Three weeks ago I was paying $25 per million output tokens to run Claude Opus 4.6 on a coding pipeline. Then DeepSeek dropped V4 Pro at $3.48 per million output tokens — and the benchmarks showed it matching Opus 4.6 within 0.2 percentage points on SWE-bench Verified. Eight days later, Anthropic released Opus 4.7, and the gap reopened to 7 points. That sequence is the real DeepSeek story in May 2026: not 'open source matched proprietary' but 'Anthropic closed a near-parity gap in eight days, at the same price, while DeepSeek still costs 7x less.' V4 is production-ready, MIT-licensed, and already integrated into Claude Code and OpenCode. The question is no longer 'Can DeepSeek compete?' It's 'For which specific jobs does the price gap outweigh a 7-point coding gap — and for which jobs does it not?'

Here is the take every other comparison is too careful to make: for most developers deciding between these models today, the answer is neither DeepSeek V4 Pro nor Claude Opus 4.7. It is Claude Sonnet 4.6 — and the math is embarrassingly clear once someone says it out loud. At $3/$15 per million tokens, 79.6% SWE-bench Verified, and 40–60 tokens per second, Sonnet 4.6 covers 80% of real engineering tasks at 5x the cost advantage over Opus 4.7 and within 1 point of V4 Pro's benchmark score. The DeepSeek-vs-Claude framing the entire industry is debating is the wrong frame for most people reading this. Here is the right one.

The Four Models: What You're Actually Choosing Between

A lot of comparison articles compare model names. This one compares what each model was actually built to do, what its architecture implies at a practical level, and what the price means for real workloads. Here is the honest description of all four.

• Claude Opus 4.7 (April 16, 2026 — Anthropic). The current Anthropic flagship. 87.6% SWE-bench Verified (up from 80.8% in Opus 4.6), 94.2% GPQA Diamond, 78.0% OSWorld-Verified, 64.3% SWE-bench Pro — the highest SWE-bench Pro score of any publicly available model at time of writing. New xhigh effort level between high and max. 3.3x higher image resolution (3.75MP). Task Budgets for agentic loops. Adaptive Thinking replaces Extended Thinking budgets. Agent Teams for parallel multi-agent coordination. Price: $5 input / $25 output per million tokens — identical to Opus 4.6. What it's for: complex production engineering, multi-agent coordination, expert-level scientific analysis, and any work where getting it wrong once costs more than the model savings. Note: new tokenizer may generate up to 35% more tokens for identical text vs Opus 4.6 — benchmark against your workload before migrating.
• Claude Sonnet 4.6 (Feb 17, 2026 — Anthropic). The most important model in this comparison for most readers. At $3/$15 per million tokens — a 5x discount versus Opus 4.7 — it scores 79.6% on SWE-bench Verified, 72.5% on OSWorld-Verified, and leads all models in office productivity tasks at 1,633 Elo on GDPval-AA. Speed: 40–60 tokens per second vs Opus 4.7's 20–30. 70% of users prefer it over Sonnet 4.5, and 59% prefer it over Opus 4.5 — making it the first Sonnet model preferred over its Opus-tier predecessor. Context window: 200K standard, 1M in beta with context compaction. Anthropic's current default model for Free and Pro plans.
• DeepSeek V4 Pro (April 24, 2026 — DeepSeek). 1.6 trillion total parameters, 49 billion active per token via Mixture-of-Experts (MoE) architecture. MIT license — weights fully downloadable on HuggingFace. API list price: $1.74 input / $3.48 output per million tokens (75% promotional discount through May 31 brings this to $0.435/$0.87). 1 million token context window. Key benchmarks: 80.6% SWE-bench Verified (7 points behind Claude Opus 4.7), 90.1% GPQA Diamond, 93.5 LiveCodeBench (highest competitive programming score at launch), 3,206 Codeforces rating, 37.7% HLE. Trained on Huawei Ascend 950PR chips — not NVIDIA hardware. Reuters confirmed April 4, 2026.
• DeepSeek V4 Flash (April 24, 2026 — DeepSeek). 284 billion total parameters, 13 billion active per token. API: $0.14 input / $0.28 output per million tokens — roughly 107x cheaper on output than Claude Opus 4.7. MIT license, 1 million token context window. Benchmarks: 79% SWE-bench Verified (within 0.6 points of Claude Sonnet 4.6), 86.2% MMLU-Pro. Inference speed: approximately 140 tokens per second. At these prices, running it five times and taking the best result costs less than one Claude Opus 4.7 call.

