Claude Opus 4.5 leads SWE‑bench in 7 of 8 languages, 15% ahead of Sonnet 4.5
In the last sprint our codebase jumped from Java to Python, then a quick fix in JavaScript, and even a tiny module in Rust. So when a new AI coding model says it’s “the best,” I want to see it handle that kind of mix, not just ace a single-language test. Claude Opus 4.5 was run through SWE-bench’s multilingual suite, which throws at the model eight different language challenges.
The numbers look interesting: Opus 4.5 beats Sonnet 4.5 in most of those languages, and the gap widens on the harder polyglot tasks where the model has to flip contexts on the fly. If you need a coding buddy that can keep up with the messy, multi-language reality of modern dev work, those results are worth a look.
---
On the SWE-bench Multilingual benchmark Opus 4.5 tops the chart in 7 of 8 languages, often pulling about 10-15% ahead of Sonnet 4.5 in Java and Python. In the Aider Polyglot test it’s roughly 10.6% better at cracking tough, cross-language problems. The Vending-Bench (Long-t…) data follows the same trend.
Multilingual Coding: On SWE-bench Multilingual, Opus 4.5 leads in 7 of 8 languages 7, often scoring ~10-15% higher than Sonnet 4.5 in languages like Java and Python. Aider Polyglot: Opus 4.5 is 10.6% better than Sonnet 4.5 at solving tough coding problems in multiple languages. Vending-Bench (Long-term Planning): Opus 4.5 earns 29% more reward than Sonnet 4.5 in a long- horizon planning task, showing much better goal-directed behavior.
Opus 4.5 has a clear lead in software engineering tasks for its competitors, and even other Anthropic models. To see how well it stacks against its contemporaries on a variety of benchmarks the following visual would assist: The heavy reliance of Anthropic on software engineering and agent tasks might not be welcomed under most contexts. But what it offers AI coding is hard to look past.
One thing that sets Claude Opus 4.5 apart isn't just how well it codes, but how reliably it behaves when the stakes rise. Anthropic's internal evaluations point to Opus 4.5 as their most robustly aligned model so far, and likely the best-aligned frontier model available today. It shows a sharp drop in "concerning behavior," the kind that includes cooperating with risky user intent or drifting into actions no one asked for.
And when it comes to prompt injection, the kind of deceptive attacks that try to hijack a model with hidden instructions, Opus 4.5 stands out even more.
The numbers look impressive, but they don’t tell the whole story. Claude Opus 4.5 tops the SWE-bench multilingual results in seven of eight languages and usually pulls a ten-to-fifteen-percent edge over Sonnet 4.5 in familiar stacks like Java and Python. On the Aider Polyglot challenge the model even shows a 10.6 % advantage on tough, cross-language coding tasks.
Still, the article only shares those benchmarks - it’s unclear whether the gains carry over to everyday developer workflows or messier, less-structured code bases. Anthropic has been quiet while Gemini 3 Pro, ChatGPT 5.1 and SAM3 have been flooding the market. Opus 4.5 is billed as the most capable member of the Claude 4.5 family, promising “maximum c” - a phrase that stops short in the source.
The Vending-Bench results get a brief mention but no hard numbers, making that claim hard to verify. So the claim of being the best AI coding model rests on a narrow set of tests. Until we see more diverse evaluations, the real-world impact of these improvements remains uncertain.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI OpenAI says AI saves knowledge workers 40‑80 minutes; use yields five‑fold gains
LLMs & Generative AI Grok Chat: AI for debugging, building, testing web apps with voice and images
LLMs & Generative AI Samsung adds Vision AI Companion, an AI Bixby, to TVs for real‑time queries
Related Topic Anthropic finds strict anti-hacking prompts increase AI sabotage and lying
Research & Benchmarks
Related Topic Tool Extracts Detailed Claude Code Transcripts via Reverse‑Engineered API
Open Source
Common Questions Answered
How does Claude Opus 4.5 perform on the SWE‑bench multilingual suite compared to Sonnet 4.5?
Claude Opus 4.5 leads in seven of eight languages on the SWE‑bench multilingual suite, often scoring about 10‑15% higher than Sonnet 4.5 in popular languages such as Java and Python. This consistent advantage demonstrates its stronger multilingual coding capabilities across diverse syntax stacks.
What advantage does Opus 4.5 show on the Aider Polyglot challenge?
On the Aider Polyglot challenge, Opus 4.5 is 10.6% better than Sonnet 4.5 at solving tough coding problems that span multiple languages. The result highlights Opus 4.5’s ability to handle cross‑language tasks more effectively than its competitor.
Why is the Vending‑Bench long‑horizon planning task significant for evaluating Opus 4.5?
The Vending‑Bench task measures a model’s goal‑directed behavior over extended planning horizons, and Opus 4.5 earns 29% more reward than Sonnet 4.5 in this scenario. This large margin indicates superior long‑term reasoning and planning abilities in software engineering contexts.
Does the article provide evidence that Opus 4.5’s benchmark gains translate to real‑world developer workflows?
No, the article only reports benchmark numbers such as the SWE‑bench multilingual lead and Aider Polyglot advantage, without presenting data on everyday developer productivity or workflow integration. Consequently, it remains unclear whether the observed gains will manifest in broader, practical coding environments.