Mistral Vibe Codex vs Claude Codex vs Cursor vs Codex: Four Agents Scored in One Scaffold-to-PR Job

The Mistral Vibe is the combined agent of Mistral for work and code, and the product known as Le Chat. Its coding environment uses the open source CLI on GitHub under Apache 2.0.

The model stack is overridden, and the differences are significant. Mistral's FAQ, Mistral Medium is a high-performance family for complex, multi-step software engineering. The Vibe CLI and IDE plugins are powered by Devstral. Codestral handles fast completion, and Codestral Embed enables semantic code searches. Remote agents work on Mistral Medium 3.5. Judging the Vibe solely by the Devstral benchmark is letting the product down.

In scaffolding, Vibe scans your file tree and Git environment for project-aware context, then performs multi-file orchestration with architecture-level reasoning. Devstral 2 is a dense 123B model with a 256K core window. The Mistral reports 72.2% in SWE-bench Verified, putting it at the state of the art among open models. Devstral Small 2 (24B, Apache 2.0) gets 68.0% and runs on consumer hardware.

First-stage feature testing: automatically generated tests that run through the codebase, matched to your existing patterns. Hooks run custom shell commands before and after each agent's turn, so you can apply policies or block patterns automatically.

The PR and async stage is well designed. Remote coding agents run in individual sandboxes for parallel use, and sessions continue while your machine is turned off. /teleport provides a live session from the local CLI to the remote agent and back. Remote agent visualization provides run history, tool calls, code changes, logs, and test traces.

The land coverage is very wide hereand this is where Vibe gets a clean 5: terminal CLI, VS Code, JetBrains, Zed, web app, mobile, and background agents. It also deploys tab-to-complete, which removes auto-complete as a reason to keep a separate tool.

Cost and control are Vibe's defining leads. Pro is $14.99/month, the cheapest premium tier of the four. Team is $24.99/user/month. There is a $5.99 student tier. Mistral claims the Devstral 2 is up to 7x more economical than the Claude Sonnet in real-world operations – again, the seller's claim. You can host it yourself, host it in a private cloud or on-premises, fine-tune proprietary code, and model training to opt out of paid plans.

Where Vibe trails: 72.2% of Devstral sits under closed models in code evals, and reviewers have flagged rate-limiting bugs and occasional instability. The CLI targets UNIX first, so Windows-first teams should call before committing.

Claude Code is an Anthropic coding tool, which uses Claude Opus 4.8 by default. Their harnesses are the deepest in the field: 30 lifecycle hooks, Skills, Plugins, Subagents, checkpoints, program mode, and MCP.

It leads to raw execution. Dynamic Workflows organize large groups of parallel subagents simultaneously. Background functions, auto-targeting, and a mature test validation loop make scaffold-test-PR its home base. One widely cited piece of evidence: Bun creator Jarred Sumner reported migrating approximately 750,000 lines from Zig to Rust with a 99.8% pass rate in 11 days.

Weaknesses are cost and control. There are no open weights, no self-control, no data retention issues for the kind of controlled teams you need. Pricing starts at $20 Pro, $100 Max 5x, $200 Max 20x, and Team seats on top. Token burning is the real bill: public postmortems document parallel subagent costs thousands of dollars, and Anthropic's own reported rate is about $13 per developer per working day before the parallels multiply it.

Codex is an OpenAI coding agent: Apache-2.0 CLI that also serves as a cloud service, IDE extension, ChatGPT app, iOS app, and, as of June 2026, Amazon Bedrock. GPT-5.6 reached general availability on 9 July 2026in three categories – Sol, Terra, and Luna – replaces the default GPT-5.5 quoted in classic comparisons.

Codex handles scaffolding and testing within a kernel-level sandbox with networking disabled by default. Submits Skills, Marketable Plugins, Subagents, Hooks, and MCP. Its standout is cross-surface async: a single task moves between CLI, cloud, app, and mobile without losing state.

Pricing is tiered: Free, Go for $8, Plus for $20, Pro 5x for $100, Pro 20x for $200, and Business and Business seats. A reliable hold is a 5-hour rolling window. Developers report heavy repo sessions that exhaust the grant in about an hour. OpenAI's scorecard documents estimate the cost of Codex approx $100–$200 per developer per monthwith wide variety in model, likeness, and fast mode.

Cursor is an AI code editor for Anysphere, a fork of VS Code. Its in-house coder 2.5 is designed for fast agent programming and, Artificial Analysis, scored 62 points in the Coding Agent Index. Cross-platform models from Anthropic, OpenAI, and Google in one place, and deploy Rules, MCP, Hooks, Skills, Plugins, and Subagents.

The cursor is great for a different task. It's IDE-first and inline-driven, and its tab completion and single-file replication are best in class. In the autonomous scaffold-test-PR loop, terminal-first and agent-first have a structural edge. That is fair judgment, not good judgment.

Costs need to be taken care of. Each paid category is a credit pool limited by its price: Hobby $0, Pro $20, Pro+ $60, Ultra $200, Groups $40/user. Once the pool is depleted, credit usage is billed at API rates, so $20 is entry, not a fee. The June 2026 teams revised the division into first-person and third-party pools and added a $120 Premium seat.

Key Takeaways

• Code's Mistral Vibe led the overall score at 22/25, driven by cost, openness, and control.
• Claude and Codex are bound for 21/25; Claude Code leads in raw boundary coding quality.
• Vibe is the only tool here that offers self-hosting, optimization, and residency of EU data.
• Vibe's open weights are Apache 2.0 research; commercial deployment requires a Mistral license.
• The benchmark scores take into account different suits and dealer claims, so they are not directly comparable.

Interactive Descriptor