Claude Opus 5.5 audits 200,000 lines in under 3 hours

Claude Opus 5.5 was announced by Anthropic on Tuesday, the 22nd. The model opens the Claude 5.5 family. The company says its performance matches Fable 5.1 in most tasks.

The most talked-about difference, however, shows up in the bill. The cost drops 40% compared to Opus 5.

In addition, it comes included in the subscription. Subscribers use the model without having to buy credits separately.

Claude Opus 5.5 saves money in two different ways

The cost reduction combines two factors. First, the price per token is lower than that of Opus 5.

Then comes the more interesting point. The model consumes fewer tokens to complete each task.

In other words, the savings happen in both the rate and the consumption at the same time. Therefore, the 40% drop shows up in real usage, not just in the price table.

The audit case that illustrates the gain in long tasks

The company highlights performance in lengthy jobs. Migrations and code audits fall into this category.

In addition, the example cited stands out. A tester audited and improved a 200,000-line codebase in under 3 hours.

For comparison, Opus 5 would take around 20 hours on the same task. It would also consume 2.5 times more tokens.

Notice what this number changes in practice. Tasks that were previously unfeasible due to cost now fit into a normal work cycle.

Claude Opus 5.5 leads the benchmarks, with two exceptions

According to Anthropic, the model outperforms Fable 5.1 and Opus 5 in all the highlighted benchmarks.

However, two tests were left out of that count. In AutomationBench and Terminal Bench Science 0.1, GPT 6 Astra came out ahead.

It's worth noting the condition of these tests. Opus 5.5 was operating in adaptive reasoning and maximum effort mode.

Text responses became more direct

In addition, the company also tweaked the writing style. The promise is about clearer and more objective messages.

According to testers, this helped during long work sessions. The most important information stands out.

In addition, the model avoids excessive jargon and phrases typical of AI agents. It also follows the writing guidelines defined by the user.

Two features come from Fable 5.1. Text responses carry an invisible watermark, and the model operates only in reasoning mode.

Price and availability on each channel

The model is already available across all Claude channels. The list includes Amazon Web Services, Google Cloud and Microsoft Azure.

In Claude Code and Claude Platform, there's also Fast mode. It promises to deliver tasks up to 2.5 times faster, at a higher cost.

The prices are as follows:

$4 per million input tokens.
$20 per million output tokens.
$0.20 per million tokens read from cache.
$5 per million tokens written to cache.

In Fast mode, input and output cost double.

The company also increased the five-hour usage limit for the Pro, Max and Team plans. Along with that, it granted a limit reset that users can save for whenever they want.

What to evaluate before migrating your workflow

Start with the cost per completed task. Price per token matters little when consumption per task changes so much.

Then, test it on real long-form work. Migration and auditing reveal much more than an isolated prompt.

Also take advantage of the cache. The difference between cache reads and writes matters in workflows with repeated context.

In addition, measure Fast mode carefully. Doubling the price only pays off where response time is holding someone back.

Finally, review your writing guidelines. The model now follows them more strictly, and old instructions may produce different output.

What's coming in the next few weeks

Anthropic has promised two upcoming releases. These are Claude Sonnet 5.5 and Claude Haiku 5.5, the latter stuck at version 4.5 since 2025.

Meanwhile, it's worth following independent evaluations of token consumption. That number determines whether the 40% savings hold true in your repository.

Follow our profile on Instagram!