Sonnet 5.5 vs Opus 5.5 in Claude Code: with Jev Router the same web app costs 61% less

With Sonnet 5.5, the result of my Jev Router test flipped. Building the same web app cost 61% less than with Opus 5.5, while in the first experiment, with Sonnet 5, the same web app had cost me more than with Opus. I changed one thing only: the Sonnet model.

What was the first Jev Router experiment?

While vibe coding, I used Jev to choose which model Claude Code should use for each task, Opus or Sonnet. Jev is the decision model from TypeSafe AI: it reads the task and returns a choice. Jev Router is the open source project that puts Jev between Claude Code and the models.

In the first test Jev almost always suggested Sonnet 5, the cheaper model: in 14 prompts out of 15. The problem is that, in the end, building the web app cost me more with Sonnet 5 than with Opus 5.5: $2.04 against $1.27 on average, 61% more. A model that is cheaper per token does not guarantee cheaper work.

What changed with Sonnet 5.5?

On the same day as the first test, Anthropic released Sonnet 5.5. A viewer asked me in the comments to repeat the experiment, so I ran it again from scratch:

• same prompts;
• same web app;
• same task;
• one difference: Sonnet 5.5 instead of Sonnet 5.

Jev kept suggesting Sonnet. This time, though, the cost of the runs with Sonnet 5.5 was 61% lower than with Opus 5.5: $0.63 against $1.61 on average, over three runs per setup. These costs are calculated at API list prices: on a Claude Pro or Max plan they are quota used, not money spent.

Why does Sonnet 5.5 cost less than Sonnet 5 if the token price is the same?

Sonnet 5 and Sonnet 5.5 have the same base token price: $2 per million input tokens and $10 per million output tokens, according to Anthropic’s pricing page. The savings come from somewhere else.

Sonnet 5.5 needs far fewer calls to finish the same work. In my test:

• Sonnet 5 made 90 calls on average;
• Sonnet 5.5 made 34.

That is where most of the savings over Sonnet 5 come from: from $2.04 down to $0.63, 69% less for the same work. Anthropic says the same in the Sonnet 5.5 announcement: Sonnet 5.5 uses fewer tokens per task and batches more tool calls together, so it finishes the work in fewer steps. The same announcement also points to one more price difference: cache reads cost half as much on Sonnet 5.5.

Why does Sonnet 5.5 cost less than Opus 5.5 too?

Against Opus 5.5, the token price adds to the number of calls. In the second test Opus 5.5 made 46 calls on average, Sonnet 5.5 made 34. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, twice as much as Sonnet 5.5. Fewer steps plus cheaper tokens add up to the 61% savings I measured.

Did Jev Router get it wrong in the first experiment?

No. Jev Router behaved the same way in both experiments: it always suggested the cheaper model. This time, the cheaper model actually turned out to be cheaper.

The lesson for anyone vibe coding is practical: measure a model’s cost on finished work, because the price list alone does not tell you enough. A router that picks the model by price per token only works well if that model completes the task without multiplying the calls. It is the same principle I applied when separating planning from execution with Grok 4.5 in Cursor.

Is Sonnet 5.5 with Jev Router worth it for web apps?

In my test, yes: same web app, 61% lower cost than Opus 5.5. Every new model can flip the result again, though, as just happened. If you build your web apps by vibe coding, rerun the measurement on your own tasks every time a model ships: it is the fastest way to know what you are really paying.