Claude Opus 5 vs DeepSeek V4 Pro
Full profiles: Claude Opus 5 (Anthropic) and DeepSeek V4 Pro (DeepSeek): every benchmark, price and rank.
Updated · Data from 12 sources
As of 26 Oct 2026, Claude Opus 5 (max) leads on quality (68.8 vs 60.8 on the BenchLeader Index), but DeepSeek V4 Pro (max) costs 5.1× less, so it is the better buy unless you need the extra quality. Claude Opus 5 leads in every category we measure both in.
Claude Opus 5 benchmarks max Anthropic
DeepSeek V4 Pro benchmarks max DeepSeekOpen in the comparison toolWhich to pick for…
The pick's numbers come first. Category scores on the BenchLeader scale (50 is the average model, 15 points one standard deviation). High confidence means a lead of half a standard deviation or more; under one point is called either.
| Use case | Pick | Why | Confidence |
|---|---|---|---|
| Coding | Claude Opus 5 | 70 vs 58 category score; SciCode 56.4% vs 51.0% | High |
| Agents & tool use | Claude Opus 5 | 64 vs 60 category score; τ³-Banking (AA) 42.1% vs 39.6% | Medium |
| Maths | Claude Opus 5 | 71 vs 56 category score; FrontierMath Tiers 1–3 85.6% vs 64.6% | High |
| Reasoning | Claude Opus 5 | 74 vs 65 category score; GPQA Diamond 93.9% vs 91.7% | High |
| Knowledge | Claude Opus 5 | 72 vs 60 category score; SimpleQA Verified 59.9% vs 52.9% | High |
| Long documents | Either | 65 vs 65 category score; AA-LCR 79.3% vs 80.3% | Either |
| Chat quality (human preference) | – | Only Claude Opus 5 has an Arena rating (69 category score) | Not comparable |
| Lowest cost | DeepSeek V4 Pro | $1.98 vs $10.00 per 1M blended; agentic coding session $0.095 vs $0.400 | – |
| Fastest first answer | DeepSeek V4 Pro | 21.7 s (DeepSeek V4 Pro (max)) vs 68.7 s (Claude Opus 5 (max)) to the first answer token | – |
| High volume on a budget | Neither flagship | A smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs | – |
Key differences
- Claude Opus 5 (max) leads on quality: 68.8 vs 60.8.
- Claude Opus 5 (max) is stronger in agents & tools, coding, composite, human preference, knowledge, maths, multimodal, reasoning.
- DeepSeek V4 Pro (max) is stronger in long context.
- DeepSeek V4 Pro (max) is 5.1× cheaper ($1.98 vs $10.00 per 1M blended).
- DeepSeek V4 Pro (max) streams 2.0× faster (100 vs 51 tokens per second).
At a glance
| Spec | Claude Opus 5 | DeepSeek V4 Pro |
|---|---|---|
| Developer | Anthropic | DeepSeek |
| Released | 24 Jul 2026 | 13 Aug 2026 |
| Context window (tokens) | 1,000,000 | 1,048,576 |
| Compared at | Claude Opus 5 (max) | DeepSeek V4 Pro (max) |
| Open weights | No | Yes |
| Price in / out / cached input per 1M | $5.00 / $25.00 / $0.500 | $1.32 / $3.96 / $0.022 |
Category by category
Category scores on the BenchLeader scale: 50 is the average evaluated model, 15 points is one standard deviation. A gap under one point is called level.
| Category | Claude Opus 5 (max) | DeepSeek V4 Pro (max) | Leader |
|---|---|---|---|
| Reasoning | 74.1 | 65.1 | Claude Opus 5 |
| Coding | 70.2 | 58.3 | Claude Opus 5 |
| Agents & tools | 64.3 | 60.2 | Claude Opus 5 |
| Maths | 70.6 | 56.1 | Claude Opus 5 |
| Knowledge | 71.6 | 59.6 | Claude Opus 5 |
| Human preference | 68.5 | – | not comparable |
| Multimodal | 67.9 | – | not comparable |
| Long context | 64.7 | 65.2 | level |
| Composite | 86.3 | 69.8 | Claude Opus 5 |
Where sources disagree
Across the 29 benchmarks both have, Claude Opus 5 leads 25, DeepSeek V4 Pro leads 4. Every overall index that rates both puts Claude Opus 5 ahead.
| Overall index | Claude Opus 5 (max) | DeepSeek V4 Pro (max) | Leader |
|---|---|---|---|
| BenchLeader Index | 68.8 | 60.8 | Claude Opus 5 |
| AA Intelligence Index v4.3.2 | 50.8 | 36 | Claude Opus 5 |
| Arena Text | 1491 | 1465† | Claude Opus 5 |
Each publisher runs its own tasks, harnesses and effort settings, so their overall verdicts can differ; BenchLeader reconciles them on one scale. † measured at another effort setting of the same model (the publisher did not test the one compared here).
