Claude Opus 4.7 Is Out: Better Vision, a New Effort Level, and a Tokenizer Change Worth Knowing About

Anthropic shipped Claude Opus 4.7 on April 16. It’s available in Claude products, the API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The API identifier is claude-opus-4-7.

Pricing is the same as Opus 4.6: $5 per million input tokens, $25 per million output tokens. There’s a catch in that, covered below.

What’s New

Vision: The biggest technical change is image handling. Opus 4.7 accepts images up to 2,576 pixels on the long edge, roughly 3.75 megapixels. That’s more than three times the prior limit. The improvement matters most for dense technical content: chemical structure diagrams, engineering schematics, and screenshots with small text that previously required multiple crops or just didn’t parse reliably.

Coding: Anthropic reports a 13% improvement on their 93-task internal coding benchmark, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could complete. Finance Agent and GDPval-AA benchmarks show similar gains. The model pays closer attention to instructions, which Anthropic describes as “substantially more literal interpretation.” If you have existing prompts that relied on Claude inferring what you meant rather than doing exactly what you said, they may behave differently now.

Effort levels: Opus 4.7 adds xhigh as a new effort level, sitting between high and max. Claude Code defaults to xhigh for all plans with this release.

Memory: File system-based memory now persists across sessions. Claude can write notes during a session and retrieve them in later conversations, reducing the amount of context you need to re-establish at the start of a new session on long-running projects.

Claude Code additions: The release includes /ultrareview, a command that runs dedicated multi-agent code review sessions. Auto mode, which lets Claude make tool permission decisions on its own, is now available for Max subscribers (previously Opus only). Task budgets, a public beta feature for managing token allocation across longer runs, also shipped with this release.

The Tokenizer Change

This matters if you’re running Opus at scale. Anthropic updated the tokenizer with Opus 4.7. The same input text now produces 1.0x to 1.35x more tokens depending on content, with a median around 1.1x. The per-token price is unchanged, but if your input consistently lands at the high end of that range, your effective cost per request is higher.

Anthropic notes the tokenizer change “on net” favors performance on coding evaluations, meaning the extra tokens reflect more useful reasoning. But “on net” doesn’t mean every workload benefits, and the impact is asymmetric: better outputs don’t always justify higher costs for tasks where the old model was already good enough.

Worth measuring in your own workloads before assuming it’s cost-neutral.

Availability

Opus 4.7 is available now in all Claude plan tiers, the Anthropic API, and major cloud model platforms. GitHub Copilot Pro+ users are getting it through the model picker as it rolls out.

If you use Claude Code, you’re already running it by default with the desktop redesign shipped earlier this week.

Source: Anthropic