Head-to-head comparison of DeepSeek V3, Claude Opus 4.6, and GPT-5 on coding, reasoning, and cost. Includes SWE-bench, Chatbot Arena, and real-world task results.
DeepSeek V3 changed the game. An open-weight model from China matching frontier performance at 1/50th the cost. We ran benchmarks to see how it actually compares to Claude Opus 4.6 and GPT-5 in real engineering tasks.
TL;DR
| Model | Coding | Reasoning | Cost/1M | Verdict |
|---|---|---|---|---|
| DeepSeek V3 | 9.2 | 8.8 | $1.37 | Best value |
| Claude Opus 4.6 | 9.4 | 9.5 | $90 | Most reliable |
| GPT-5.2 | 9.1 | 9.0 | $40 | Most versatile |
Winner by use case:
- Budget-conscious: DeepSeek V3
- Mission-critical: Claude Opus 4.6
- Multimodal needs: GPT-5.2
Benchmark Results
SWE-bench Verified (Code Bug Fixing)
| Model | Score | Notes |
|---|---|---|
| Claude Opus 4.6 | 72.5% | Best on complex refactors |
| DeepSeek V3 | 71.2% | Strong on standard patterns |
| GPT-5.2 | 70.8% | Good overall coverage |
| Gemini 3 Pro | 74.2% | Current leader |
SWE-bench measures ability to fix real GitHub issues. DeepSeek trails Claude by only 1.3 percentage points despite costing 98% less.
Chatbot Arena Elo (Crowdsourced Quality)
| Model | Elo Rating | Rank |
|---|---|---|
| GPT-5.2 High Reasoning | 1420 | #1 |
| Claude Opus 4.6 | 1405 | #3 |
| DeepSeek V3 | 1385 | #8 |
| Gemini 3 Pro | 1410 | #2 |
Chatbot Arena captures subjective quality perception. DeepSeek ranks in the top 10 globally — the highest-ranked open-weight model ever.
GPQA Diamond (Graduate-Level Reasoning)
| Model | Score |
|---|---|
| Claude Opus 4.6 | 71.2% |
| GPT-5.2 | 69.8% |
| DeepSeek V3 | 65.4% |
| Gemini 3 Pro | 68.9% |
GPQA tests PhD-level reasoning in biology, physics, and chemistry. Claude maintains a 5.8-point lead over DeepSeek here.
HumanEval (Code Generation)
| Model | Pass@1 |
|---|---|
| DeepSeek V3 | 85.2% |
| Claude Opus 4.6 | 88.1% |
| GPT-5.2 | 86.5% |
| Gemini 2.5 Pro | 84.9% |
HumanEval measures functional correctness on 164 Python problems. DeepSeek excels here — it’s optimized for code synthesis.
Our Task-Level Evaluation
We tested all three models on our standard eval suite:
Coding: Pagination Bug Fix
| Model | Score | Notes |
|---|---|---|
| Claude Opus 4.6 | 9.4 | Best validation, thorough explanation |
| DeepSeek V3 | 9.2 | Clean diff, minor style issues |
| GPT-5.2 | 9.1 | Correct but verbose |
Reasoning: Build vs Buy Decision
| Model | Score | Notes |
|---|---|---|
| Claude Opus 4.6 | 9.5 | Decisive, excellent tradeoff matrix |
| GPT-5.2 | 9.0 | Good analysis, slightly hedged |
| DeepSeek V3 | 8.8 | Solid reasoning, less structured output |
Tool Use: Stripe Webhook Setup
| Model | Score | Notes |
|---|---|---|
| Claude Opus 4.6 | 9.3 | Best security emphasis |
| GPT-5.2 | 8.9 | Complete steps |
| DeepSeek V3 | 8.7 | Correct but missed webhook secret note |
Cost Analysis
Price per Million Tokens
| Model | Input | Output | Total (1:1 ratio) |
|---|---|---|---|
| DeepSeek V3 | $0.27 | $1.10 | $1.37 |
| Claude Opus 4.6 | $15 | $75 | $90 |
| GPT-5.2 | $10 | $30 | $40 |
DeepSeek is 65x cheaper than Claude Opus and 29x cheaper than GPT-5.
Cost for 1M Complex Queries
Assuming 2,000 input tokens and 1,000 output tokens per query:
| Model | Cost for 1M queries |
|---|---|
| DeepSeek V3 | $1,640 |
| Claude Opus 4.6 | $105,000 |
| GPT-5.2 | $50,000 |
Where DeepSeek Wins
- High-volume code generation — Near-frontier quality at 2% of the cost
- Self-hosting — Open weights mean you can run it on your own infrastructure
- Chinese-language tasks — Strong multilingual performance
- Budget prototyping — Iterate cheaply, upgrade to Claude/GPT for production
Where Claude/GPT Still Lead
- Complex reasoning chains — Claude’s 9.5 vs DeepSeek’s 8.8 on reasoning tasks
- Security-critical code — Claude’s better at catching edge cases
- Structured output — Claude more reliably follows formatting instructions
- Multimodal tasks — GPT-5 and Gemini have better image/audio integration
- Enterprise support — Anthropic and OpenAI offer SLAs, DeepSeek does not
The Convergence Story
The gap between open and closed models is closing fast:
| Year | Open vs Closed Gap (Chatbot Arena) |
|---|---|
| 2024 | 8.0% |
| 2025 | 2.5% |
| 2026 | 1.7% |
DeepSeek V3 proves you don’t need a $10B training run to reach frontier performance. This is the new normal.
Our Recommendation
| Scenario | Pick |
|---|---|
| Startup, cost-sensitive | DeepSeek V3 |
| Enterprise, reliability-critical | Claude Opus 4.6 |
| Need multimodal | GPT-5.2 or Gemini 3 |
| Self-hosting required | DeepSeek V3 |
| Chinese-language focus | DeepSeek V3 or GLM-5 |
Data Sources
Related: See our pricing guide for detailed cost breakdowns.