AI・テクノロジー
Mistral AI Releases "Codestral 2.5" — 62.4% on SWE-bench Reshapes the Options for Autonomous Code Repair
機械翻訳 / Machine-translated
機械翻訳 / Machine-translated
On August 22, 2026, Mistral AI officially released "Codestral 2.5," a code-specialized model. Its SWE-bench Verified score of 62.4% is considered the highest level currently achieved among open-weight models in the 32B parameter class. Available in both API and open-weight formats, it can also be deployed in closed, on-premises data center environments — a development that is beginning to shake up the procurement structures of enterprises that previously had no choice but to rely on proprietary solutions.
On the afternoon of August 22 local time, Mistral AI made its announcement via its official X account and blog. Key specifications are as follows:
"It took three hours to integrate Codestral 2.5 into our CI/CD pipeline. The bug detection rate changed noticeably. The fact that it's open-weight and can be kept entirely within our internal network was the deciding factor."
— Engineer at an unnamed SaaS company (via X post)
HumanEval is a benchmark that measures traditional single-function completion, and its difficulty differs from SWE-bench, which handles repository-scale problems closer to real-world work. The SWE-bench score of 62.4% is notable for surpassing Qwen 2.5 Coder's 59.1% among open models in the same size class, and for also exceeding the proprietary GPT-4o's 57.8%.
Since releasing the first version of Codestral in June 2024, Mistral has maintained its code-specialized direction. During that time, Anthropic's o4 and OpenAI's Claude line took the lead in agentic reasoning, while Mistral positioned "open-weight models unconstrained by execution environment" as its key differentiator.
The biggest architectural change in version 2.5 is said to be support for the inter-agent communication specification standardized in MCP v2.0. The simultaneous achievement of improved tool-calling accuracy and stable long-context processing has expanded the model's use cases from a simple code completion tool to a "foundation for autonomous repair agents."
From an EU regulatory perspective, demand for open-weight models that can be operated within European territory without transmitting data externally has surged since the EU AI Act came into force. This is one of the structural factors driving Mistral to become the default choice for European enterprises.
Among researchers, surpassing 60% on SWE-bench Verified is seen as indicating a level at which an agent "can autonomously repair a meaningful proportion of existing bugs that would otherwise be unmanageable." This is close to the critical point at which the transition from a supplementary tool that assists developers to an agent that autonomously generates pull requests becomes realistic.
Configurations in which Codestral 2.5 runs as a local node on an MCP v2.0-compliant orchestrator have already been attempted by multiple developers since the model's release. The ability to operate agents entirely within a closed network is likely to be especially valued in use cases spanning finance, healthcare, and the public sector.
The input price of $0.28 per million tokens is approximately 40% lower than GPT-4o's current input price (around $0.47). Since code generation is token-intensive and monthly costs can accumulate quickly, unit pricing becomes a practical selection criterion when integrating into large-scale CI/CD workflows.
Mistral's official blog contains no explicit mention of Japanese language support, and quality in codebases that mix Japanese text remains to be verified. For domestic Japanese enterprises making adoption decisions, this is seen as a branching point that requires measured real-world data.
We do not believe it is coincidental that this release came immediately after the GA of GitHub Copilot Agent Mode and the beta of the Anthropic Claude Agent SDK. The coding agent market has begun to clearly split along the axis of "API + cloud dependency" versus "open-weight + on-premises."
The dynamic of proprietary players leading on scores while open-weight players compete on data sovereignty and pricing has persisted since last year — but with SWE-bench scores now reaching the 60% range, that gap is entering a phase where it becomes visible to enterprise decision-makers as a difference in actual experience. The clearer the cost-security-accuracy tradeoff becomes, the more unavoidable the question of "which side to lock into" will be in internal AI coding infrastructure selection meetings.
For Japanese system integrators and software houses, whether they can build differentiated services leveraging open-weight models is a directly relevant question for their order pipelines next year.
Codestral 2.5 is not merely a benchmark story — it is an implementation-level development that influences the architectural decision of "at which layer to build autonomous coding." As the MCP v2.0 standard spreads, the choice of which model to use as an agent foundation may converge over the coming months. The next focal points are how quickly the Meta Llama and Qwen lines catch up to this score range, and how Mistral structures its commercial licensing framework going forward.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
