Head-to-head
DeepSeek-V4 vs Claude Opus 5: which AI model wins in 2026?
DeepSeek-V4 ($0.87/1M out) and Claude Opus 5 ($25/1M out) are two of the most-used AI models in 2026. Across 7 community votes, Claude Opus 5 leads with 63% approval.
Quick verdict
On Reasoning, pick Claude Opus 5: the arena rates it 5/5 against 4.5/5 for DeepSeek-V4. On budget, DeepSeek-V4 wins: it starts at $0.87/1M out versus $25/1M out for Claude Opus 5.
Line-by-line comparison
Strengths and weaknesses
DeepSeek-V4
- 1M-token context window (8x the 128K of V3.2) with up to 384K output tokens, standard on the official API
- Aggressive pricing: $0.435/$0.87 per 1M tokens (V4-Pro), roughly 28.7x cheaper per output token than Claude Opus 4.8; cache-hit input drops to $0.003625/1M (over 99% discount)
- MIT-licensed open weights for both V4-Pro and V4-Flash on Hugging Face: commercial use, fine-tuning and redistribution allowed
- Open-source SOTA on agentic coding: 80.6 on SWE-bench Verified (Think Max config), tied with Gemini 3.1 Pro, plus Codeforces rating 3206 (~rank 23 vs humans)
- Ranks #3 of 93 on the Artificial Analysis Intelligence Index (score 44), well above the 25 average
- Sparse-attention stack cuts 1M-context inference to 27% of V3.2's FLOPs and 10% of its KV cache
- Intermittent malformed tool calls: function calls sometimes emitted as plain text in content instead of the tool_calls field (GitHub issue deepseek-ai #1244)
- Thinking mode breaks long multi-turn tool-call chains with 400 errors in agent frameworks (OpenClaw issue #72044, fix still incomplete)
- Developers report it fabricating nonexistent APIs in custom codebases and acting on hallucinated user input in agent loops
- Very verbose (180M eval output tokens vs 95M median) and mid-pack speed at 54.6 tok/s (#39/93), which erodes the low per-token price in practice
- Text only (no vision or audio) and still a preview: the official release planned for mid-July 2026 adds peak-hour pricing that doubles listed API rates during Beijing business hours
Claude Opus 5
- Ranked #1 in composite intelligence across 190 models on Artificial Analysis at launch, and 43.3% on Frontier-Bench v0.1 vs 34.4% for GPT-5.6 Sol and 33.7% for Claude Fable 5
- 3.9x better than GPT-5.6 Sol on ARC-AGI-3 novel reasoning (30.2% vs 7.8%), and Elo 1861 on GDPval-AA v2 economic knowledge work, ahead of Fable 5 (1747)
- 79.2% on SWE-bench Pro, within a point of Fable 5 (80.0%) and 10 points above Opus 4.8 (69.2%), at half Fable's price; Cursor's co-founder calls it 'near Fable 5 intelligence at Opus speed and cost'
- Same $5/$25 pricing as Opus 4.8 with a bigger window: 1M context is now the default and only tier, with 128K max output and prompt caching from 512 tokens
- Dual-use safety classifiers trigger 85% less often than on Fable 5, and the new default fallback mode avoids the silent mid-session refusals that plagued Fable's launch
- Self-verifies its work without being told, handles mid-conversation tool changes (beta) without busting the prompt cache, and ships day one on Claude.ai, the API, Bedrock, Vertex and Microsoft Foundry
- Notably slow and very verbose: 52.6 output tokens/s and 68 seconds to first token on Artificial Analysis, and it consumed ~100M output tokens during their eval vs a 63M median
- The verbosity is a real-world cost problem: CodeRabbit measured it reading ~50% more and writing ~65% more than reference frontier models per code-review call
- No actual price cut despite the 'cost-efficient' narrative: identical to Opus 4.8 ($5/$25), nearly GPT-5.6 money ($5/$30), and more than 2x comparable Gemini or Grok tiers
- Reviewers describe a 'brilliant but annoying' personality: over-verification, hedging, and occasional refusals of mundane tasks like resolving a merge conflict (Lenny's Newsletter field review)
- Breaking API change for Opus 4.8 migrants: thinking is on by default and cannot be disabled at xhigh or max effort (returns a 400); Fast mode ($10/$50, ~2.5x faster) is API-only, not on Bedrock or Vertex
Cast your verdict
One recommendation per tool per gladiator. It reshapes the crowd score everyone sees.
