Claude Sonnet 5.5: Anthropic speeds up and cuts the cost of its model for everyday work

Anthropic launched Claude Sonnet 5.5 today, September 28, 2026. Just six days after introducing Opus 5.5, we already have the second piece of the new family.

And there's something that caught my attention from the first numbers: Anthropic didn't try to sell Sonnet 5.5 simply as a cheaper Opus.

The pitch is far more practical. They want a fast model for the tasks we do every day, capable enough to get close to Opus in several scenarios, and with a considerably lower real cost per job. Anthropic

The 30% savings needs an explanation

There's an important detail here because the announcement can be misleading.

Anthropic did not cut token prices by 30%.

Sonnet 5.5 keeps Sonnet 5's rates:

The savings Anthropic is announcing come from somewhere else: Sonnet 5.5 typically needs fewer tokens to complete the same job. According to their tests, this can mean up to 30% less cost per task. Anthropic

To me, this metric is far more interesting.

When we use a model through an API, the same cost per token doesn't mean two models cost the same to do the job. One may need several iterations, more context, and a huge output to solve something another finishes much sooner.

It's like charging the same hourly rate to two technicians when one finishes the job in three hours and the other in five.

It's also quite a bit faster

Anthropic reports that Sonnet 5.5 generates output more than 30% faster than Sonnet 5, making it the fastest Sonnet they've released so far. Anthropic

This seems like a secondary detail until you work with AI all day.

On an isolated question, waiting a few extra seconds makes practically no difference. When you're coding and chaining dozens or hundreds of interactions with an agent, latency starts to get in the way.

That's when I do want speed.

The leap in coding is huge

This part was what first made me stop at the table.

On Terminal-Bench 4.0, a coding test using agents working in a terminal, the results published by Anthropic are:

• Sonnet 5.5: 70.6%
• Sonnet 5: 10.3%
• Opus 5.5: 66.4%

Yes, the gap between generations is that big in this evaluation.

But I also wouldn't take a single benchmark as definitive proof that Sonnet is better than Opus. Anthropic itself warns that Opus 5.5 remains clearly superior on complex, open-ended work that requires maintaining judgment over long periods. Anthropic

That matters.

Benchmarks are useful for comparing specific behaviors under specific conditions. They're not a universal score for intelligence.

When we look at more tests, a more interesting picture emerges

The official results published by Anthropic include:

This is where we start to see where Sonnet 5.5 fits.

On GDPval-AA it's just two points behind Opus 5.5. On OSWorld the difference is 1.7 points. On CursorBench it's 2.3.

Sonnet no longer looks like just the budget sibling you settle for when you don't want to pay for Opus. On several tasks it's getting very close to the top-tier model. Anthropic

So what's left for Opus 5.5?

The hard work.

Anthropic lays out a fairly clear division.

Sonnet 5.5 is aimed at well-defined everyday tasks: fixing bugs, coding, working with documents, creating presentations and spreadsheets, analyzing images, and collaborating quickly.

Opus 5.5 is designed for complex, open-ended problems that require maintaining judgment over long stretches. Anthropic

As a developer, this split makes sense to me.

I don't need to use the most powerful model available for absolutely everything. If I'm fixing an isolated bug, reviewing a function, or generating documentation, I want speed and accuracy at a reasonable cost.

If I have an architecture problem, need to analyze many dependencies, or I'm delegating a task that could take quite a while, that's when it makes sense to move up to Opus.

There's another new variable: the effort level

Claude lets you adjust how much effort the model puts into the work.

In the Claude and Claude Code apps, Sonnet 5.5 uses Medium by default. On Claude Platform, the default is High.

A lower level reduces time and token consumption. A higher one lets the model work longer and review what it's doing more thoroughly. Anthropic

This opens up an option that interests me much more than always choosing between models.

We can use the same model in different ways depending on the problem.

Anthropic claims, for example, that Sonnet 5.5 on Low or Medium beats Sonnet 5's best result on some evaluations while spending roughly a tenth per task. On FrontierCode, Sonnet 5.5 with High achieves considerably better results than Sonnet 5 at a much lower cost per task. Anthropic

That kind of optimization can actually move the bill for an application running thousands of operations.

There are also important security changes

There's another detail that can go unnoticed among so many benchmarks.

Sonnet 5.5's cybersecurity capabilities increased enough that Anthropic is launching it with protections similar to those implemented in Opus 5.5.

According to the company, these measures target a small set of high-risk requests, and normal software development should remain practically unchanged. Anthropic

Anthropic also evaluated the model's behavior across roughly 1,850 automated scenarios related to alignment, deception, resistance to abuse, and actions contrary to the user's intent. The company reports results equal to or better than Sonnet 5 on most of those tests, though it acknowledges that no evaluation can catch every possible failure. Anthropic

That last part is worth keeping in mind whenever we look at any security table: they're evaluations, not guarantees.

And Haiku 5.5 is still to come

The family isn't complete yet.

Anthropic confirmed that Claude Haiku 5.5 will arrive in the coming weeks, focused on high-volume applications where speed and cost matter even more. Anthropic

So the strategy starts to become pretty clear:

• Haiku for volume and cost.
• Sonnet for most everyday work.
• Opus for complex problems.

The difference is that Sonnet got quite close to Opus in several scenarios, and that makes the decision of which model to use much more interesting.

My takeaways

For a long time we've compared models by looking first at who gets the highest percentage.

With Sonnet 5.5 I'm interested in a different metric: how much it costs to solve a complete task.

A model that charges exactly the same per token can end up being considerably cheaper if it needs less context, fewer attempts, and less output to reach a correct result.

And if it also responds 30% faster, that improvement starts to be felt directly while you work.

For those of us who use AI as a development tool and not just as a chatbot, that's probably the most important part of this launch.

Conclusion

Claude Sonnet 5.5 seems like a much more important update than its ".5" might suggest.

The leap compared to Sonnet 5 shows up in programming, professional work, computer use, visual interpretation, speed, and efficiency. At the same time, Anthropic keeps Opus 5.5 as the option for problems that require more sustained reasoning and judgment.

I especially want to test that difference by working on real projects. Benchmarks tell me where to start looking; several hours of programming with both models will tell me which one I end up using.

The technical information and full evaluations are in the official Claude Sonnet 5.5 announcement and in Anthropic's System Cards.