Price, speed and latency
List prices blended at three input tokens per output token. Per-task prices use the token counts on the methodology page; suite cost is what Artificial Analysis paid to run its whole Intelligence Index.
| Measure | Claude Opus 5 (max) | DeepSeek V4 Pro (max) |
|---|---|---|
| Blended price per 1M tokens | $10.00 | $1.98 |
| Input / output per 1M | $5.00 / $25.00 | $1.32 / $3.96 |
| Cached input per 1M | $0.500 | $0.022 |
| Price per task: chat reply | $0.0095 | $0.0017 |
| Price per task: summarize a 30-page report | $0.075 | $0.018 |
| Price per task: code edit | $0.068 | $0.014 |
| Price per task: agentic coding session | $0.400 | $0.095 |
| Price per task: structured extraction | $0.015 | $0.0034 |
| Cost to run a full benchmark suite | $5.86 | $0.674 |
| Output speed | 51 tok/s | 100 tok/s |
| Time to first answer | 68.7 s | 21.7 s |
| Time to a complete answer | 78.4 s | 26.7 s |
Per token is not per task: on Artificial Analysis' full suite Claude Opus 5 (max) cost $5.86 and DeepSeek V4 Pro (max) $0.674, because Claude Opus 5 (max) used about 1.7× as many tokens for the same work, which a per-token price hides.
Each model is compared at its best-ranked reasoning-effort setting. Claude Opus 5 (max) scores 68.8 and takes 69 s to its first answer token; Claude Opus 5 (high) scores 67.9 in 17 s, Claude Opus 5 (xhigh) scores 67.4 in 31 s. DeepSeek V4 Pro (max) scores 60.8 and takes 22 s to its first answer token; DeepSeek V4 Pro (high) scores 57.5 in 1 s, DeepSeek V4 Pro scores 53.4 in 1 s. Higher effort means more thinking before the answer: better scores, longer waits and more tokens billed. How reasoning effort changes scores and cost.
Every benchmark
141 benchmarks, grouped by category; each name links to the board and its leaders. † measured at another effort setting of the same model, shown for reference and not counted as a win.
| Metric | Publisher | Claude Opus 5 (max) | DeepSeek V4 Pro (max) |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 68.8 | 60.8 |
| Agents & tools score | BenchLeader | 64.3 | 60.2 |
| Coding score | BenchLeader | 70.2 | 58.3 |
| Composite score | BenchLeader | 86.3 | 69.8 |
| Human preference score | BenchLeader | 68.5 | – |
| Knowledge score | BenchLeader | 71.6 | 59.6 |
| Long context score | BenchLeader | 64.7 | 65.2 |
| Maths score | BenchLeader | 70.6 | 56.1 |
| Multimodal score | BenchLeader | 67.9 | – |
| Reasoning score | BenchLeader | 74.1 | 65.1 |
| Blended price $/M | BenchLeader | $10.00 | $1.98 |
| Output speed | BenchLeader | 51 tok/s | 100 tok/s |
| Time to first answer | BenchLeader | 68.7 s | 21.7 s |
| Context window | BenchLeader | 1,000,000 tokens | 1,048,576 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 93.9% | 91.7% |
| SimpleBench | SimpleBench | 80.6%† | – |
| Arena Hard Prompts | Arena (LMArena) | 1515 | 1480† |
| LiveBench 1.0 Reasoning | LiveBench | 91.2% | 85.8%† |
| GPQA Diamond (AA) | Artificial Analysis | 93.2% | 92.8% |
| Humanity's Last Exam (AA) | Artificial Analysis | 54.9% | 41.0% |
| GPQA Diamond (Vals) | Vals AI | 93.4%† | 92.4% |
| Kagi LLM Benchmark | Kagi LLM Benchmark | – | 53.5%† |
| ARC-AGI-1 | ARC Prize | 97.5% | 90.0% |
| ARC-AGI-2 | ARC Prize | 90.4% | 61.3% |
| ARC-AGI-3 | ARC Prize | 30.2%† | – |
| CritPt | Artificial Analysis | 29.1% | 18.0% |
| MysteryMechanism | Vals AI | 37.4%† | – |
| Chess Puzzles | Epoch AI Benchmarking Hub | 43.0% | 46.0% |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 59.0% | 43.0% |
