DeepSeek V4 vs Claude Opus 4.6 - Open Weight Meets Proprietary

Claude Opus 4.6 is the best agentic model money can buy. It leads on Terminal-Bench 2.0, GDPval-AA, BrowseComp, and Humanity's Last Exam. It has agent teams, a 1M context window that actually retrieves accurately, and the highest Chatbot Arena Elo. It costs $5.00 per million input tokens and $25.00 per million output tokens.

DeepSeek V4 is expected to cost roughly $0.14 per million input tokens and $0.28 per million output tokens - if the leaked pricing estimates hold. That is 36x cheaper on input and 89x cheaper on output. The leaked SWE-bench Verified score of 80%+ would match Opus 4.6's 80.8%.

This comparison asks the most important question in AI right now: can an open-weight model match the proprietary frontier on the benchmarks that matter while costing two orders of magnitude less?

Note: DeepSeek V4 has not been officially released. All V4 specifications are based on leaked benchmarks and reporting from FT/Reuters/CNBC. This comparison will be updated with verified data after launch.

Quick Comparison

Claude Opus 4.6: The Agentic Benchmark

Opus 4.6 is the model other models are measured against on agentic work. GDPval-AA at 1,606 Elo puts it 144 points ahead of GPT-5.2 and 411 points ahead of Gemini 3.1 Pro on real-world office tasks. Terminal-Bench 2.0 at 65.4% leads for sustained coding execution. BrowseComp at 84.0% is the highest for web research. tau2-bench Retail at 91.9% shows near-perfect tool calling reliability. These aren't marginal leads - they represent decisive advantages on the kind of tasks enterprises pay for.

Agent teams are the headline capability. Opus 4.6 can spawn and coordinate parallel sub-agents through Claude Code, decomposing complex tasks into independently executed subtasks. Anthropic demonstrated this with a 100,000-line C compiler built by 16 parallel agents. This is infrastructure-level coordination that V4 has no known equivalent to.

The 1M context window is not just a number on a spec sheet. MRCR v2 at 76.0% shows Opus actually retrieves accurately from million-token contexts - versus Gemini 3.1 Pro's 26.3% on the same test. If your workload requires processing entire codebases or legal document sets in a single pass with reliable retrieval, Opus has the strongest verification data.

Where Opus falls short is on cost. At $5.00/$25.00 per million tokens, it is the most expensive frontier model. The long-context tier jumps to $10.00/$37.50 above 200K tokens. For high-volume production deployments, the bill adds up fast. The Batch API (50% discount) and prompt caching help, but Opus remains fundamentally a premium-priced product.

The other gap is modality. Opus handles text and images but not video or audio. V4 is described as natively processing all four modalities.

DeepSeek V4: The Economics Argument

V4's case rests on a simple proposition: if you can match the proprietary frontier on the benchmarks that drive production value, the pricing gap becomes indefensible.

The leaked SWE-bench Verified score of 80%+ is the critical data point. Opus scores 80.8%. If V4 genuinely hits 80%+, the coding gap between the cheapest and most expensive frontier models effectively disappears. HumanEval at ~90% versus Opus's 88% reinforces this picture - V4 may actually lead on code generation while costing 89x less on output.

The 1M context window matches Opus's advertised capacity. Whether V4 retrieves as accurately from million-token contexts as Opus's 76% MRCR v2 score is unknown - Engram Conditional Memory is the theoretical basis, but there's no independent benchmark data.

Native multimodal processing across text, image, video, and audio gives V4 a modality advantage over Opus's text-and-image limitation. For workloads involving video analysis or audio transcription integrated with text reasoning, V4 would eliminate the need for separate models.

The open-weight licensing is the strategic differentiator. If V4 ships under MIT or Apache 2.0, teams can self-host the model, customize it, fine-tune it, and deploy it without per-token API costs. For organizations processing billions of tokens per month, self-hosting V4 could reduce inference costs to the hardware level - a category shift in economics.

Benchmark Comparison

The pattern is the same as every V4 comparison right now: we have almost no leaked benchmark data beyond coding. On SWE-bench and HumanEval - the two benchmarks where V4 has leaked numbers - V4 is competitive or slightly ahead. On every other benchmark, Opus wins by default because V4 hasn't published data.

The critical unknowns are agentic performance (Terminal-Bench, BrowseComp, GDPval-AA, tau2-bench) and reasoning (GPQA Diamond, MRCR). These are Opus's strongest categories. If V4 matches Opus on coding but trails by 20+ points on agentic execution and enterprise tasks, the comparison becomes workload-specific rather than clear-cut.