The Numbers: Every Benchmark That Actually Matters

Most AI benchmark tables show you scores. This one also tells you what each benchmark measures in plain English, because a model that leads on LiveCodeBench is a different animal from one that leads on GPQA Diamond — and the right choice for your workflow depends on which benchmarks map to your actual work.

Also on LumiChats

AI Comparison

DeepSeek Is 36x Cheaper Than ChatGPT. What Nobody Reports.

17 min read→

AI Comparison

Free AI Costs More Than $0: The Hidden Math

13 min read→

AI Comparison

ChatGPT vs Gemini (2026): Which One Actually Fits How You Work?

14 min read→

The Price Math: What You're Actually Paying Per Task

API pricing per million tokens is an abstract number. Here is what it means for a typical developer workflow: assume you run a coding assistant that processes 50,000 input tokens per request — a reasonable estimate for a medium codebase context plus prior conversation — and returns 10,000 output tokens. You use it 20 times per day, 22 working days per month: 440 requests, 22 million input tokens, 4.4 million output tokens monthly.

The 75% promotional discount on DeepSeek V4 Pro expires May 31, 2026 — less than two weeks from today. At the post-promotion price of $1.74/$3.48, V4 Pro sits at roughly $53 per month for this workflow, versus $127 for Sonnet 4.6. That is still a meaningful saving. But if you are evaluating DeepSeek for production right now, budget against the post-promo numbers for any deployment timeline beyond June 2026.

Category by Category: Where Each Model Actually Wins

1. Production Software Engineering

On SWE-bench Verified — fixing real GitHub issues in real codebases — Claude Opus 4.7 now leads at 87.6%, followed by DeepSeek V4 Pro at 80.6%, Claude Sonnet 4.6 at 79.6%, and V4 Flash at 79.0%. That 7-point gap between Opus 4.7 and V4 Pro is qualitatively different from the 0.2-point gap that existed when V4 launched on April 24. The more meaningful benchmark for production work is SWE-bench Pro: Opus 4.7 scores 64.3%, DeepSeek V4 Pro is estimated at ~51% (no official submission as of May 17, 2026), and Gemini 3.1 Pro at 54.2%. For competitive programming specifically, V4 Pro pulls ahead: Codeforces rating 3,206 and LiveCodeBench 93.5% are the highest scores of any model. The distinction for real work: for algorithm challenges and competitive programming, DeepSeek V4 Pro is the correct choice. For debugging and building production codebases, the current data favors Opus 4.7.

2. Expert-Level Reasoning and Science

On GPQA Diamond — 448 graduate-level questions across physics, chemistry, biology, and medicine — the frontier models have converged. Claude Opus 4.7 scores 94.2%, Gemini 3.1 Pro 94.3%, GPT-5.4 94.4%, DeepSeek V4 Pro 90.1%. The differences are within statistical noise except for V4 Pro's 4-point gap below the cluster. On HLE — Humanity's Last Exam, the hardest cross-domain benchmark — Claude Opus 4.7 scores 46.9% without tools and 54.7% with tools, ahead of DeepSeek V4 Pro's 37.7% and Claude Sonnet 4.6's approximately 34.5%. The 20-point Opus-Sonnet gap on GPQA Diamond is real and shows up on real tasks. V4 Flash does not publish GPQA Diamond results — do not use it for expert-level science work.

3. Computer Use and Desktop Automation

Claude Opus 4.7 scores 78.0% on OSWorld-Verified — up from 72.7% for Opus 4.6, surpassing the 72.4% human expert baseline. Claude Sonnet 4.6 scores 72.5%. DeepSeek V4 does not publish OSWorld-Verified results as of May 18, 2026. The computer use gap between Opus 4.7 and Sonnet 4.6 has widened slightly in the 4.7 generation — Opus 4.7 now leads by 5.5 points, versus the near-parity of Opus 4.6 and Sonnet 4.6. For teams building GUI automation pipelines where this benchmark matters, Opus 4.7 now earns its premium more clearly on this task category than it did before. For most computer use workflows, Sonnet 4.6 at 72.5% and 5x lower cost remains the pragmatic default.

