If you are still agonizing over whether to buy API credits for Claude Opus 4.6, this article might save you some money.
Anthropic released Claude Sonnet 4.6 in February at the exact same price as Sonnet 4.5 — $3 per million input tokens and $15 per million output tokens — yet its benchmark scores directly approach those of the five-times-more-expensive Opus 4.6.
By the numbers
Here is the most straightforward comparison:
| Benchmark | Sonnet 4.6 | Opus 4.6 | Gap |
|---|---|---|---|
| SWE-bench Verified (coding) | 79.6% | 80.8% | -1.2% |
| OSWorld-Verified (computer use) | 72.5% | 72.7% | -0.2% |
| Math (MATH-500) | 89% | Not disclosed | — |
| GPQA Diamond (scientific reasoning) | 74.1% | 91.3% | -17.2% |
On coding and computer use — the tasks most enterprises rely on Claude for — Sonnet 4.6 is already nearly on par with Opus 4.6. The real gap appears in more academic tasks like scientific reasoning, where Opus 4.6's 91.3% versus Sonnet's 74.1% is a gap you cannot ignore.
But if your use case is primarily writing code, controlling a browser, or processing long documents, paying five times more for a 1.2 percentage point improvement in coding is not, by any calculation, a smart move.
How steep the computer-use curve is
Anthropic published data showing the evolution of OSWorld-Verified scores — a standard test measuring an AI's ability to autonomously control a computer — over the past 16 months:
- Sonnet 3.5: 14.9%
- Sonnet 3.5 v2: 28.0%
- Sonnet 3.6: 42.2%
- Sonnet 4.5: 61.4%
- Sonnet 4.6: 72.5%
From 14.9% to 72.5% in a year and a half — this curve is steeper than most people realize. In insurance workflow automation tests, Sonnet 4.6 achieved 94% accuracy, covering tasks like manipulating complex Excel spreadsheets, filling out multi-step web forms, and interacting with legacy desktop applications.
A big leap in math
Sonnet 4.5 scored 62 on math; Sonnet 4.6 jumped directly to 89. That span is the largest single improvement within the same generation. Anthropic has not disclosed the specific reasons, but given the overall improvement in reasoning capability, it is likely tied to better chain-of-thought quality.
User preference data
Anthropic ran a series of internal Claude Code tests where developers compared output quality without knowing which model was behind each response:
- Sonnet 4.6 vs Sonnet 4.5: 70% of users preferred Sonnet 4.6
- Sonnet 4.6 vs Opus 4.5: 59% of users preferred Sonnet 4.6
The second comparison is more interesting. Opus 4.5 is significantly more expensive than Sonnet 4.6, yet in subjective evaluations of real coding output, Sonnet 4.6 was actually more popular. Reasons cited include more accurate instruction following, fewer hallucinations, and less over-engineering — a persistent issue with the Opus line, which tends to overcomplicate simple requirements.
1M context is not a gimmick
Sonnet 4.6 also comes with a 1-million-token context window (beta), requiring no additional headers. Requests exceeding 200k tokens automatically use this channel, and pricing follows the standard rate.
What does 1 million tokens roughly equate to? The entire codebase of a mid-sized enterprise, or dozens of research papers, can be processed in a single request. Scenarios that previously required RAG chunking can now simply be thrown in for the model to find relevant connections on its own.
Which one should you choose
Choose Sonnet 4.6 if:
- Your primary tasks are writing code, controlling a browser, or processing long documents
- You have high volume and are price-sensitive
- You do not need PhD-level scientific reasoning
Choose Opus 4.6 if:
- You need intensive scientific reasoning or academic tasks
- Your context requirements are extremely complex
- Price is no object and you want the absolute best result
For most developers and enterprises, Sonnet 4.6 is currently the most cost-effective entry point to Claude. Opus 4.6 is priced at $15/$75, while Sonnet 4.6 is $3/$15 — a fivefold difference for the same coding capability. The math is not hard.
Sources: Introducing Claude Sonnet 4.6 (Anthropic official); CocoLoop, Claude Sonnet 4.6: 79.6% SWE-bench at $3/MTok — Complete Guide (NxCode); Claude Sonnet 4.6: 1M context and stronger computer use (Gend.co)