DeepSeek V4 Pro 0813: Benchmark Results and Hands-On Test
DeepSeek V4 Pro 0813 benchmarked against GPT, Claude, and Gemini rivals, with official scores, independent coding tests, and pricing breakdown.
What is DeepSeek V4 Pro 0813?
DeepSeek V4 Pro 0813 is the full release of the V4 Pro model line, following a preview version shown back in April. DeepSeek pushed it live on their API with no launch event and minimal official documentation, which is typical of how the company ships models. Specs circulating in the community point to a 1.6 trillion parameter model with 49 billion active parameters and a 1 million token context window, though DeepSeek hasn’t confirmed these figures directly, so they should be treated as community-sourced rather than official.
TL;DR
- DeepSeek V4 Pro 0813 jumped from 72.1 to 87.9 on Terminal Bench 2.1 compared to its April preview, a 15.8 point gain that puts it near Kimi K3 (88.3) and Claude Opus-class performance.
- On official benchmarks it posted the top score on the table for Cyberjim (83.3) and Automation Bench (31.8), beating out rivals rather than just matching them.
- Independent testing on an eight-question coding and reasoning benchmark gave it 61 out of 80 (76.25%), tying with Muse Spark 1.2 and landing just below Kimi K3 and Opus 5.
- The model’s pricing sits around 43 cents per million input tokens and 87 cents per million output tokens, dramatically cheaper than premium rivals like Fable 5 at $10.50.
- Hands-on testing found the model overthinks simple problems and can be overeager, rewriting more code than a task requires, which sometimes makes its smaller sibling V4 Flash the better everyday choice.
- Front-end generation, task planning, and asking clarifying questions before running off with a wrong assumption stood out as genuine strengths.
- Compared to the V4 Pro preview, which scored just 24.8% on the same independent benchmark, the jump to 76.25% represents one of the larger generational leaps seen in a single model update.
How does DeepSeek V4 Pro 0813 perform on official benchmarks?
The official benchmark table circulating after release shows DeepSeek V4 Pro 0813 competing directly with frontier models rather than trailing them. On Terminal Bench 2.1, a benchmark focused on command-line and terminal-based task completion, it scored 87.9, essentially matching Fable 5 (88) and Kimi K3 (88.3).
It actually leads the pack on two benchmarks: Cyberjim, where it scored 83.3 (edging out Fable 5), and Automation Bench, where it scored 31.8, ahead of both Kimi K3 and Fable 5. On HLE (Humanity’s Last Exam), it scored 42.7 without tool use and 60 with tools, which trails Fable 5’s 53.3 and 63 but still holds up respectably. On DSBench-style evaluations and NL2 repo tasks, Opus 4.8 keeps a clear lead, and DeepSeek sits a few points behind Fable 5 and Opus in that category.
The overall pattern: DeepSeek V4 Pro 0813 isn’t uniformly the best model on every agentic benchmark, but it’s within a few points of the frontier almost everywhere, and outright wins on a couple of them. For a model priced the way this one is, that’s a notable result.
Is DeepSeek V4 Pro 0813 worth it on price?
This is where the model gets hard to ignore. At roughly 43 cents per million input tokens and 87 cents per million output tokens, DeepSeek V4 Pro 0813 costs a fraction of what comparable frontier models charge. Fable 5, one of its closest competitors on raw benchmark scores, runs at $10.50, which independent testers have pointed out works out to roughly 57 times the cost for similar performance on several tasks.
That price gap is why people, including members of the Klein team, have started calling this the best price-to-performance model currently available. It’s not automatically the best model overall, but for teams running high volumes of agentic or coding tasks where API costs stack up fast, the economics change the calculation significantly.
How did DeepSeek V4 Pro 0813 do on independent testing?
Official benchmarks are one thing, third-party testing tells a fuller story. An independent eight-question benchmark covering front-end coding, 3D rendering, SVG generation, math reasoning, long-horizon agentic tasks, and a notoriously difficult 3D wristwatch challenge, gave DeepSeek V4 Pro 0813 a final score of 61 out of 80, or 76.25%.
