Codex vs Claude Code with DeepSeek: Real-World Test Results — 5.7x Speed Gap, Stability Score 40 vs 100

Background: Does Codex Really Crush Claude Code?

Recently, the developer community has been flooded with claims that "Codex crushes Claude Code," but have these conclusions been rigorously tested? One developer decided to put it to the test with a scientific approach — when both AI coding tools are paired with the DeepSeek model, which one actually performs better?

The results surprised many: the Claude Code + DeepSeek combination comprehensively outperformed Codex + DeepSeek in both stability and speed, with a staggeringly large gap.

Industry Context: The AI coding tool market is undergoing a "model decoupling" transformation. Early tools were tightly bound to specific models (e.g., Copilot to OpenAI, Claude Code to Anthropic), but with the rise of cost-effective models like DeepSeek, developers are increasingly seeking the flexibility to choose their own underlying models. This has given rise to two technical approaches: one where tools natively support multiple third-party protocols, and another that uses intermediate proxy layers for protocol conversion. The differences in stability and performance between these two approaches are exactly the core tension revealed by this benchmark.

Test Design: Using AI to Test AI

The methodology behind this benchmark is worth noting. The tester used VS Code Copilot with OPAI 4.7 to design the comparison test plan, then used Sonnet 4.6 to execute the tests and generate reports — the entire process embodied the concept of "using AI to test AI."

This "AI-on-AI" automated evaluation paradigm is gaining traction. Its strengths lie in more systematic test case design, reproducible execution, and efficient report generation. However, it also has limitations: AI-designed test scenarios may have coverage blind spots, and 30 samples yield a relatively wide confidence interval statistically. For engineering decisions, this type of benchmark is better suited as a "quick screening" tool rather than a definitive conclusion — supplementary validation with your own real-world business scenarios is recommended.

The benchmark took about two hours and consumed roughly 5 RMB (less than $1 USD) in DeepSeek API costs across 30 total test runs. While the sample size isn't massive, it provides meaningful reference value for tool-level comparisons. DeepSeek has rapidly become the go-to choice for developers in China thanks to its extremely low API costs (approximately 90% cheaper than the GPT-4 series) and code generation capabilities approaching top-tier models. Running 30 tests for about 5 RMB vividly demonstrates this cost advantage — and it's the core reason developers are eager to integrate it into various AI tools.

Test Results: The Gap Between Codex and Claude Code Is Striking

Response Speed: Claude Code Is 5.7x Faster

Codex + DeepSeek's response speed was 5.7 times slower than Claude Code + DeepSeek. In everyday coding scenarios, this means waiting several times longer for each interaction, severely impacting development efficiency and user experience.

Stability: 100 vs 40

Even more critical is the stability metric. Claude Code + DeepSeek scored a perfect 100, while Codex + DeepSeek scored only 40.

What does a stability score of 40 mean? Simply put — "basically unusable." You can't predict whether each API call will return correct results, which is unacceptable in real-world development.

Root Cause Analysis: Third-Party Model Protocol Support Is the Key

Why is the performance gap so large when both tools are paired with DeepSeek? The core reason lies in the different approaches to third-party model protocol support.

Claude Code: Native Third-Party Model Protocol Support

Claude Code natively supports third-party protocols. Users only need simple configuration to connect to third-party models like DeepSeek. The entire communication pipeline is well-optimized, resulting in stable and efficient performance.

Codex: Protocol Conversion Introduces Compatibility Issues

Codex supports OpenAI's newer protocol, but DeepSeek doesn't yet support this protocol. To understand this issue, you need to know the protocol evolution background: OpenAI's API has gone through multiple iterations, with the latest version introducing features like the Responses API, which differs significantly from the older Chat Completions API. Third-party models like DeepSeek typically prioritize compatibility with the older Chat Completions protocol (since it's more stable and better documented), and support for the newer protocol lags behind.

This means users must handle protocol conversion themselves (typically through an intermediate proxy layer). When Codex calls the new protocol while DeepSeek hasn't fully implemented it, the proxy layer needs to perform field mapping, request format conversion, and other operations. Each step can introduce latency or trigger edge cases that cause request failures — this extra conversion step directly causes the dual decline in both speed and stability, and is the fundamental reason behind the 40-point stability score.

Practical Selection Guide for Developers in China

For developers in China, if DeepSeek is your primary model, Claude Code is currently the recommended AI coding tool. DeepSeek's V3 and R1 series perform excellently across multiple coding benchmarks, with particular advantages in Chinese code comments, understanding of domestic frameworks, and other localized scenarios. Combined with Claude Code's native protocol support, you can maximize the strengths of both. The reasons are straightforward:

1. Native protocol support: No need for an extra protocol conversion layer; configuration is simple
2. Guaranteed stability: A perfect stability score means you can confidently use it in production environments
3. Significant speed advantage: A 5.7x speed difference is extremely noticeable in daily use

Of course, this conclusion has a specific scope — it applies only to scenarios involving DeepSeek. If each tool uses its native model (Codex with GPT series, Claude Code with Claude series), the results could be entirely different. Tool selection should ultimately be based on your actual use case.

Conclusion

This real-world comparison between Codex and Claude Code offers an important lesson: don't blindly follow the community's "X crushes Y" narratives. A tool's actual performance depends on the specific use case and configuration. When it comes to third-party model integration, Claude Code currently does a better job. Until Codex resolves its compatibility issues with the newer protocol, its experience when paired with DeepSeek has significant room for improvement.

Key Takeaways

• Claude Code + DeepSeek scored 100 on stability; Codex + DeepSeek scored only 40 — a massive gap
• Codex + DeepSeek was 5.7x slower than Claude Code + DeepSeek
• The core reason is a protocol support difference: Claude Code natively supports third-party protocols, while Codex requires extra protocol conversion (a generational gap between the new Responses API and DeepSeek's currently compatible Chat Completions protocol)
• Developers in China using DeepSeek are better off choosing Claude Code
• The benchmark was based on 30 tests, took two hours, and cost approximately 5 RMB in DeepSeek API fees