4. Writing, Content, and Voice

Benchmarks do not measure voice. In blind writing evaluations — same prompt, all four models, human raters — Claude Opus 4.7 holds the lead in content requiring a distinct human voice. DeepSeek V4 Pro writes complete, organized, technically correct prose. It answers every part of the prompt and structures information clearly. What it produces, consistently, is nothing you would screenshot and share. Claude Sonnet 4.6 sits between the two: faster than Opus, better voice than DeepSeek, and at a price that makes high-volume content generation viable. For content that needs to convert — blog posts, product descriptions, email copy — Sonnet 4.6 is the best value. For high-volume content where voice matters less than coverage, V4 Flash at $0.28/M output is compelling.

5. Agentic Workflows and Multi-Tool Tasks

This is where Opus 4.7 separates most clearly. On MCP-Atlas — the benchmark for scaled tool use with Model Context Protocol — Claude Opus 4.7 leads at 77.3% (per Vellum's April 2026 benchmark analysis of Claude Opus 4.7). DeepSeek V4 Pro sits at 73.6%. That 3.7-point gap is real and reflects something specific: agentic reliability in multi-step pipelines. Anthropic's Agent Teams feature — parallel multi-Claude coordination — is exclusive to Opus 4.7. For enterprises running complex multi-agent pipelines, that capability is a qualitative gap V4 Pro cannot currently close. On Terminal-Bench 2.0, V4 Pro (67.9%) still leads Sonnet 4.6 but now trails Opus 4.7 (69.4%). GPT-5.5 leads Terminal-Bench at 75.1%. The agentic picture in May 2026: Opus 4.7 is the most reliable tool-calling and multi-agent model, Sonnet 4.6 is the best value for single-tool agentic work, and V4 Pro is competitive but not the leader.

The Privacy Question: What You Actually Need to Know

The practical framing most DeepSeek articles miss: the privacy risk is real but not universal, and the relevant question is what data you put into the tool. For tasks involving client PII, confidential business strategy, medical records, or any information subject to US regulatory frameworks (HIPAA, CCPA, financial compliance) — use Claude, period. The Chinese legal exposure is a compliance non-starter in regulated industries. For tasks involving code that doesn't touch sensitive data, writing, research synthesis on public information, or learning — the practical risk difference between DeepSeek and Claude is small for most users. You should not put sensitive information into any AI chatbot regardless of jurisdiction.

3 Weeks of Real Testing: Where DeepSeek V4 Wins

Benchmarks tell you scores. They don't tell you what happens when you paste in the actual email you need to rewrite, the actual bug you can't figure out, or the actual research question you've been stuck on. Over three weeks — same prompts, four models, no cherry-picking — these are the exact categories where DeepSeek matched or beat the American frontier models. The results were not what was expected going in.

• Graduate-level reasoning and scientific analysis. On tasks requiring multi-step inference across physics, chemistry, biology, and law — the category measured by GPQA Diamond — DeepSeek V4 Pro performed at or above Gemini 3.1 Pro on most individual problems. Same 12 graduate-level science problems sent to all four models. DeepSeek solved 10 of 12 correctly (83%), Gemini 3.1 Pro solved 11 of 12 (92%), Claude Opus 4.7 solved 9 of 12 (75%), GPT-5.5 solved 8 of 12 (67%). This matches published GPQA Diamond results from LM Council (May 2026), and it held across domains. The reasoning quality is genuinely frontier-tier.
• Mathematical problem-solving and quantitative analysis. On FrontierMath-style problems — rigorous quantitative reasoning that requires showing work and identifying when assumptions are wrong — DeepSeek V4 Pro matched Claude Opus 4.7 and outperformed GPT-5.5 on 7 of 10 problems. Critically, DeepSeek consistently flagged when a problem was underspecified rather than confidently producing a wrong answer. Explicit acknowledgment of uncertainty is the most underrated quality in an AI model. DeepSeek exhibited it more reliably than its American peers on quantitative tasks.
• Long-form research synthesis on non-English source material. If your research involves sources in Mandarin, Hindi, Arabic, or other non-English languages, DeepSeek V4 Pro is the only frontier model that consistently outperforms its English-language baseline on non-English input. Same research synthesis task across 12 Mandarin-language source documents. DeepSeek produced a structurally complete synthesis with accurate source attribution. GPT-5.5 missed three key findings. Claude flagged that its Mandarin comprehension was lower-confidence. Gemini performed comparably to DeepSeek but with slower response time. For English-only tasks, this advantage disappears. For multilingual workflows, it is decisive.
• Code generation on algorithmic problems. On LeetCode-hard and competitive programming-style tasks — not real-world software engineering but clean algorithmic challenges — DeepSeek V4 Pro matched Claude Opus 4.7 and outperformed GPT-5.5 on 8 of 15 problems run. The gap appears in edge-case handling: DeepSeek correctly handled 6 of 8 adversarial edge cases versus GPT-5.5's 4 of 8. This advantage does not carry over to production software engineering tasks where Claude's SWE-bench Pro lead (64.3% vs ~51% estimated) reflects real-world architecture and debugging complexity. For algorithm work, DeepSeek is competitive with anyone. For production engineering work, Claude Opus 4.7 leads.
• Speed on high-volume tasks. DeepSeek V4 Flash generates output at approximately 140 tokens per second — roughly 2x the speed of Claude Opus 4.7 and 1.5x the speed of GPT-5.5 at comparable quality tiers. A 400-word response comes back in under 9 seconds. For applications requiring real-time responses — chat, live document processing, streaming agents — this latency advantage is functional, not cosmetic.

