On 22 September, Anthropic released Claude Opus 5.5 and retired Opus 5 with it. Two things stand out right away: the model costs a fifth less than its predecessor and thinking can no longer be switched off. Opus 5.5 is available on ChatX from today.
What changes compared to Opus 5
According to Anthropic, Opus 5.5 costs around twenty percent less than its predecessor, on input as well as on output. Context window and output limit stay the same, one million tokens of context and 128,000 tokens of output. The cache is even more striking: reads now cost five percent of the input price instead of ten. In long conversations, where ChatX reuses the history, that accounts for most of the saving.
On substance, Anthropic names three improvements that show without any effort on your side. At the same level, the model thinks more than Opus 5, and it reads values off dense charts, diagrams and screenshots far more precisely. On top of that, it is built for long tasks with many steps. Its knowledge runs to June 2026.
Thinking is now mandatory
The most important change is not the price. On Opus 5 you could switch thinking off. On Opus 5.5 the interface rejects that with an error: adaptive thinking is always on and only its depth is steered. Anthropic names Medium as the default.
On ChatX that means only Low, Medium and High are left under “Depth of thought” for Opus 5.5. The “Off” is gone because it no longer exists technically. We start on Low so that short questions do not burn tokens needlessly. If you deliberately used Opus 5 without thinking, Low is now the closest thing to that behaviour.
One quirk goes with it. The short notes the model writes while using tools no longer arrive as normal text. With web search on, it therefore stays quiet between the search steps until the answer begins. That is not a fault, it is the new response format.
What a request costs on ChatX
The table converts a typical request of 600 input tokens and a 400 token answer into ChatX tokens and puts the frontier models side by side.
| Model | Tokens per request | Ratio |
|---|---|---|
| Claude Sonnet 5 | 1,156 | 0.5× |
| Claude Opus 5.5 | 2,312 | 1× |
| GPT-5.6 Sol | 2,312 | 1× |
| Claude Fable 5.1 | 5,778 | 2.5× |
| GPT-6 Astra | 5,778 | 2.5× |
Opus 5.5 therefore sits exactly at the level of GPT-5.6 Sol and costs two and a half times less than Fable 5.1 or GPT-6 Astra. One thing to keep in mind: more thinking means more output tokens. On High an answer can cost a multiple of the table values, on Low it stays in this range.
When to use Opus 5.5 and when another model
- Opus 5.5 on Low: programming, reading code, multi-step analysis, reading charts and screenshots. The standard case.
- Opus 5.5 on Medium or High: tricky bugs, tasks with several possible solutions, long refactorings. Only here does the step up from Low pay off.
- Sonnet 5: everyday texts, summaries, translations. Half the price and barely behind on those tasks.
- Fable 5.1: only for the cases where Opus 5.5 fails twice. That is what the two and a half times price is for, not for daily use.
- Images and files: Opus 5.5 understands images, so you can drag screenshots and charts straight into the chat.
Opus 5 is thus retired on ChatX and no longer selectable. Anthropic guarantees Opus 5.5 until at least 22 September 2027. If you want to know what sits behind the switch technically, the list of changes is in the release notes and in the migration guide.