Breaking that down by task type:
- Elevator simulation (multi-elevator logic simulation): 6/10, core logic mostly worked but had gaps in edge cases.
- 3D contact lens case (click-to-open interaction): 8/10, tying top-tier models on this task.
- Folding table animation: 9/10, tied for the best score on this question alongside Fable 5, Kimi K3, GLM 5.2, and Sonnet 5.
- SVG panda illustration: 5/10, a clear weak spot, well behind models like Muse Spark 1.2, which scored a perfect 10.
- Bow and arrow game: 6/10, functional but not polished compared to top scorers at 10/10.
- Hard math permutation problem: 10/10, correct answer (2460), matching Fable 5, Opus 4.8, Kimi K3, and Qwen 3.8 Max.
- Long-horizon agentic task (autonomous dataset generation, model fine-tuning, and local web UI build): 10/10, completed fully without intervention.
- 3D wristwatch with real-time hands and dual time zones: 7/10, the highest score any model has achieved on this particular test, ahead of Fable 5’s earlier best of 4/10 and V4 Flash’s 6/10.
One coffee. One working app.
You bring the idea. Remy manages the project.
For context, the earlier V4 Pro preview scored just 24.8% on this same benchmark. Moving to 76.25% in one release cycle is a substantial jump, and it puts the model in a tie with Muse Spark 1.2, just above GLM 5.2 (75%) and its own smaller sibling V4 Flash (72.5%), and just below Kimi K3 and Opus 5 (77.5% each). Fable 5 still leads the pack at 82.5%.
What are the model’s real-world strengths and weaknesses?
Benchmark charts don’t capture everything. In hands-on use, a few patterns emerged that matter for anyone considering the model for daily work.
The model tends to overthink simple problems. Given a trivial task, it will sometimes reason far longer than necessary, and unlike its sibling V4 Flash, that extra reasoning doesn’t always help; it can actually lead to a worse final answer. Flash, by comparison, tends to solve straightforward problems directly without the same risk of second-guessing itself into a mistake.
It also shows a tendency toward overengineering. Given a task that needs a one-line fix, it may restructure entire files or add abstractions that weren’t requested, which means more code to review than the task warranted.
On the positive side, front-end generation is a clear strength, visually competitive with frontier-tier models, as reflected in its scores on the folding table and wristwatch tests. Task planning and breaking down larger jobs into sequenced steps was also called out as genuinely strong. One particularly useful behavior: when a prompt is ambiguous, the model tends to ask clarifying questions rather than guessing and burning tokens on the wrong approach, a habit that saves real back-and-forth time in practice.
Frequently Asked Questions
What is DeepSeek V4 Pro 0813’s context window?
Community-sourced specs put it at 1 million tokens, though DeepSeek has not officially confirmed this figure.
How does DeepSeek V4 Pro 0813 compare to Claude Opus?
On agentic benchmarks like NL2 repo and DSBench, Opus 4.8 maintains a clear lead. But on Terminal Bench 2.1, Cyberjim, and Automation Bench, DeepSeek V4 Pro 0813 performs competitively or better, at a much lower price point.
Is DeepSeek V4 Pro 0813 better than DeepSeek V4 Flash?
Not universally. For simple, everyday tasks, V4 Flash often performs better because it doesn’t overthink and rarely makes mistakes from excessive reasoning. V4 Pro pulls ahead on front-end work, planning, and larger agentic tasks.
What does DeepSeek V4 Pro 0813 cost to run?
Pricing sits around 43 cents per million input tokens and 87 cents per million output tokens, substantially cheaper than comparably performing models like Fable 5.
How much did DeepSeek V4 Pro improve over the April preview?
On an independent benchmark, the preview scored 24.8%, while the 0813 release scored 76.25%, a jump of over 50 percentage points on the same test suite.