The arena’s verdict on DeepSeek-V4
Choose DeepSeek-V4 if you want near-frontier reasoning and agentic coding at 3x to nearly 30x below Claude Opus or GPT-5.5 pricing, or if MIT-licensed weights for self-hosting and fine-tuning matter to you. It is a decisive upgrade over V3.2: 8x longer context, far cheaper long-context inference and stronger coding, and the legacy deepseek-chat/reasoner endpoints are deprecated on July 24, 2026 anyway. Avoid it for production agents that depend on rock-solid multi-turn tool calling, where users still report malformed tool calls and fabricated APIs, and for any vision or audio work since it is text only. Latency-sensitive apps should also test first, as its verbosity and mid-pack 54.6 tok/s output speed offset some of the cost advantage, and budget for the peak-hour price doubling arriving with the official mid-July release.
The arena’s verdict on Claude Opus 5
Claude Opus 5 is the sane default of the Series 5 range: most of Fable 5's intelligence (and more than Fable on Frontier-Bench and GDPval) at exactly half the token price, with classifiers that trigger 85% less often. If you migrated workloads to Fable 5 for capability but resent the bill or the false-positive refusals, move them here; if you are still on Opus 4.8, the upgrade is 10 SWE-bench Pro points for free. The two honest reasons to look elsewhere: latency and verbosity. At 52.6 tokens/s with 68s to first token it is a poor fit for interactive UX, and its token appetite quietly inflates real costs beyond the sticker price, so budget-sensitive high-volume pipelines still belong on Sonnet 5, Gemini or DeepSeek. Keep Fable 5 only for the longest autonomous runs where its slight SWE-bench Pro edge compounds.
What the crowd says
On DeepSeek-V4
“Tool calling is flaky. Function calls sometimes land as plain text instead of the tool_calls field, and thinking mode 400s on long multi-turn chains. Not agent-ready yet.”
“MIT license on both Pro and Flash weights is the real story. Fine-tune, redistribute, ship commercially, no lawyer needed. Plus 384K output tokens for long-doc generation.”
“MIT weights, 1M context, and output tokens roughly 29x cheaper than Opus 4.8. Cache hits make input basically free. Moved my bulk pipelines over and the bill collapsed.”
On Claude Opus 5
“The sticker price is unchanged but my invoice is not: it writes essays where Opus 4.8 wrote answers. 68 seconds to first token killed it for our support chat, we went back to Sonnet 5.”
“Fed it a 40-tab financial model with cross-sheet formulas and asked for a scenario deck. It got the edge cases the analysts missed. For document-heavy enterprise work this is the best model I have used.”
“We moved off Fable 5 because the bio classifier kept flagging our genomics tooling. Opus 5 does the same work at half the price and I have not seen a single silent reroute since.”
“Migrated our agents from Opus 4.8 the day it dropped. Same bill, and tasks that used to stall at the planning stage now just finish. The 10-point SWE-bench jump is not marketing, our merge queue feels it.”
Keep comparing
Frequently asked questions
Is DeepSeek-V4 better than Claude Opus 5?
The crowd currently sides with Claude Opus 5: 63% recommend it, versus 57% for DeepSeek-V4 (7 votes). On Reasoning, Claude Opus 5 rates higher (5/5 vs 4.5/5). The right pick depends on your use case. The line-by-line comparison on this page breaks down pricing, key specs and arena ratings.
Which is cheaper, DeepSeek-V4 or Claude Opus 5?
DeepSeek-V4 is cheaper: it starts at $0.87/1M out, while Claude Opus 5 starts at $25/1M out.
How much do DeepSeek-V4 and Claude Opus 5 cost per 1M tokens?
DeepSeek-V4: $0.435/1M in (cache hit $0.003625) per 1M input tokens, $0.87/1M out per 1M output tokens. Claude Opus 5: $5/1M in per 1M input tokens, $25/1M out per 1M output tokens.