On April 16, 2026, Anthropic dropped Claude Opus 4.7. Same day this chapter was written, which means by the time you read it some of the rough edges may have been sanded off and a few new ones added.
That is the nature of writing a book about a moving target. I am going to tell you what changed, what matters for a Delphi developer, and what I would not lose sleep over.
The headline numbers
Same price as 4.6: five dollars per million input tokens, twenty-five per million output tokens. No surprise tax for the upgrade.
The coding gains are real but not dramatic. About ten to thirteen percent better than 4.6 across the standard coding benchmarks, with SWE-bench Pro climbing from roughly 53% to 64%. Decent.
Those benchmarks are written in Python and Go, not Delphi, so take the exact percentages with the grain of salt they deserve.
To set expectations honestly: the jump from 4.6 to 4.7 is smaller than the jump from 4.5 to 4.6 was. If you were hoping for another revolution, this one is an evolution.
Alignment is noticeably better
The single biggest improvement I noticed is not on any benchmark slide – it is alignment to instructions. When you write a CLAUDE.md telling the model to never use the with statement, to prefer FreeAndNil over .Free, to skip trivial property declarations, 4.7 follows those rules a bit better.
That sounds like a small thing. It is not. Every time the model violates a project rule, you spend a turn correcting it, which costs tokens and breaks your flow. Cut those corrections in half and the upgrade pays for itself.
What is actually new in Claude Code
A few things landed alongside the model that you will run into within your first hour.
There is a new effort level called xHigh, sitting between High and Max. Claude Code now defaults to it for every plan. In practice this means the model thinks longer before it answers, which costs more tokens but produces fewer of those WTF moments.
If you are doing serious Delphi work, Max is still the only setting worth using.
A new slash command, /ultrareview, is dedicated to code review. Pro and Max subscribers get (only) three free runs of it. The output is noticeably more thorough than asking “review this file” yourself, because it pulls in more context and runs more passes internally before reporting back. Worth saving for files where you actually want a second opinion, not for routine work.
Task budgets are now in public beta. You set a token ceiling for a long-running task and Claude tries to stay under it. Useful when you tell the model to “go fix all the warnings in this project” and you want to avoid waking up to a credit card alert.
Auto-mode is now extended to Max users. If you are on the Pro plan, no change there.
There are also hints that DocX handling has improved on the model side, separate from MCP servers. I have not confirmed this end-to-end yet – if you write your books in Word like I do, it is worth testing.
The tokenizer change you should not miss
This one is buried in the release notes but it matters: the tokenizer was updated. The same Delphi source file may now consume between 1.0 and 1.35 times as many tokens as it did under 4.6. That is up to a third more, on identical input.
Combined with xHigh (or Max) thinking by default, expect your token usage to climb noticeably in the first week. Same sticker price, bigger bill.
Cost cuts are still available the same ways they always were:
- Prompt cache: up to 90% off
- Batch processing: 50% off
Neither of those helps if you are doing interactive Delphi development in Claude Code. Both help if you are building agents that run repeatedly with similar prompts.
The thing nobody warns you about
Anthropic mentions, almost in passing, that “prompts written for earlier models can sometimes now produce unexpected results.” This is real. Opus 4.7 is more literal. Where 4.6 might have inferred what you meant, 4.7 does what you said. If you have a CLAUDE.md or a personal prompt template that has been working for months, give it a re-read. Vague instructions that used to be charitably interpreted may now be followed to the letter, including the parts you did not actually mean.
For a Delphi codebase this shows up in places like “use modern patterns” – 4.6 understood you meant TThread and anonymous methods, 4.7 may ask you which patterns you have in mind. Be explicit.
A surprise outside Delphi
Worth noting for anyone who, like me, does work that touches molecular biology or bioinformatics: the model improved by over one hundred percent in that domain. Not all of you will care, but you should as this means better medicine!
What I actually want next
A better model is welcome. What I want more is a better agent and a better harness around it. Claude Code itself – the planner, the tool router, the way subagents coordinate – matters more for daily Delphi work than another five percent on a benchmark. The model is already smart enough. The agent infrastructure around it is what limits how much real work you can offload in one session.
Should you upgrade?
Yes. The model auto-upgrades in Claude Code anyway. The jump from 4.6 to 4.7 is not the kind of leap that changes what you can do. It is the kind that makes long sessions less painful, refactors less leaky, alignment more reliable, and code review more useful. If you are doing solo Delphi work with Claude Code, you will feel the difference within a few days.