DeepSeek's own release note:
- Agent upgrades with claimed production gains
- Reasoning-effort control on both V4-Pro and V4-Flash:
lowfor simple tasks,highfor daily agent workflows,maxfor complex work - Native OpenAI Responses API support, with one-click Codex setup
- Availability in the app and web UI via "Expert Mode"
- Model names unchanged —
deepseek-v4-pronow resolves to the 0813 checkpoint
Specifications
| V4-Pro (0813) | V4-Flash (0731) | |
|---|---|---|
| Context | 1,048,576 | 1,048,576 |
| Max output | 384K | 384K |
| Weights | MIT, 892.7 GB | MIT, 166.9 GB (284B / 13B active) |
| Precision | FP8 | FP8 |
| Modality | Text only — no vision | |
| Concurrency limit | 500 | 2,500 |
| Current price (in/out) | $0.435 / $0.87 | $0.14 / $0.28 |
The knowledge cutoff is not published anywhere in DeepSeek's documentation — a notable omission.
Architecturally, the 0813 config is essentially identical to the preview apart from four new DSpark keys and about 51.7B additional parameters — the speculative-decoding module we covered in our DSpark explainer. This is a serving-efficiency release more than a new base model.
How good is it, really?
Two answers, and both are true.
On patch-style coding, it is genuinely near the frontier
Vals AI runs all 82 systems on its board through one neutral bash-only harness. On SWE-bench Verified:
| Model | SWE-bench Verified | Cost per test |
|---|---|---|
| Claude Opus 5 | 97.00% | $1.29 |
| DeepSeek V4-Pro-0813 | 96.40% ±0.83 | $0.022 |
| GPT-5.6 Sol | 96.20% | — |
| Grok 4.6 | 95.60% | $0.785 |
| DeepSeek V4-Flash-0731 | 88.80% | $0.010 |
Second place, 0.6 points off the leader, at roughly 1/36th of Grok 4.6's cost per test and 1/59th of Claude Opus 5's.
What makes this result unusually credible: DeepSeek never claimed a SWE-bench number. This is purely third-party measurement on a neutral harness. There is no vendor chart to discount.
On agentic work, it is last in class
DeepSeek claims 87.9% on Terminal-Bench 2.1 for V4-Pro-0813, measured on its own "DeepSeek Harness minimal mode" — a harness that has not been released.
Under Terminus 2, the benchmark's own reference harness, Vals measures 54.68%. That is a 33-point gap.
Some context is essential here, because a 33-point drop sounds like fraud and mostly is not. Every lab loses ground moving to Terminus 2 — Kimi K3 by 7.4 points, Claude Fable 5 by 7.5, Claude Opus 4.8 by 13.1, GLM-5.2 by 13.2. Harness quality is a large, legitimate part of any agentic score.
But 33 points is not within that range. It is a categorical outlier, and it lands V4-Pro 12 points below its own cheaper sibling — V4-Flash scores 67.04% on the same neutral harness. A flagship losing to its own budget model is a strong signal that the claimed number reflects scaffolding rather than capability.
LiveBench corroborates independently: on agentic coding, DeepSeek ranks last of seven frontier peers — V4-Pro at 54.95 and V4-Flash at 46.77, against Claude Opus 5's 65.20 and Qwen3.8-Max's 64.65 — while sitting mid-pack on general benchmarks.
The fair reading
DeepSeek did something the other labs did not: it open-sourced its benchmark harness under MIT on release day. An independent public harness then reproduced V4-Flash's 82.7% Terminal-Bench figure exactly.
The caveat matters though — that reproduction ran with inflated task timeouts, nearly six hours on a task with a one-hour limit, which would be rejected by the official leaderboard. So the claim is real if your scaffolding is as generous as DeepSeek's, and DeepSeek has not shipped that scaffolding.
Neither model appears on the official verified Terminal-Bench board.
The cache-pricing advantage nobody talks about
This is the most important economic fact about DeepSeek and it is almost never mentioned.
V4-Pro's cache-read price is $0.003625 per million tokens — 0.83% of its input price. Every major Western lab prices cache reads at exactly 10% of input. DeepSeek prices them at roughly a twelfth of that ratio.
Why it matters: agentic workloads are dominated by cache reads. Every turn re-sends the accumulated conversation. On a realistic agentic request shape — around 750 uncached input tokens, 290 output, and 82,000 cache-read — V4-Pro costs about $0.00088 per request against Claude Opus 5's $0.052. That is 59x.
On stateless one-shot work the same comparison is only 16x. DeepSeek's advantage roughly quadruples the moment you put it in an agent loop — which is precisely why the August 16 price change, which raises cache reads 7.84x blended, hits harder than the headline rates suggest.
