Opus 4.7 Is The Worst Release Anthropic Has Ever Shipped.

Boris Cherny posted a thread today on how to get the most out of the newly released Opus 4.7 (it's a great insight).

6 tips on how to improve usage, no mention of the breaking changes that shipped along with it.

Here is what he didn't mention.

Not a Medium member? Read for free by clicking here.

3 Breaking Changes To Your Workflow

Anthropic just released Opus 4.7, they also released a migration guide (which most people won't read).

There are 3 changes you need to know about right now:

budget_tokens now returns a 400 error.
New tokenizer is charging 35% more tokens.
Thinking tokens are hidden by default.

Let's look at each one.

1. budget_tokens Is Broken

If your code has this:

thinking={"type": "enabled", "budget_tokens": 32000}

thinking={"type": "enabled", "budget_tokens": 32000}

On Opus 4.7 that now returns 400. No deprecation notice it simply returns 400.

This is a massive deal when using Opus at scale. Thinking budgets (used as to help with cost) gone.

This is the replacement.

thinking={"type": "adaptive"}
output_config={"effort": "xhigh"}

thinking={"type": "adaptive"}
output_config={"effort": "xhigh"}

Effort levels: low, medium, high, xhigh (new), max.

Adaptive thinking is turned off by default on Opus 4.7.

The model will run without thinking, be very careful when upgrading your model name from 4.6 to 4.7, it's a lot worse now by default.

2. Tokenizer Now Uses More Tokens

At the same price per token, same 1M context window (nope that didn't increase).

However the new tokenizer uses up 1.35x more tokens for the same text.

A lot of reports suggesting, it's affected by context rot more too.

So effectively the price is gone up but not much has improved from our side. Don't be fooled by the trust me bro benchmarks.

Another 3 massive things that matter now are:

Any hardcoded context budgets are now wrong.
Any client-side token estimation is now wrong.
Your API bill is higher for the same prompt.

3. Thinking Tokens Are Still Hidden

This one is still nasty, I did a recent post on this.

Opus 4.6, until recently thinking content defaults to display "summarised".

Opus 4.7, it's now defaulted to "omitted". Thinking blocks appear empty in the response, but you still fully pay for them.

"You're still charged for the full thinking tokens. Omitting reduces latency, not cost."

That is straight from Anthropic.

Your bill will include tokens you cannot see anymore.

Long Context Retrieval Just Fell Off A Cliff

On the MRCR v2 benchmarks at 1M tokens:

Opus 4.6: 78.3%
Opus 4.7: 32.2%

That's a 46 point regression on a benchmark Anthropic themselves have published.

Here was Boris's response, however I don't completely buy it and I personally think it's misleading.

What Developers Are Actually Reporting

The overall response has been brutal.

Even Opus 4.7 is confessing to making up garbage.

Other patterns coming up in real usage:

Hallucinating colleagues and random people

Let's call it a day.

Configured preferences being ignored.

Anthropic Raised The Rate Limits (Apparently)

Following backlash, Anthropic announced "permanent increase in rate limits".

However I am very skeptical around this. As there are no numbers or percentages as of yet.

But let's say they increase rate limits 1.0–1.35x well that now matches the new thinking usage (So Enjoy? I think not).

What Boris Said (And Didn't)

Boris's thread is worth reading:

Auto mode for longer unsupervised runs.
/fewer-permission-prompts skill to cut down on approval loops.
Recaps for returning to long sessions.
Focus mode to hide intermediate work.
Effort tuning (start at xhigh for coding).
Give Claude a way to verify its work.

However don't expect to upgrade with Boris's workflow for it to just work.

Start By Doing This

If you have __5 minutes: Grep your codebase for budget_tokens. Set effort to xhigh for coding work (In Claude Code set this to the highest too).

If you have __15 minutes: Read through Anthropic's release note and migration guide.

NOTE

I love using Claude Code and Claude, however these recent moves are really making me nervous and I'm not the only one. Everyone I talk to recently has noticed a recent big dip in performance, while we are told to jump on the hype train.

I am not affiliated with Anthropic or Claude. All thoughts are my own.

One more thing

If you want to follow along with how I'm navigating all of this, I write about it on my Substack. Real workflows.

I'm also building a course for engineers who want to be better at shipping with AI. 10 years of building distilled into something I wish I had when I started. You can join the waitlist.

Contents