elvis
I have been testing DeepSeek-V4-Pro with the Pi coding agent. I am mindblown by how well it works out of the box.
Task Built an LLM wiki with an agent powered entirely by DeepSeek-V4-Pro.
anecdotalhigh-signal user
DeepSeekDeepSeek
DeepSeek V4 Pro
DeepSeek · DeepSeek. Released Apr 24, 2026.
DeepSeek V4 Pro is a model from DeepSeek in the DeepSeek family, released Apr 24, 2026. evals.report tracks 25 reported DeepSeek V4 Pro benchmark scores across DeepSWE, LMArena, LiveBench, Humanity's Last Exam, WeirdML, SWE-bench Verified, GPQA Diamond, SWE-bench Pro, and 16 more — each shown with its benchmark, metric, source status, and date, and never combined into a single ranking.
Open25 results
Benchmark results 25
Compare this model| Benchmark | Category | Score | Metric | Status | Date | |
|---|---|---|---|---|---|---|
| DeepSWE | Coding | 7.52% | % resolved | Official | Apr 24, 2026 | Details |
| LMArena | Chat preference | 1446 | source-defined rating | Official | Apr 24, 2026 | Details |
| LiveBench | Reasoning | 73.58% | score | Official | Apr 24, 2026 | Details |
| Humanity's Last Exam | Reasoning | 32.4% | accuracy | Official | Apr 24, 2026 | Details |
| WeirdML | Coding | 48.9% | average accuracy | Official | Apr 24, 2026 | Details |
| SWE-bench Verified | Coding | 80.6% | % resolved | Unverified | Apr 24, 2026 | Details |
| GPQA Diamond | Reasoning | 90.1% | accuracy | Unverified | Apr 24, 2026 | Details |
| SWE-bench Pro | Coding | 55.4% | % resolved | Unverified | Apr 24, 2026 | Details |
| LiveCodeBench Pro | Coding | 3206 | Codeforces Elo | Unverified | Apr 24, 2026 | Details |
| Artificial Analysis Intelligence Index | Reasoning | 51.5 | Index | Unverified | Apr 24, 2026 | Details |
| τ²-bench (Telecom) | Tool use | 96.2% | pass^1 | Official | Apr 24, 2026 | Details |
| AIME 2026 | Reasoning | 95.83% | accuracy | Official | Apr 24, 2026 | Details |
| GDPval | Agents | 1558 | Elo | Official | Apr 24, 2026 | Details |
| SciCode | Coding | 50.0% | accuracy | Unverified | Apr 24, 2026 | Details |
| AA-Omniscience: Knowledge and Hallucination Benchmark | Reasoning | -10 | AA-Omniscience Index | Official | Apr 24, 2026 | Details |
| IFBench | Reasoning | 76.5% | accuracy | Official | Apr 24, 2026 | Details |
| WebDev Arena | Chat preference | 1464 | Elo | Verified | Apr 24, 2026 | Details |
| EQ-Bench Creative Writing v3 | Chat preference | 1570 | Elo | Verified | Apr 24, 2026 | Details |
| Design Arena | Chat preference | 1302 | Elo | Verified | Apr 24, 2026 | Details |
| MathArena HMMT February 2026 | Reasoning | 93.94% | accuracy | Official | Apr 24, 2026 | Details |
| Vectara Hallucination Leaderboard | Other | 8.6% | Hallucination Rate | Official | Apr 24, 2026 | Details |
| Vibe Code Bench | Coding | 49.93% | Overall accuracy | Verified | Apr 24, 2026 | Details |
| Terminal-Bench 2.0 | Agents | 67.9% | task success | Verified | Apr 24, 2026 | Details |
| FrontierSWE | Agents | 29% | dominance score | Official | Apr 24, 2026 | Details |
| DeepSWE | Coding | 5.3% | % resolved | Unverified | Apr 24, 2026 | Details |
In the wild 1
See allReal-world feedback on DeepSeek V4 Pro from people using it on actual prompts — praise and criticism alike, each linked to its source. Qualitative, never scored.
Report tone
Report type
Topic
Positive