Anthropic just shipped Claude Opus 4.7, and unlike the usual model-update noise, this one actually changes the calculus for power users. Here's what's new, what's hype, and whether it's worth switching.
What's genuinely new
Opus 4.7 ships with a one-million-token context window — the same headline capacity as the rest of the current Opus tier, and the thing that changes day-to-day work most. A million tokens is roughly a large codebase or a stack of long documents held in view at once, which turns "chunk it and hope" into "just paste it."
The more interesting change is how reasoning is controlled. Earlier models took a budget_tokens figure — you guessed a thinking allowance up front. That is gone. Opus 4.7 uses adaptive thinking, where the model decides how much to reason based on the problem, and you steer overall spend with a separate effort setting rather than a token count.
Effort has five levels, and 4.7 is where xhigh was introduced — a tier sitting between high and max. For most coding and agentic work it is the sweet spot: high for general use, xhigh when the task is genuinely hard, max when correctness matters more than cost, and low for cheap mechanical sub-tasks.
"The shift from guessing a thinking budget to setting an effort level is the kind of change that sounds minor and rewrites how you tune a system."
Agentic workflows, finally usable
The practical headline is agent reliability over long tool-use chains. Two features do most of that work.
Compaction handles conversations that outgrow even a million tokens: the API summarises earlier context server-side as you approach a threshold, rather than failing or silently dropping history. Context editing is the sharper instrument — it clears stale tool results or old thinking blocks outright instead of summarising them, which matters in agents that make hundreds of tool calls whose intermediate output stops being relevant.
Together they mean a long-running agent degrades gracefully instead of hitting a wall, which is the difference between a demo and something you leave running.
Coding: the new default for many devs
Pair-programming feels different. Opus 4.7 understands large repos better, suggests fewer hallucinated APIs, and is markedly better at refactors. If you've been on GPT-5 for code, this is the first Claude release that genuinely deserves a side-by-side trial.
Pricing and access
On the first-party Anthropic API, Opus 4.7 runs $5 per million input tokens and $25 per million output — the standard Opus-tier rate. It is reachable through Claude.ai's paid plans, the Anthropic API, Amazon Bedrock, and Google Vertex AI, though feature availability varies by platform: some capabilities are first-party API only, so check before assuming parity on a cloud provider.
Two things commonly break when migrating older code to this generation, both worth knowing before you switch:
budget_tokens is rejected. Sending it returns a 400. Use adaptive thinking plus an effort level instead.- Assistant prefills are rejected. The old trick of seeding the assistant's turn to force a format returns a 400 too. Use structured outputs or a system instruction.
Should you switch?
- Heavy coders: Try it for a week. Most will stay.
- Writers and analysts: Marginal gain. Stick with what you have unless context length matters.
- Agent builders: Switch now. The reliability jump is the biggest in any 2026 release.
FAQ
How does it compare to newer Claude models?
Anthropic has shipped further releases since — the Opus line continued past 4.7, and there are Sonnet and Haiku tiers at different price and speed points. Opus 4.7 remains a capable model; whether it is the right one depends on your cost and latency budget more than on capability alone.
Is a million tokens of context actually usable?
Usable, yes; free, no. You pay for every token you send, so filling the window on every request gets expensive fast. Prompt caching is what makes large stable context economical — cache the unchanging prefix and you pay a fraction for it on repeat requests.
Do I need to change my prompts?
Often, yes, and this is the step people skip. Prompts written for older models tend to over-instruct — long "think step by step" scaffolding and rigid formatting rules that a reasoning model does not need and can be actively hindered by. Migrating the model without revisiting the prompt leaves quality on the table.
The bottom line
Opus 4.7 isn't a revolution, but it moves long-running agents closer to boring — in the best way. The million-token window gets the attention; the reasoning-control changes are what you actually feel three weeks in.
If you are comparing across providers rather than across Claude versions, ChatGPT vs Claude vs Gemini is the more useful read.