Another model. Measured.
Which AI model to use for what, at what effort, and what it costs you in money or in subscription quota. Measured where we can, dated always, sourced every time.
Sonnet 5.5 is cheaper than Sonnet 5, until you turn it up to max
Anthropic says Sonnet 5.5 needs far fewer tokens than Sonnet 5 for the same work. Artificial Analysis says it uses more output tokens per task than any model it has measured. Both are right, at different effort settings.
At medium, the Claude Code default, Sonnet 5.5 costs 41% less per task than Sonnet 5 on Artificial Analysis's index, and scores higher than Sonnet 5 does at max. At max, where Anthropic's headline scores come from, it costs 49% more. Above high, Opus 5.5 a step or two lower gets about the same score for less than half the cost.
The whole page: charts, every number, sourcesClaude Sonnet 5.5 is out, at Sonnet 5's price per token
It costs $2 per million input tokens and $10 per million output tokens. The Claude apps and Claude Code start it on medium effort, the API on high.
Source: Introducing Claude Sonnet 5.5 (launch page), read 28 Sep 2026, 21:17 CEST · our video
At max, Sonnet 5.5 writes more tokens per task than any model Artificial Analysis has measured
It used ~193k output tokens per Intelligence Index task. Artificial Analysis ran it on a pre-release deployment with a structured-output bug and says it will re-run.
Source: Artificial Analysis, launch article on Sonnet 5.5, read 28 Sep 2026, 21:19 CEST · our video
Ember-1: "half the tokens, same answers"
That's Fireworks' claim for its new model, from its launch page. Our video on it comes next.
Not measured yetSource: Fireworks, Introducing Ember-1, read 28 Sep 2026, 18:55 CEST
| setting | Sonnet 5.5 | Sonnet 5 |
|---|---|---|
| low | $0.41 | $0.51 |
| medium | $0.59 | $1.00 |
| high | $1.08 | $1.79 |
| xhigh | $2.74 | $2.87 |
| max | $7.60 | $5.09 |
Artificial Analysis Intelligence Index, average cost per task at list prices. The marked row is the Claude Code default. Pre-release deployment of Sonnet 5.5; Artificial Analysis says it will re-run. Read 28 Sep 2026, 21:29 CEST.
By question
Every video answers one of four questions. They're listed here under the one they answer.
Which model
for…
No video yet.
How much
effort…
should you run Sonnet 5.5 at?
What does
it cost…
No video yet.
Can I
trust…
No video yet.