| CL-bench Life | Epoch AI Benchmarking Hub | – | 13.5%† |
| LMCA | Epoch AI Benchmarking Hub | 63.3% | 41.2% |
| DTBench | Epoch AI Benchmarking Hub | 97.6% | 90.7% |
| ForecastBench | ForecastBench | 66.5% | 60.9%† |
| BTF-3 | Epoch AI Benchmarking Hub | 12.0%† | – |
| Coding | |||
| SWE-bench Verified (Epoch) | Epoch AI Benchmarking Hub | – | 77.6% |
| SciCode | SciCode | 56.4% | 51.0% |
| WeirdML | WeirdML | 91.8% | 66.2% |
| FrontierCode | Cognition | 53.4% | 28.6%† |
| Arena Coding | Arena (LMArena) | 1528 | 1501† |
Show all 141 benchmarks
| Metric | Publisher | Claude Opus 5 (max) | DeepSeek V4 Pro (max) |
|---|---|---|---|
| Coding | |||
| Arena WebDev | Arena (LMArena) | 1691 | 1447† |
| LiveBench 1.0 Coding | LiveBench | 81.5% | 77.2%† |
| LiveCodeBench | Vals AI | 89.0%† | 87.5% |
| IOI | Vals AI | 84.3%† | 51.6% |
| SWE-bench (Vals) | Vals AI | 97.0%† | 96.4% |
| Code Migration | Vals AI | 57.5%† | 41.5% |
| Code Migration v1.1 | Vals AI | – | 52.0% |
| ProgramBench | Vals AI | 3.0%† | 0.0% |
| Vibe Code Bench 1-100 | Vals AI | 28.5%† | 17.5% |
| Vibe Code Bench v1.1 | Vals AI | 88.4%† | 82.3% |
| SWE Atlas: Codebase QnA | Scale AI SEAL | 63.2%† | 27.1%† |
| SWE Atlas: Test Writing | Scale AI SEAL | 62.2%† | 27.1%† |
| Surface Evolver Bench | Epoch AI Benchmarking Hub | – | 40.0%† |
| DeepSWE v1.1 | Epoch AI Benchmarking Hub | 73.7% | – |
| CursorBench | Cursor | 46.6% | – |
| ALE-Bench | Epoch AI Benchmarking Hub | 2164.6† | 1403.2 |
| FrontierSWE | Epoch AI Benchmarking Hub | 52.0% | – |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 49.0% | 14.1% |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | 89.1% | 78.7% |
| Startrise Frontend | Startrise | 82.3%† | 62.7%† |
| FrontierCode 1.1 (AA) | Artificial Analysis | – | 38.8% |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | 51.8% | – |
| OSWorld-Verified 2.0 | OSWorld | 31.4% | – |
| APEX-Agents | Mercor | 65.8% | 47.3%† |
| Arena Agent | Arena (LMArena) | 8.2 | -0.3† |
| LiveBench Agentic Coding | LiveBench | 65.2% | 55.0%† |
| Terminal-Bench 2.1 (Vals) | Vals AI | 84.6%† | 54.7% |
| MCP Atlas | Scale AI SEAL | 85.8%† | – |
| HiL-Bench | Scale AI SEAL | 57.0%† | 38.5%† |
| τ³-bench Banking | τ²-bench | 48.7% | – |
| GDPval-AA v2.1 | Artificial Analysis | – | 47.5% |
| τ³-Banking (AA) | Artificial Analysis | 42.1% | 39.6% |
| Analyst Agent (AA) | Artificial Analysis | 53.8% | – |
| CUA-bench | Vals AI | 9.0% | – |
| CyberBench | Vals AI | 65.4%† | 72.9% |
| Finance Agent v2 | Vals AI | 58.6%† | 50.4% |
| Finance Agent v2.1 | Vals AI | – | 57.1% |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | – | 10.8% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | 6.7%† | 7.5% |
| Legal Research Bench | Vals AI | 55.3%† | 40.9% |
| SkillsBench | Vals AI | 60.4%† | 53.8% |
| SREBench | Vals AI | 12.2%† | 1.9% |
| Tax Agent Bench | Vals AI | 45.5%† | 26.8% |
| Terminal-Bench 2.0 (Vals) | Vals AI | – | 56.2%† |
| Terminal-Bench 4.0 (Vals) | Vals AI | 53.5%† | 14.1% |
| Terminal-Bench Science | Vals AI | 27.1%† | 4.3% |
| Time Horizon Index: KSP | Vals AI | 18.8%† | – |
| Arena Search | Arena (LMArena) | 1214† | – |
| BALROG | BALROG | 63.4% | – |
| PostTrainBench | Epoch AI Benchmarking Hub | 35.0%† | – |
| GBAEval | Epoch AI Benchmarking Hub | 79.6%† | – |
| Vending-Bench 2 | Andon Labs | 11181.9† | 3284.5† |
| AutomationBench | Artificial Analysis | 56.6% | 56.7% |
