deepseek-v4-pro
Meant to excel at
-
Ten out of ten on a long-horizon agentic task (dataset generation, fine-tuning, web UI) with no intervention, and good front-end generation and clarifying questions, in one 8-question test.
Practitioner report mindstudio.ai -
Met every rule on a strict multi-constraint memo where V4 Flash missed arithmetic and length limits.
Practitioner report glbgpt.com
Watch out for
-
Overthinks simple problems and rewrites more code than a task needs; Flash may be the better everyday pick.
Practitioner report mindstudio.ai -
Used its whole 2,048-token budget on reasoning and returned no visible answer on a strict task until the limit was raised.
Practitioner report glbgpt.com -
Retired in practice: since 2026-09-14 deepseek-v4-pro requests are routed to V4.1 Flash and billed at Flash rates until a V4.1 Pro ships.
Vendor claim, relayed thenextweb.com
Benchmarks
Agentic coding percentile 0.54
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| swe-bench pro | 55 | 0.25 | Independent evaluator morphllm.com | 2026-07-24 |
| swe-bench verified | 80.6 | 0.70 | Vendor claim, relayed morphllm.com | 2026-07-24 |
| terminal-bench 2.1 | 87.9 | 0.55 | Vendor claim, relayed anotherwrapper.com | 2026-10-19 |
| terminal-bench 2.1 (aggregator, anotherwrapper.com) | 87.9 | 0.67 | Independent evaluator anotherwrapper.com | 2026-10-19 |
Overall percentile 0.13
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 36 | 0.13 | Independent evaluator artificialanalysis.ai | 2026-10-19 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +
DeepSeek V4 Pro 0813 was the fastest in a four-model HTML test, about 100 to 114 tokens per second.
Compared with glm-5.3 | 2026-10-04
Practitioner report empiriolabs.ai - +/-
A vendor-run 8-question test scored V4 Pro 61/80: perfect on a long-horizon dataset, fine-tune and web UI task, good on front-end, weak on an SVG drawing, and prone to overengineering. The tester is unnamed, scores are unreplicated, and MindStudio promotes its own product.
Task: agentic-long-running | 2026-08-13
Practitioner report mindstudio.ai - +/-
On a 3-file agentic Python pricing bug, Pro and Flash took the same six tool calls and both fixed it (8/8 checks), but Flash finished in about 47s vs Pro's 95s. Single run per model at temperature 0, via a reseller gateway.
Task: bug-fix | 2026-08-13
Practitioner report glbgpt.com - +
On a strict launch memo with word limits, exact table, arithmetic and JSON, Pro passed 12/12 checks while Flash got 8/12 (miscalculated cost, summary under the minimum length). Both first exhausted a 2,048-token cap on reasoning; single run, published by a reseller.
Task: writing-content | 2026-08-13
Practitioner report glbgpt.com