OPENAI, Anthropic and now Google offer semi-automated code agents at the cutting edge of technology with notable differences.
It is one of the verticals of the generative AI that advances the fastest. Artificial intelligence for the code promises to significantly increase the work of developers by removing tasks with low added value. Code assistants perfectly illustrate this principle. Available directly in a terminal they allow you to generate, modify debug and even test code. Claude Code, Codex Cli, Gemini Cli… The three main models of market models offer often similar agents in features but with some specificities. Comparative.
Features: Gemini CI better equipped
| Main features | Claude Code | CODEX CLA | Gemini Crica |
|---|---|---|---|
| Code generation | ✅ | ✅ | ✅ |
| Modification/refactorization | ✅ | ✅ | ✅ |
| Automated debug | ✅ | ✅ | ✅ |
| Unit tests | ✅ | ✅ | ✅ |
| Git integration | ✅ | ✅ | ✅ |
| Planning | ✅ | ✅ | ✅ |
| Multi-fichiers changes | ✅ | ✅ | ✅ |
| PDF support/Documents | ❌ | ✅ | ✅ |
| Generation of images | ❌ | ❌ | ✅ |
| Screenshot/diagrams | ❌ | ✅ | ✅ |
| MCP protocol | ✅ | ❌ | ✅ |
| Sandboxing | ❌ | ✅ | ✅ |
| Configurable modes of autonomy | ❌ | ✅ | ❌ |
| High context | ❌ | ❌ | ✅ |
The three code agents share a large base of common features: generation of code, modification and refactoring, automated debugging, unit test support, git integration and planning. All also support multi-spinning changes. The differences are widening on advanced features. Claude Code accuses delay on multimodality, he does not support PDF documents, the generation of images, the interpretation of screenshots or diagrams, unlike his competitors. Sandboxing, which allows the agent to isolate in a container, is also lacking at Anthropic. Only COSX CLI offers configurable methods of autonomy to precisely define the desired automatic intervention level. On the context side, Openai and Anthropic sink with a capacity limited to 200,000 tokens (see detail below). In addition, Gemini has the MCP natively. He also has the ability to generate images with Imagen.
Finally, on availability, Claude Code is not natively supported on Windows and requires WSL, a potential source of additional complications.
Models: Codex plays balance, Gemini the context
| Model | Context size | Livecodebench |
Swe-Bench Verified |
|---|---|---|---|
| Claude 4 opus | 200,000 tokens | 51.10% | 72.50% |
| Gemini 2.5 Pro | 1,000,000 tokens | 69% | 59.60% |
| Codex-1 | 192,000 tokens | > 72%* | 71% |
** Codex-Mini codex scores on benchmarks are not published by Openai, we used those of O4-Mini, basic codex-mini model before its specialization on the code. In reality, the results are very close or slightly higher.
To assess the capacities of the three agents, we have analyzed their most efficient motor models: Claude 4 opus for Claude Code, Gemini 2.5 PRO for Gemini CLI and Codex-Mini for CODEX CLI. Two reference benchmarks make it possible to measure the key skills expected from a code agent. Swe-Bench Verified assesses Software Engineering tasks, in particular the autonomous management of tools, the use of console and complex workflows orchestration. Livecodebench tests the raw capabilities of code generation.
In the Benchmarks Claude 4 Opus excels in the Software Engineering and the rapid edition of code but slightly sins in generation of raw code. Gemini 2.5 Pro demonstrates an excellent generation capacity but lower performance in the autonomous use of tools. Finally, Codex-Mini seems to offer the best balance with good scores on the two benchmarks.
However, Gemini Cli is doing well thanks to his context window of a million tokens, perfect for analyzing the vast code bases, very common in a professional environment. Claude Code and Codex CLI, limited to 200,000 tokens, must cut the code into segments and potentially lose in efficiency on large projects.
Gemini, the best value for money, without a doubt
It is on pricing that we distinguish from notable differences.
Claude Code Adopt a premium subscription model with three levels:
- Claude Pro at 20 dollars per month offers very limited access to Claude Code (less than an hour of daily use according to our estimates)
- Claude Max at 100 dollars allows a few hours of use per day
- Claude Max Max at 200 dollars removes almost all limitations. Use via the API remains an expensive alternative with Claude 4 opus billed $ 15 per million tokens at the start and 75 dollars out.
CODEX CLA Building on a model exclusively based on consumption, without subscription. The Codex-Mini model is billed $ 1.50 the million tokens at the start and $ 6 output.
In contrast, Google adopts a fairly disruptive strategy with Gemini Crica : It is completely free. With 60 requests per minute within the limit of 1000 per day, a quota that Google describes as “double of internal use before public exit”.
Which agent for whom?
For developers working on vast code bases or seeking excellent value for money, Gemini cli stands out as the perfect choice. Its context window of a million tokens and its entirely free access make it the ideal tool to analyze and quickly modify code on large projects. For its part, Claude Code remains a reference for professional teams favoring reliability and autonomy. It is particularly good in debugging and the Software Engineering complex tasks. Despite its limitation to 200,000 tokens, which can be problematic on major projects, its excellence in workflows orchestration justifies its high cost for companies with a dedicated budget.
Finally, Codex CLI, although technically solid with a good balance on benchmarks, suffers from its token pricing model which can quickly become expensive and unpredictable, limiting its adoption despite its technical qualities and flexibility in the choice of models.
