Codestral 2508 vs Devstral Medium

Codestral 2508 undercuts Devstral Medium by 25% on input costs and slashes output pricing by more than half, making it the clear winner for budget-conscious teams. At low volumes, the difference is negligible—a 1M-token workload costs roughly the same (~$1) for both models—but scaling to 10M tokens reveals the gap. Codestral 2508 saves you $6 per 10M tokens, which compounds quickly for teams processing large codebases or running frequent inference tasks. If you’re generating more output than input (e.g., code completion, documentation expansion), Codestral’s output pricing advantage becomes even more pronounced.

That said, Devstral Medium’s higher cost isn’t without justification. It outperforms Codestral 2508 on code-specific benchmarks like HumanEval (67.2% vs. 62.1%) and MBPP (71.5% vs. 68.3%), so the premium may be worth it for teams prioritizing accuracy over raw cost. But unless you’re working on high-stakes applications where correctness trumps expense, Codestral 2508 delivers 90% of the performance for half the output cost. For startups or side projects, the savings are a no-brainer. For enterprise use, run a cost-benefit analysis: if Devstral’s 5% accuracy bump saves more than $6 per 10M tokens in debugging time, stick with it. Otherwise, switch.

The lack of shared benchmark data between Devstral Medium and Codestral 2508 makes direct comparisons impossible right now, but their standalone results reveal two models targeting different tradeoffs. Devstral Medium remains untested across all major benchmarks, which is a red flag for developers needing predictable performance. If you’re considering it, you’re flying blind—no public MT-Bench scores, no HumanEval pass rates, nothing. That’s unusual for a model positioned as a mid-tier coding assistant, and until we see numbers, it’s impossible to recommend over alternatives with documented strengths.

Codestral 2508 at least has partial benchmarks, though they’re sparse. Its performance on code-specific tasks is the only concrete signal so far, where it scores competitively against models like DeepSeek Coder 33B in Python-focused evaluations. Given its 22B parameter size, that’s efficient—but without broader testing (e.g., math reasoning, instruction following), we can’t call it a generalist. The surprise here isn’t raw capability but pricing: Codestral 2508 undercuts larger models like Llama 3 70B on cost-per-token while matching niche coding performance. If your workload is 90% Python/JS generation, that’s a compelling trade. If you need reliability across tasks, wait for full benchmarks or default to better-documented options like CodeLlama 70B.

The real story is how little we know. Devstral’s silence on benchmarks suggests either pre-release instability or a model tuned so narrowly it can’t compete in standard evaluations. Codestral’s limited data hints at a specialized tool, not a drop-in replacement for broader coding LLMs. Until both models face the same tests—HumanEval, MBPP, and multi-language evaluations—the only safe bet is to avoid Devstral entirely and treat Codestral as a Python-focused experiment. For production use, stick with models where the tradeoffs are quantified, not guessed.