A trap worth knowing
Third-party V4-Flash endpoints on OpenRouter advertise cheaper input and output ($0.08/$0.18 versus DeepSeek's $0.14/$0.28). For agentic use they are about 3.4x more expensive — roughly $47 a month against $13.72 on the same workload — because their cache reads cost 5.7x more.
If you moved to a third-party host for data-residency reasons, that is a defensible trade. If you moved there to save money on agent workloads, you are paying more.
Prices rise on August 16
From 16:00 UTC on August 16, 2026, DeepSeek switches to peak/off-peak billing. Blended, that is roughly 1.96x on input, 2.94x on output and 7.84x on cache reads. Peak hours are 01:00–04:00 and 06:00–10:00 UTC.
Full rate tables and a worked example are in our DeepSeek price change guide. If you are running DeepSeek in production, read that before Sunday — after the change, GPT-5.6 Luna becomes cheaper than V4-Pro on agentic coding.
Can you run it locally?
V4-Pro: no. The weights are 892.7 GB. They are MIT-licensed and downloadable, but nobody is running this on their own hardware.
V4-Flash: yes, genuinely. At 166.9 GB with 284B total and 13B active parameters, community reports have it running at 50–70 tokens/second with the full 1M context on two DGX Sparks, or about 33 tokens/second on roughly $6,000 of used 3090s paired with a quad-Xeon board. One caveat from the same reports: time to first token stretches to about 42 seconds at 18K context.
Both are MIT-licensed, which remains one of the most permissive positions among frontier-adjacent models.
What about reliability?
The "DeepSeek's API falls over" reputation is out of date. Measured uptime on DeepSeek's own endpoints is 99.99%, while several third-party hosts sit between 88% and 92%. The most-cited outage complaint turned out to be an OpenRouter routing issue, corrected in its own thread.
The real constraint is published and rarely discussed: concurrency limits of 500 for V4-Pro and 2,500 for V4-Flash. Latency, not availability, is the live complaint.
Worth noting for evaluation purposes: there is no named enterprise deployment of DeepSeek V4 in the public record. Every production report we could find is an individual developer or a small team.
Should you use DeepSeek V4-Pro?
Yes, if your work is patch-shaped — bug fixes, targeted changes, PR-scale edits. Second place on SWE-bench Verified at 1/36th the cost of the nearest competitor is an exceptional offer, and it is independently measured.
Yes, if you run high-volume cache-heavy loops and can schedule around the peak windows. The cache-read pricing is structurally better than anything a Western lab offers, even after Sunday.
No, if you need long-horizon autonomous agents. Last of seven on LiveBench agentic coding, and a 33-point gap between claimed and neutral Terminal-Bench, both point the same direction.
No, if factual reliability is critical. AA-Omniscience scores V4-Pro-0813 at 0.83 against Claude Opus 5's 37.07 — near the floor. This is the least-discussed weakness and the one most likely to hurt in production.
Check first, if data residency matters. DeepSeek's own API is where the pricing advantage lives; moving to a third-party host for residency reasons costs you most of it.
FAQ
Is DeepSeek V4-Pro out of preview?
Yes. The April 24, 2026 release was a preview; general availability arrived with the 0813 checkpoint on August 13, 2026.
How much does DeepSeek V4-Pro cost?
Currently $0.435 per million input tokens, $0.87 output, and $0.003625 for cache reads. From 16:00 UTC on August 16, 2026 it moves to peak/off-peak billing at $0.66/$1.98 off-peak and $1.32/$3.96 at peak.
Is DeepSeek V4-Pro good at coding?
For patch-style work, yes — #2 on SWE-bench Verified at 96.40% on a neutral harness, behind only Claude Opus 5. For agentic coding, no: it ranks last of seven frontier peers on LiveBench agentic coding.
Why is DeepSeek's Terminal-Bench score disputed?
DeepSeek claims 87.9% using its own unreleased harness. Under the benchmark's reference harness, Vals measures 54.68% — a 33-point gap. All models lose ground on the reference harness, typically 7–13 points, so 33 is an outlier.
Can I run DeepSeek V4-Pro locally?
Not realistically — the weights are 892.7 GB. V4-Flash at 166.9 GB is viable, running around 33 tokens/second on roughly $6,000 of used consumer GPUs. Both are MIT-licensed.
Is DeepSeek's API reliable now?
Yes. Measured uptime on DeepSeek's own endpoints is 99.99%, better than several third-party hosts. The real limit is concurrency — 500 for V4-Pro, 2,500 for V4-Flash.
Should I use V4-Flash or V4-Pro?
Pro is stronger on SWE-bench (96.40% vs 88.80%) and, counterintuitively, cheaper per successful task on LiveBench because Flash generates far more tokens. Flash is the choice for local hosting or very high volume.
Are DeepSeek V4 weights open?
Yes, both V4-Pro and V4-Flash are released under the MIT licence — among the most permissive terms of any frontier-adjacent model.