Mistral Vibe for Code vs Claude Code vs Cursor vs Codex: One Task, Four Agents, One Winner

MarkTechPost ran a head-to-head this week that skips the usual benchmark-suite approach entirely: one practical task, four coding agents, scored side by side. The task was deliberately ordinary — scaffold a FastAPI endpoint, generate tests, and open a pull request, the kind of ticket that fills most agentic coding sessions. Mistral Vibe for Code came out on top at 22/25 on a cost/openness/control rubric, with Claude Code and OpenAI Codex tied at 21/25 (Claude Code rated strongest on raw frontier coding quality), and Cursor trailing at 16/25. If you have been picking your daily-driver agent off leaderboard vibes rather than a task that looks like your actual work, this comparison is worth a closer read than the headline ranking suggests.

What You'll Learn

Why a single scaffold-to-PR task can be a more useful signal than an aggregate benchmark score; what the cost/openness/control rubric actually rewards and why that is a different axis than raw coding quality; how Claude Code and Codex tying at 21/25 while scoring differently under the hood should change how you read a single composite number; and a practical way to weigh Mistral Vibe for Code's win against Claude Code's frontier-quality edge when neither is a universal answer.

Step 1: Separate the Rubric From the Task

The scaffold-to-PR task (build a FastAPI endpoint, write tests, open a PR) is the same for all four agents, but the scoring rubric — cost, openness, control — is not measuring "which agent wrote the best code." It is measuring how cheaply and controllably each agent got a reasonable result. Mistral Vibe for Code's 22/25 win reflects that framing: strong marks on cost and openness, not necessarily the single sharpest diff. Before you treat this as "Mistral beat Claude Code," check which axis actually matters for your use case.

Step 2: Read the Claude Code / Codex Tie Correctly

Claude Code and Codex tying at 21/25 hides a real difference — the piece notes Claude Code was rated strongest on raw frontier coding quality. A tied composite score with a stated quality edge for one side means the gap is being made up elsewhere, most likely cost or control overhead. If your task is closer to "gnarly refactor that needs to be right the first time" than "routine scaffold," the quality sub-score matters more than the tied total.

Step 3: Don't Write Off Cursor From One Task

Cursor's 16/25 on this specific task is the headline most people will remember, but a single scaffold-to-PR run is a narrow slice of what any of these tools do across a real week of work — IDE integration, multi-file navigation, and inline review UX do not show up in a task scored purely on cost/openness/control and PR output. Treat a 16/25 on one task as a data point to investigate, not a verdict to act on for tools you already rely on for other workflows.

Step 4: Match the Rubric to What You Actually Optimize For

If cost-per-task and control over the agent's execution path are what matter most to your team — think high-volume, low-risk scaffolding work — Mistral Vibe for Code's win is directly actionable. If you are optimizing for the agent least likely to produce a diff you need to rewrite on a hard problem, Claude Code's frontier-quality edge is the more relevant signal even inside a tied composite score. Pick the axis that matches your actual failure mode, not the one the headline ranking implies.

Common Challenges

"Doesn't a tied score mean the agents are basically interchangeable?" — Only on this task, on this rubric. A composite score can tie while the underlying quality, cost, and control numbers diverge meaningfully, which is exactly what happened between Claude Code and Codex here. "One scaffold-to-PR task seems like a small sample size" — It is, and that is the honest trade-off of task-specific evaluations versus benchmark suites: you get a concrete, inspectable result instead of an abstract aggregate, but you lose generalizability. Use it as one input, not the whole decision. "Should we switch our default agent based on this?" — Not off a single article. Run the same kind of task — one you actually repeat — through your current agent and whichever one scored better here before changing a team default.

Advanced Tips

Build your own scoring rubric before your own bake-off. Cost, openness, and control are reasonable axes for this evaluation's goals, but your team might weight review-time-saved or IDE integration higher — define your rubric before running the comparison, not after, so you are not unconsciously scoring toward whichever tool you already prefer. Run the losing tool on a task it should be good at. Cursor's 16/25 here was on a backend-scaffolding task; if your team leans on Cursor for its editor-native multi-file refactor UX, that is a different workload than what this comparison tested — don't let one score generalize past its task. Track cost per merged PR, not cost per token. The rubric's "cost" axis is more useful when translated into your team's actual unit economics — a cheaper agent that needs two follow-up prompts to get a mergeable PR is not automatically the cheaper choice in practice.

Conclusion

The real value in this shootout is not the 22/25 vs 21/25 vs 16/25 scoreboard — it is the reminder that a single, concrete task run head-to-head across agents surfaces trade-offs an aggregate benchmark suite smooths over. Mistral Vibe for Code's win on cost and control, alongside Claude Code's tied score with a quality edge, both point toward the same conclusion: pick the axis you actually optimize for before picking the tool. For more on how to read agentic coding benchmarks without over-indexing on one headline number, see our recent piece on the Sonnet 5 / Opus 4.8 benchmark comparison, and for the continuously-updated tool-by-tool breakdown, check the Tool Comparison Matrix in the ebook. For daily coverage of comparisons like this as they land, subscribe to the EndOfCoding newsletter.