Opus 4.7 — the worst release in Anthropic's history?

Anthropic released a migration guide, which, of course, no one will read. Too bad. Here are three changes that will flip your workflow right now:

1. The budget_tokens parameter now returns a 400 error.

2. The new tokenizer takes 35% more tokens for the same text.

3. Thinking tokens are now hidden by default.

Let’s break down this mess point by point.

1. The budget_tokens parameter is broken

If your code specifies something like:
thinking={"type": "enabled", "budget_tokens": 32000}
...then on Opus 4.7, you will get a 400 Bad Request error. No warnings, no “deprecated” label. Just an error.

For those using Opus on an industrial scale, this is a disaster. You can no longer strictly limit the budget for “thinking” to control costs.

In exchange, they are pushing this on us:
thinking={"type": "adaptive"}
output_config={"effort": "xhigh"}

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

By default, in Opus 4.7, adaptive thinking is disabled. The model will run without deep contemplation. Be very careful when renaming the model from 4.6 to 4.7 in your config — by default, it will perform much worse.

2. The tokenizer now consumes more

The price per token is the same, the context window of 1 million tokens is the same (no, it hasn’t been increased). But there’s a catch.

The new tokenizer uses 1.35 times more tokens for the same text.

Plus, there are a ton of complaints that the model has started to “forget” the beginning of the dialogue faster (context rot). So essentially, the price has gone up, and the benefit for us is zero. Don’t fall for their internal benchmarks of the “I swear on my mother” variety.

What this means for you:

• All your hardcoded context limits are now incorrect.

• Any client-side token estimation is now lying.

• Your API bill will increase with the same prompts.

3. Thinking tokens are still hidden

This is the most insidious part.
In Opus 4.6, thinking blocks were displayed by default as “summarised”.
In Opus 4.7, the default is “omitted”.

Thinking thinking blocks in the response look empty, but you still pay full price for them.

Quote from Anthropic: “You are still charged tokens for thoughts. Omitting text reduces latency, but not cost.”

Your bill will contain tokens that you can't even see.

Working with long context has just collapsed

In the MRCR v2 tests (context understanding of 1 million tokens):

• Opus 4.6: 78.3%

• Opus 4.7: 32.2%

This is a 46-point drop in the benchmark published by Anthropic themselves.

Boris Chernoy is trying to justify this, but developers on X (formerly Twitter) have already mocked him. This is not progress; it is regression.

What developers say "in the field"

The feedback is just brutal.

Opus 4.7 hallucinates out of nowhere:

• Makes up non-existent colleagues (who the hell is Anton?!).

• Admits to spouting complete nonsense.

• Simply ignores user preference settings.

It feels like the model wants to finish the workday early and produce some kind of result.

“Handout” with limits

In the wake of the criticism, Anthropic announced a “permanent increase in request limits” (rate limits).

I am feeling skeptical. There are no specific numbers. But even if they raised the limits by 1.35 times — it simply compensates for the fact that the new model consumes 35% more tokens. There’s nothing to be happy about here.

What to do right now?

If you have 5 minutes:
Run a grep on the code for budget_tokens. Replace it with effort: xhigh for coding tasks (also set to maximum in Claude Code).

If you have 15 minutes:
Carefully read the release notes and migration guide, otherwise your API bills at the end of the month will be very unpleasantly surprising.