Where DeepSeek Still Loses — Specifically

• Production software engineering. On SWE-bench Pro — the benchmark most predictive of real-world software engineering performance — Claude Opus 4.7 scores 64.3%, GPT-5.5 scores 58.6%, and DeepSeek V4 Pro scores approximately 51% based on third-party evaluations (no official submission as of May 17, 2026). The gap reflects something specific: DeepSeek struggles with tasks that require understanding implicit project conventions, navigating undocumented API behavior, and making judgment calls about architectural tradeoffs. These are the tasks where Claude's RLHF training on human code review feedback produces an advantage. Source: SWE-bench Pro Leaderboard, April 2026.
• Writing quality and long-form content requiring a distinct voice. In blind writing evaluations — same prompt, four models, human raters who didn't know the source — Claude Opus 4.7 beat DeepSeek V4 Pro across 40 writing tasks. The failure mode is specific: DeepSeek writes complete, organized, technically correct prose. It lacks personality. It answers every part of the prompt, structures information clearly, and produces nothing you would screenshot. For structured reports, this doesn't matter. For content that moves people to share it, it does.
• Agentic and multi-tool workflows. On MCP-Atlas, Claude Opus 4.7 leads at 77.3% versus DeepSeek V4 Pro's 73.6% (per Vellum's April 2026 benchmark analysis). For enterprises running multi-agent pipelines where agents need to coordinate across specialized subtasks, Anthropic's Agent Teams is a qualitative capability gap DeepSeek V4 cannot currently close.
• US-context knowledge and cultural specificity. DeepSeek's training data underrepresents American legal specifics, US healthcare system details, and US regulatory frameworks compared to Claude and ChatGPT. Testing 20 US-specific knowledge questions — IRS tax categories, state-level employment law, American insurance terminology, US banking regulations — DeepSeek answered 13 of 20 correctly. Claude answered 19. GPT-5.5 answered 18. If your work touches anything distinctly American in its rules or institutional structure, this gap shows up on the first task that requires it.

The Factor Most Comparisons Skip: Open Weights Changes What the Product Is

Both DeepSeek V4 models are released under the MIT license with full weights downloadable from HuggingFace. V4 Flash is 160GB. V4 Pro is 865GB. Downloadable weights change what the product actually is. Claude Opus 4.7 and Sonnet 4.6 are API-only — you use them through Anthropic's infrastructure, subject to Anthropic's terms, pricing, and availability. DeepSeek V4 Pro can run on your own infrastructure: no per-token billing, no data leaving your environment, no dependency on a provider's uptime. Self-hosting V4 Pro requires serious hardware — typically 8x H100 or A100 GPUs even quantized. But for enterprises with sensitive data, compliance requirements against sending data to any third-party API, or interest in fine-tuning on proprietary datasets, the model's existence as a downloadable asset changes the calculus entirely. You pay in infrastructure and maintenance overhead. You own everything else.

Stop Asking 'DeepSeek or Claude.' Ask This Instead.