| GDP.pdf (AA) | Artificial Analysis | 21.6% | 11.4% |
| EnterpriseOps-Gym | Artificial Analysis | 47.5% | 49.6% |
| AA-Briefcase v1.1 | Artificial Analysis | 1657 | 1256 |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | 28.6% | 5.7% |
| DeepSec-Bench (AA) | Artificial Analysis | 25.9% | – |
| FACTS Search | 85.3%† | – | |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 85.6% | 64.6% |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 73.2% | 26.8% |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 98.9% | 98.6% |
| ProofBench | Vals AI | 99.0% | 16.0% |
| LiveBench Mathematics | LiveBench | 95.7% | 95.1%† |
| AIME 2026 | MathArena | – | 96.7% |
| HMMT February 2026 | MathArena | – | 93.9% |
| USAMO 2026 | MathArena | – | 60.7% |
| MathArena Apex | MathArena | – | 28.1% |
| Arena Maths | Arena (LMArena) | 1517 | 1444† |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 59.9% | 52.9% |
| LiveBench 1.0 Data Analysis | LiveBench | 74.5% | 79.2%† |
| AA-Omniscience | Artificial Analysis | 37.1 | 0.8 |
| MMLU-Pro | Vals AI | 91.6%† | 87.3% |
| LegalBench | Vals AI | 87.0%† | 82.4% |
| CorpFin | Vals AI | 73.2%† | 65.4% |
| TaxEval | Vals AI | 75.1%† | 73.1% |
| BioMysteryBench | Vals AI | 79.3%† | – |
| CaseLaw v2 | Vals AI | – | 59.4% |
| Excel Modeling Benchmark | Vals AI | 73.6%† | 52.8% |
| Excel Modeling Benchmark v1.1 | Vals AI | – | 65.1% |
| MedCode | Vals AI | 63.6%† | 42.5% |
| MedScribe | Vals AI | 91.0%† | 80.2% |
| Public Benefits Bench v1.1 | Vals AI | 76.9%† | 62.9% |
| DrugDiscoveryBench | Scale AI SEAL | 53.7% | 31.7%† |
| EBR-bench | Epoch AI Benchmarking Hub | 45.7% | – |
| GDP.pdf | Epoch AI Benchmarking Hub | 24.0% | – |
| AA-Omniscience: accuracy | Artificial Analysis | 60.9% | 49.1% |
| AA-Omniscience: non-hallucination | Artificial Analysis | 39.2% | 5.2% |
| Instruction following | |||
| LiveBench Language | LiveBench | 88.7% | 82.1%† |
| LiveBench Instruction Following | LiveBench | 63.8% | 67.7%† |
| Arena Instruction Following | Arena (LMArena) | 1495 | 1453† |
| Human preference | |||
| Arena Text | Arena (LMArena) | 1491 | 1458† |
| EQ-Bench 4 | EQ-Bench | 1360† | 1160† |
| Arena Creative Writing | Arena (LMArena) | 1474 | 1446† |
| Arena Multi-turn | Arena (LMArena) | 1487 | 1475† |
| Arena Longer Queries | Arena (LMArena) | 1503 | 1472† |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | 1320† | – |
| MMMU-Pro | Artificial Analysis | 84.7% | – |
| MMMU-Pro (Vals) | Vals AI | 89.9%† | – |
| MortgageTax | Vals AI | 72.1%† | – |
| SAGE | Vals AI | 49.4%† | – |
| Blueprint-Bench 2 | Epoch AI Benchmarking Hub | 30.4%† | – |
| Long context | |||
| AA-LCR | Artificial Analysis | 79.3% | 80.3% |
| Arena Document | Arena (LMArena) | 1490† | – |
| MLCR | Artificial Analysis | 11.1% | – |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 163.9† | 155.6† |
| LiveBench 1.0 | LiveBench | 80.1% | 77.4%† |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 50.8 | 36 |
| Vals Index | Vals AI | – | 53.0 |
| Vals Finance Index | Vals AI | – | 52.9% |
| Vals RSI Index | Vals AI | 33.0% | – |
| Vals Science Index | Vals AI | 47.9%† | – |
| Vals Multimodal Index | Vals AI | 73.9%† | – |
| Vals Router Index | Vals AI | – | 53.0%† |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | 16.8%† | – |
| Metadata | |||
| AA Openness Index | Artificial Analysis | – | 44.4 |
Data as of 2026-10-26. Best configuration of each model; every score links to its source on the model pages.