Pricing Analysis

The numbers are stark. Processing 1 million output tokens costs $0.28 with V4 versus $25.00 with Opus - an 89x difference. For a team generating 100 million output tokens per month:

• DeepSeek V4: ~$28/month
• Claude Opus 4.6: ~$2,500/month
• Opus Batch API: ~$1,250/month

Annualized, the difference is $336 versus $30,000 (or $15,000 with batch). Even accounting for the Batch API's 50% discount, Opus remains 44x more expensive.

This gap only widens for teams that can self-host. Running V4 on owned GPU infrastructure eliminates per-token costs entirely, reducing the marginal cost to electricity and hardware amortization. For organizations processing billions of tokens per month, the self-hosting option is the difference between AI being a line item and AI being a capital investment. See our open source vs proprietary AI guide for the full framework.

DeepSeek V4: Pros and Cons

Pros (expected):

• 36-89x cheaper than Opus on per-token API costs
• Open weights enable self-hosting, fine-tuning, and unlimited customization
• Leaked SWE-bench 80%+ matches Opus on the most important coding benchmark
• HumanEval ~90% potentially exceeds Opus's 88%
• Native text, image, video, and audio versus Opus's text and image only
• MIT or Apache 2.0 licensing versus Opus's proprietary API-only model

Cons (expected/potential):

• Not yet released - all performance claims are unverified
• No leaked data on agentic tasks where Opus leads decisively
• No known agent teams or multi-agent coordination capability
• Huawei Ascend optimization may mean suboptimal Nvidia performance at launch
• Enterprise integration (PowerPoint, Excel) - Opus has native support, V4 unknown
• Tool calling reliability (Opus scores 91.9% on tau2-bench) is completely unknown for V4
• Trillion-parameter self-hosting requires massive GPU infrastructure

Claude Opus 4.6: Pros and Cons

Pros:

• Best-in-class agentic execution across every benchmark category
• Agent teams enable genuine multi-agent coordination through Claude Code
• 1M context with proven retrieval accuracy (76% MRCR v2)
• 128K output tokens - the longest output window at the frontier
• Native PowerPoint and Excel support for enterprise workflows
• Production-stable since February 5, 2026, with extensive documentation
• Strongest tool calling (91.9% tau2-bench) and enterprise task performance (1,606 GDPval-AA)

Cons:

• 36-89x more expensive per token than V4's estimated pricing
• Proprietary - no self-hosting, no fine-tuning, no open weights
• Text and image only - no video or audio input
• Parameters not disclosed - opaque architecture
• Long-context pricing ($10/$37.50 above 200K) adds significant cost for full-window usage
• Usage caps on subscription plans (Claude Max) can bottleneck heavy workloads

Verdict

Choose Claude Opus 4.6 if agentic reliability is your primary requirement. Opus leads every agentic benchmark by decisive margins, has proven tool calling at 91.9%, and agent teams enable multi-agent workflows that no open-weight model currently matches. If your workload involves sustained autonomous execution - code repair pipelines, research orchestration, enterprise document processing - Opus is the safer choice today and probably still will be after V4 launches. The cost premium is real, but for agentic work where failures are expensive, Opus's reliability advantage is worth paying for.

Choose DeepSeek V4 if you are building high-volume production systems where per-token cost dominates total cost of ownership. If V4's leaked coding benchmarks are accurate, you get 99% of Opus's SWE-bench performance at 1% of the cost. For batch processing, text analysis, code generation at scale, multimodal workloads with video and audio, or any deployment where you can tolerate slightly less agentic reliability in exchange for dramatically lower costs, V4 will likely be the rational choice. The open-weight license also enables self-hosting that eliminates per-token costs entirely.

The honest assessment: these models are not really competing for the same buyer. Opus is for teams where the quality of each individual output justifies premium pricing - enterprise knowledge work, complex agentic pipelines, mission-critical code repair. V4 is for teams where aggregate throughput and cost efficiency matter more than peak single-query performance. Most organizations will use both, routing queries by complexity and cost sensitivity. For a full breakdown of model routing strategies, see our guide on how to choose the right LLM in 2026.

Sources

• DeepSeek to release V4 next week - Financial Times (via Techmeme)
• Introducing Claude Opus 4.6 - Anthropic
• Claude Opus 4.6 System Card - Anthropic (PDF)
• Claude API Pricing - Anthropic
• DeepSeek API Pricing
• DeepSeek V4 Lite Leak - Awesome Agents