Is Claude Sonnet 5 Actually Worth It?
You will know what Sonnet 5 improved, why its lower token price does not make a finished task cheaper, and which Claude model I recommend running for now.
O que você precisa
Access to Claude (claude.ai) or the models you currently run
A rough idea of which model your workflows use today (Sonnet 4.6, Opus 4.8 or a GPT model)
Passo a passo
Know what was released
Anthropic released Sonnet 5, the most powerful version of its Sonnet line. Before switching anything, look at two things I walk through: the launch benchmarks, and what a completed task actually costs.
Read the benchmarks against Sonnet 4.6
On Anthropic's first benchmark chart, Sonnet 5 beats Sonnet 4.6 on every metric shown: computer usage, agentic coding and general reasoning. On raw scores, it is a clear upgrade over the previous Sonnet.
Compare it with Opus 4.8
Check the Opus column. Everywhere except agentic coding, Sonnet 5 lands near Opus 4.8, which was the best model before Fable 5. On paper, that looks like Opus-level intelligence at a lower price.
Look at cost per task, not cost per token
Open Anthropic's second graph comparing Sonnet 5 and Opus 4.8. You pay less per token, but per task you pay about the same as Opus 4.8. Judge the model on what a finished task costs you.
Check it against GPT 5.5
A graph shared on X shows the same pattern: Sonnet 5 cost $2.30 for a task where GPT 5.5 cost $1. That is more than twice the price for the same job.
Understand why the bill goes up
Sonnet 5 misbehaves less than the previous version, but you pay for it. I explain that because it is less intelligent than Opus, it has to think longer before it does a task, so it spends more tokens.
Choose what to run for now
If you run on Sonnet today, stay on Sonnet 4.6 until this gets fixed (maybe with Opus 5.1). If you need more intelligence, I'd say you might as well use Opus 4.8, since Sonnet 5 costs about the same per task.
Open Claudehttps://claude.ai
Treat launch charts as marketing
Take launch metrics with a grain of salt: they are marketing. In this case, the problem shows up even in Anthropic's own charts, so read cost per task before you trust any headline number.
Fique atento a
A lower price per token does not mean a cheaper completed task: Sonnet 5 costs about the same per task as Opus 4.8.
A less intelligent model can think longer and spend more tokens, which pushes the cost up.
I do not recommend Sonnet 5 for business use cases until the pricing issue is fixed.
Launch benchmarks are marketing, so check the cost charts, not just the scores.
