Coding agents are the most contested category in developer tooling right now. Four names dominate the shortlist: Mistral Vibe for Code, Claude Code, Cursor, and OpenAI Codex. Each claims to take a feature from prompt to pull request.
This comparison runs all four against one practical workflow. Not a toy script. A real unit of engineering work: scaffold a feature across multiple files, generate and run tests, then open a pull request.
The task
Prompt: ‘In our existing Python/FastAPI service, add a /subscriptions endpoint. Scaffold the route, the Pydantic models, and a service layer across the correct files. Generate unit and integration tests, run them, fix any failures, then open a pull request with a clear description.’
That prompt maps to three stages every agent must clear: scaffold, test, ship.
Methodology, stated honestly
This is a capability comparison, not a single timed run on one machine. Scores reflect documented features, published benchmarks, and vendor specifications as of July 14, 2026. Each dimension is rated 1 to 5, for 25 points maximum. Every score carries a one-line reason.
Three caveats that matter more than the scores:
- Benchmark numbers here are not directly comparable. SWE-bench Verified and SWE-Bench Pro are different suites with different difficulty. Terminal-Bench is a third. Never read across them as one scale.
- Vendor claims are labeled as such. Mistral’s cost-efficiency figure is Mistral’s own published claim, not an independent result.
- These products ship weekly. Model defaults and prices verified today may move. Every claim below links to its source so you can re-check.
The five dimensions: feature scaffolding, test generation and run loop, PR and async workflow, surface coverage, and cost/openness/control.
Results
Mistral Vibe is Mistral’s unified agent for work and code, and the product formerly known as Le Chat. Its coding surface runs an open-source CLI on GitHub under Apache 2.0.
The model stack is layered, and the distinction is important. Per Mistral’s own FAQ, Mistral Medium is the most performant family for complex, multi-step software engineering. The Vibe CLI and IDE plugins are powered by Devstral. Codestral handles fast completions, and Codestral Embed powers semantic code search. Remote agents run on Mistral Medium 3.5. Judging Vibe solely by Devstral’s benchmark understates the product.
On scaffolding, Vibe scans your file tree and Git status for project-aware context, then does multi-file orchestration with architecture-level reasoning. Devstral 2 is a 123B dense model with a 256K context window. Mistral reports it at 72.2% on SWE-bench Verified, which it positions as state of the art among open models. Devstral Small 2 (24B, Apache 2.0) scores 68.0% and runs on consumer hardware.
Testing is a first-class feature: auto-generated tests that evolve with the codebase, matched to your existing patterns. Hooks run custom shell commands before and after each agent turn, so you can enforce conventions or block patterns automatically.
The PR and async stage is well-designed. Remote coding agents run in isolated sandboxes for parallel execution, and sessions persist while your machine is off. /teleport hands a live session from local CLI to a remote agent and back. Remote agent observability gives run history, tool calls, code changes, logs, and audit trails.
Surface coverage is the widest here, and it is where Vibe scores a clean 5: terminal CLI, VS Code, JetBrains, Zed, web app, mobile, and background agents. It also ships tab-to-complete, which removes autocomplete as a reason to keep a separate tool.
Cost and control is Vibe’s decisive lead. 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 Devstral 2 is up to 7x more cost-efficient than Claude Sonnet on real-world tasks — again, a vendor claim. You can self-host, deploy in a private cloud or on-premises, fine-tune on proprietary code, and model training is opt-out on paid plans.
Where Vibe trails: Devstral’s 72.2% sits below frontier closed models on coding evals, and reviewers have flagged rate-limiting bugs and occasional instability. The CLI targets UNIX first, so Windows-first teams should pilot before committing.
Claude Code is Anthropic’s agentic coding tool, running Claude Opus 4.8 by default. Its harness is the deepest in the field: 30 lifecycle hooks, Skills, Plugins, Subagents, checkpoints, plan mode, and MCP.
It leads on raw execution. Dynamic Workflows orchestrate large fleets of parallel subagents in a single session. Background tasks, automatic checkpointing, and a mature test-verify loop make the scaffold-test-PR arc its home turf. One widely cited proof point: Bun creator Jarred Sumner reported porting roughly 750,000 lines from Zig to Rust at a 99.8% test pass rate in 11 days.
The weakness is cost and control. No open weights, no self-hosting, no data residency story of the kind regulated teams need. Pricing runs $20 Pro, $100 Max 5x, $200 Max 20x, with Team seats above that. Token burn is the real bill: public postmortems document parallel subagent runs costing thousands of dollars, and Anthropic’s own reported average is about $13 per developer per active day before parallelism multiplies it.
Codex is OpenAI’s coding agent: an Apache-2.0 CLI that also runs as a cloud service, IDE extension, ChatGPT app, iOS app, and, since June 2026, on Amazon Bedrock. GPT-5.6 reached general availability on July 9, 2026, in three tiers — Sol, Terra, and Luna — replacing the GPT-5.5 defaults cited in older comparisons.
Codex handles scaffolding and testing inside a kernel-level sandbox with network off by default. It ships Skills, Plugins with a marketplace, Subagents, Hooks, and MCP. Its standout is cross-surface async: one task moves between CLI, cloud, app, and mobile without losing state.
Pricing is a ladder: Free, Go at $8, Plus at $20, Pro 5x at $100, Pro 20x at $200, plus Business and Enterprise seats. The honest catch is the 5-hour rolling window. Developers report serious-repo sessions exhausting the allowance in about an hour. OpenAI’s own rate-card documentation estimates Codex costs roughly $100–$200 per developer per month, with wide variance by model, parallelism, and fast mode.
Cursor is Anysphere’s AI code editor, a VS Code fork. Its in-house Composer 2.5 is tuned for fast agentic editing and, per Artificial Analysis, scored 62 on its Coding Agent Index. It routes across frontier models from Anthropic, OpenAI, and Google in one surface, and ships Rules, MCP, Hooks, Skills, Plugins, and Subagents.
Cursor is excellent at a different shape of work. It is IDE-first and inline-driven, and its tab completion and single-file iteration are best-in-class. On an autonomous scaffold-test-PR loop, terminal-first and agent-first tools have the structural edge. That is a fit judgment, not a quality judgment.
Cost needs care. Every paid tier is a credit pool sized to its price: Hobby $0, Pro $20, Pro+ $60, Ultra $200, Teams $40/user. Once the pool is spent, usage bills in arrears at API rates, so $20 is the entry, not the bill. A June 2026 Teams update split usage into first-party and third-party pools and added a $120 Premium seat.
Key Takeaways
- Mistral Vibe for Code led on total value at 22/25, driven by cost, openness, and control.
- Claude Code and Codex tied at 21/25; Claude Code leads on raw frontier coding quality.
- Vibe is the only tool here offering self-hosting, fine-tuning, and EU data residency.
- Vibe’s open weights are Apache 2.0 for research; commercial deployment needs a Mistral license.
- Benchmark scores span different suites and vendor claims, so they are not directly comparable.
Interactive Explainer
How we compared: A capability comparison scored 1–5 across five dimensions (25 max), based on documented features, published benchmarks, and vendor specifications — not a timed head-to-head run. No agents were executed against the prompt, and no timing or pass-rate results are claimed. Package versions, licenses, default models, and shipped token prices verified July 14, 2026 against mistral-vibe 2.19.1 via PyPI and npm. Vendor claims are labeled. Benchmark figures span different suites and are not directly comparable. Prices subject to change.
