Claude Opus 4.7: 4 things that moved the needle in April
Anthropic's most capable public model reached general availability on April 16, landing simultaneously on Amazon Bedrock, Google Vertex and Microsoft Foundry.
Is it worth switching from Sonnet 4.6 to Opus 4.7 for serious coding? It is, but only if you are on the API or the Pro plan.
Anthropic quietly released Opus 4.7 on April 16 through its own channel and Amazon Bedrock. Within the first few weeks, developers were already reporting that the model handled tasks that had previously needed multiple attempts.
Opus is the more expensive, smarter sibling of Sonnet in Anthropic’s lineup. Three times slower on a typical query, but roughly five times more expensive per output token (the basic unit used to bill API usage, about four characters or half a word).
For normal back-and-forth in an app, it is not worth it. For a multi-hour project that needs precision, it is.
1. Stronger SWE-Bench results
SWE-Bench measures how well a model resolves real bugs filed in actual software projects. It has become the standard benchmark for AI coding ability.
Opus 4.7 set a new public record, clearly above Sonnet 4.6, which was already strong.
Development teams report that Opus 4.7 will now check its own output before declaring “done.” That is a behavior shift you feel immediately.
A concrete example: working with a large repository (the storage for a project’s source code) where the model needs to find every place a function is used, edit them all, and track the changes. The old Opus would routinely lose track of what it had already changed. The new one does not.
2. High-resolution image analysis
Opus 4.7 sees images at much higher resolution than the previous version. In practice:
- You can send it a photo of a complex chart and the model will read every label
- You can send it a screenshot of an old site and ask for a table of changes
- You can send it a photo of a magazine page and the model will pick up small text and layout
The previous version would solve that task halfway. For design work, translation or building presentations, the difference is significant.
3. Pricing stays the same
Five dollars (about 588 dinars) per million input tokens (the text you send to the model).
Twenty-five dollars (about 2,940 dinars) per million output tokens (the text the model sends back).
That matches the Opus 4.6 pricing from last November. Since Opus typically replaces the previous version automatically, Pro plan users do not have to change anything.
For claude.ai subscribers paying 20 dollars per month (about 2,350 dinars), Opus 4.7 comes with slightly higher rate limits. Anthropic adjusts these periodically as it scales infrastructure.
4. Safety against cyber misuse
A less visible but important change. Opus 4.7 includes a built-in system that recognizes and blocks prompts that would lead to attacks on third-party infrastructure.
Cybersecurity professionals who legitimately use the model for pentesting (planned testing of your own system’s security) or vulnerability research now have to go through verification under the new Cyber Verification Program to get unrestricted access.
In practice, a normal dev team will not notice this. For someone building a vulnerability scanner, there is now a clearer route.
What it means in practice
If you have been choosing between Opus 4.6 and Sonnet 4.6, Opus 4.7 widens the gap. For most work, Sonnet remains the right choice on speed and price. For the hardest tasks that go beyond trivial scripting, Opus 4.7 now solves them in a single pass more reliably.
Anthropic has publicly acknowledged that its internal version, codenamed Mythos, is stronger still. The company is not releasing it until stronger safeguards exist. Until then, Opus 4.7 is the strongest model available to a regular user.
What users in Serbia need to know
Serbian payment cards work normally on claude.ai and through the API, with the same functionality as in the US. There are no geographic restrictions on the model.
The Mac app is a free download. Subscriptions are managed through the site.
There is no official Windows or Linux desktop app, but the web version runs in any modern browser.
Do you use Claude for work, and how often do you switch up to Opus from Sonnet?