The DeepSeek-vs-Claude framing collapses the moment you ask the right question: which model for which specific job, at which price, with which risk tolerance. Here are the clear calls — no hedging.

• Default daily coding and development work → Claude Sonnet 4.6. 79.6% SWE-bench Verified, $3/$15 per million tokens, 40–60 tokens/second. Replaces Opus for 80% of coding tasks at 5x lower cost. Start here.
• High-volume automated pipelines, bulk processing, cost-constrained builds → DeepSeek V4 Flash. $0.14/$0.28 per million tokens, approximately 140 tokens/second, 79% SWE-bench Verified. At this price point, running the same task three times and taking the best result costs less than one Claude Sonnet 4.6 call. The reliability caveat: tool calling in multi-step pipelines shows more error-recovery gaps than Claude — budget for retry logic in your architecture.
• Algorithmic and competitive programming work → DeepSeek V4 Pro. 93.5 LiveCodeBench (highest of any model), 3,206 Codeforces rating, $1.74/$3.48 per million tokens list price. The MIT license makes it the first frontier-class model you can self-host without licensing risk.
• Complex production engineering, real-world software architecture, multi-agent coordination → Claude Opus 4.7. 87.6% SWE-bench Verified, 64.3% SWE-bench Pro (13 points ahead of V4 Pro), 77.3% MCP-Atlas. Agent Teams for parallel multi-Claude orchestration. Extended Thinking for deep step-by-step reasoning.
• Expert scientific research, medical analysis, graduate-level reasoning → Claude Opus 4.7. 94.2% GPQA Diamond, 54.7% HLE with tools. All frontier models have converged on GPQA — the differentiation is in applied reasoning depth at the extreme edge, where Opus 4.7 still leads.
• Desktop and GUI automation → Claude Opus 4.7 for maximum performance (78.0% OSWorld), Claude Sonnet 4.6 for cost-effective automation (72.5% OSWorld). DeepSeek has no published OSWorld results. For teams where 5.5 percentage points of computer use performance is worth 5x the cost, choose Opus 4.7.
• US regulated industries (healthcare, finance, legal, government) → Claude Sonnet 4.6 or Opus 4.7, no exceptions. Data routing requirements and regulatory compliance in HIPAA, CCPA, financial services, and government contracting make DeepSeek an unsuitable choice regardless of benchmark performance.
• Content generation at scale where cost matters → DeepSeek V4 Flash. At $0.28 per million output tokens, high-volume content operations that would cost $15,000/month on Sonnet 4.6 run for $700 on V4 Flash — with SWE-bench scores within 1 point of Sonnet. For content workflows without regulatory sensitivity, the price case is compelling.

Frequently Asked Questions

The Honest Bottom Line

The DeepSeek V4 launch is the most significant pricing event in the AI industry since DeepSeek R1 moved markets in January 2025. The difference is that R1 was primarily a benchmark story — impressive numbers, not yet a clear replacement for production workflows. V4 is different: MIT license, open weights, 1M token context, production-ready API with OpenAI and Anthropic format compatibility, and benchmark scores that match Claude Opus 4.6 on the most important software engineering benchmark. The cost gap — 7x on V4 Pro, over 100x on V4 Flash versus Opus 4.7 — is structural, not promotional. It comes from an architectural choice (MoE with 49B active parameters out of 1.6T total) that was itself forced by US export controls that denied DeepSeek access to Nvidia's best chips. The policy designed to limit Chinese AI capability may be the reason DeepSeek built the most cost-efficient frontier model in the world.

Claude Sonnet 4.6 remains the right daily driver for most developers — faster, US-based, with the most mature agentic infrastructure. Claude Opus 4.7 is Anthropic's current flagship and now holds meaningful leads on SWE-bench Pro, OSWorld, and MCP-Atlas that did not exist in the Opus 4.6 generation. DeepSeek V4 Flash should be in your stack for high-volume non-sensitive workloads, and DeepSeek V4 Pro deserves serious evaluation for algorithmic programming and any pipeline where cost is the constraint and data doesn't touch regulatory requirements. The honest update as of May 18, 2026: the AI cost structure just broke in half. The teams that re-route their workloads this month will have a cost advantage their competitors will not understand until 2027. The teams that wait will pay 7x to 107x more for the same output — and call it 'being cautious.'