Claude Opus 4.7: What Changed, and Should You Care?

Three categories of improvement, in descending order of how much you'll notice.

Coding got measurably better. Per Anthropic's release notes, CursorBench (a coding benchmark scored on real-world tasks inside the Cursor IDE) moved from 58% on 4.6 to 70% on 4.7. Rakuten's internal SWE-Bench saw 4.7 resolve three times as many production tasks as 4.6. CodeRabbit's automated code-review benchmark logged a 10%+ recall improvement. These are the only numbers in the release that genuinely matter for daily use; the model is noticeably better at the kinds of long, multi-file tasks that an AI coding assistant gets thrown.

Vision got bigger. 4.7 accepts images up to roughly 3.75 megapixels (the long edge can be 2,576 pixels), more than three times the previous limit. That's the difference between "I can paste a screenshot" and "I can paste a full-resolution Figma export". Useful if you regularly hand the model design comps; irrelevant if you don't.

Finance-agent and reasoning got the rest. 4.7 hits state-of-the-art on Anthropic's Finance Agent eval and improves on a handful of mixed reasoning benchmarks. The practical impact for non-specialists is small; it tells you the model is broadly stronger, but it won't change a workflow that wasn't doing financial analysis to begin with.

The headline is that 4.7 is no more expensive than 4.6: $5 per million input tokens, $25 per million output tokens. In sterling and at typical coding-session sizes, that translates to numbers that surprise people in both directions.

A casual chat session (a few prompts, a few hundred words of replies) costs pennies. A typical Claude Code, Anthropic's terminal-based agent coding assistant for software engineering, session that loads your CLAUDE.md, reads ten files, writes two, runs a few tests, and produces a 200-line edit will spend maybe 50,000 input tokens and 5,000 output tokens. That's about 30p per session at current exchange rates. Run ten sessions a day and you're at £3 a day, or roughly £60 a month.

The Claude Pro subscription (£15-ish a month at UK pricing) bundles a generous quota that covers far more than that. For a hobbyist, the subscription wins on cost as long as you're not running batch jobs through the API. For an API user processing documents or running production agents, the per-token cost dominates and the maths gets more interesting.

Compared to running a similar-quality open-source model locally on a beefy Mac: locally it's free per call, but you're amortising several thousand pounds of hardware and you're not getting 4.7-class output. For most people who write code professionally, the cost-per-month of Claude Pro is cheaper than an hour of their own time.

Three workflow features ship alongside the model, all relevant if you live in Claude Code day-to-day.

If you're already on a Claude paid plan: yes, automatically. The model is the default in Claude.ai, the API, and Claude Code. There's no opt-in step.

If you're on the free tier: probably not just for this release. 4.7's gains are concentrated on long, multi-step coding tasks where the free tier's rate limits already get in the way. The model is technically better; you'll likely not have enough headroom to feel the difference. If you start hitting limits often, the £15 Pro tier becomes the obvious next step, but "because 4.7 launched" isn't by itself a reason.

If you're paying for ChatGPT Plus or Gemini Advanced instead: the comparison is genuinely close in 2026. Claude leads on coding agents and long-context work. GPT-5 is stronger on raw reasoning benchmarks and arguably better at creative writing. Gemini 3 has the longest context window in mainstream use and tighter Google Workspace integration. Pick by which workflow you're actually trying to support, not by which model topped the latest benchmark.

Anthropic ships three tiers: Opus (most capable), Sonnet (balanced), Haiku (fastest and cheapest). Opus 4.7's launch doesn't change when each is the right choice.

Sonnet 4.6 is the default for almost everything that isn't a hard agentic task. Drafting prose, summarising documents, answering questions about a codebase you've already explained, light refactors: Sonnet handles them at a fraction of the cost and latency. The quality gap to Opus is small enough on bounded tasks that most users would not notice.

Haiku 4.5 is for batch jobs and high-volume API workloads where the per-token cost matters and the task is well-defined (classification, extraction, simple summarisation). Anything where you're running thousands or millions of calls and each one is small.

Opus 4.7 earns its keep on long-running coding tasks, multi-step agent flows, and analysis that needs to keep many things in mind at once. If your typical session is "write me a quick helper function", Sonnet does it for less. If it's "refactor this auth flow across twelve files and update all the tests", Opus is the right tool.