What changed
- Claude Opus 5 (max): 1 new result (Vals RSI Index).
- DeepSeek V4 Pro (max): 5 new results (Code Migration v1.1, Excel Modeling Benchmark v1.1, Finance Agent v2.1…).
- DeepSeek V4 Pro (max): BenchLeader Index 62.6 → 60.8.
- Claude Opus 5 (max): 1 score revision.
- Claude Opus 5 (max): 1 new result (ForecastBench); 1 score revision.
- DeepSeek V4 Pro (max): 1 new result (FrontierCode 1.1 (AA)); 2 score revisions.
- Claude Opus 5 (max): BenchLeader Index 68.3 → 68.9.
- DeepSeek V4 Pro (max): BenchLeader Index 61.0 → 62.6.
- Claude Opus 5 (max): 4 new results (GDPval-AA v2.1, AA-Omniscience: accuracy, AA-Omniscience: non-hallucination…).
- DeepSeek V4 Pro (max): 7 new results (USAMO 2026, IFBench, τ²-Bench Telecom (AA)…); 4 score revisions.
Related comparisons
- Claude Opus 5 vs Claude Opus 5.5
- Claude Opus 5 vs Gemini 4 Argon
- Claude Opus 5 vs Muse Wonder Full
- DeepSeek V4 Pro vs Claude Opus 5.5
- DeepSeek V4 Pro vs Gemini 4 Argon
- DeepSeek V4 Pro vs Muse Wonder Full
- DeepSeek V4 Pro vs Claude Fable 5.1
- Claude Fable 5 vs DeepSeek V4 Pro
Rankings: Best LLM overall · Best LLM for coding · Best LLM for agents · Best cheap LLMs
Claude Opus 5 vs DeepSeek V4 Pro: questions
- Is Claude Opus 5 better than DeepSeek V4 Pro?
- Claude Opus 5 (max) leads on quality: 68.8 vs 60.8. The BenchLeader Index combines every independent quality benchmark; Claude Opus 5 (max) is ahead overall as of 2026-10-26, but check the category scores for your use.
- Which is cheaper, Claude Opus 5 or DeepSeek V4 Pro, and will that change?
- DeepSeek V4 Pro is cheaper today: $1.98 against $10.00 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is Claude Opus 5 better than DeepSeek V4 Pro for coding?
- Claude Opus 5 scores higher in coding (70 vs 58 on the category index, where 50 is average). On SciCode, the top shared coding benchmark, it is 56.4% for Claude Opus 5 against 51.0% for DeepSeek V4 Pro.
- Is Claude Opus 5 better than DeepSeek V4 Pro for agents and tool use?
- Claude Opus 5 scores higher in agentic tasks (64 vs 60 on the category index, where 50 is average). On τ³-Banking (AA), the top shared agents & tools benchmark, it is 42.1% for Claude Opus 5 against 39.6% for DeepSeek V4 Pro.
- Which is better for long documents?
- They are level on long-context benchmarks (65 vs 65). Both take about 1M tokens: Claude Opus 5 1,000,000, DeepSeek V4 Pro 1,048,576.
- Which has lower latency?
- DeepSeek V4 Pro (max) answers first: 21.7 s against 68.7 s to the first answer token, as Artificial Analysis measures it. Reasoning effort drives most of this: a higher setting thinks longer before it answers.
- Which is faster, Claude Opus 5 or DeepSeek V4 Pro?
- DeepSeek V4 Pro streams faster: 100 against 51 output tokens per second.
- Do independent benchmarks agree on Claude Opus 5 vs DeepSeek V4 Pro?
- Across the 29 benchmarks both have, Claude Opus 5 leads 25, DeepSeek V4 Pro leads 4. Every overall index that rates both puts Claude Opus 5 ahead.
- When was DeepSeek V4 Pro released?
- DeepSeek V4 Pro was released on 13 Aug 2026.