There is a new model again. Claude Opus 4.8, the successor to 4.7. I do not read these releases for the benchmarks. I read them to see what changes about the way I work, in the sessions I run and in the things I build with them myself.
This time the difference is not a bigger number on a slide. It is in the details you only notice after a few hours of real work.
What is new
The core, briefly. Claude Opus 4.8 runs on a one million token context window by default and can produce up to 128k tokens of output. That is a lot of room for long documents and long working sessions.
Beyond that, a few things sound smaller than they are:
- Adaptive thinking. The model only reasons when the question calls for it. A simple lookup gets answered straight away, a complicated problem gets thought through first. Fewer wasted reasoning tokens, at the same setting.
- Fast mode. A research preview that gives you up to 2.5 times more output tokens per second from the same model. It is priced at a premium, but for work where speed matters it can be worth it.
- Steering mid-conversation. You can now add a system message in the middle of a running conversation, without resending the whole instruction. On long agent loops that saves money, because the cache on the earlier turns stays intact.
- A lower bar for caching. Prompts from 1,024 tokens can now be cached. Short prompts that were just too small on 4.7 now benefit without you changing anything.
- Better at long jobs. Fewer interruptions during long coding sessions, and fewer cases where the model skips a tool it was supposed to use.
The full list lives in Anthropic’s release notes.
My favourites
Not all of the above touches my daily work equally. Three things do.
Adaptive thinking is the big one for me. I often work with a mix of simple and heavy questions back to back. On the old model you paid for reasoning on every question, even when it was not needed. Now the model decides per turn. That sounds technical, but the effect is plain: less waiting and lower cost, without me touching a dial.
Fewer interruptions on long jobs is the second. When I build something for a whole day, the old version would sometimes stall on trimming the context. I would lose the thread and lose time. Those moments are rarer now, and when they happen the model recovers better. You do not hear that in an announcement, but you feel it at the end of a working day.
Fast mode is not something I leave on all the time, but I am glad it exists. For a live session, where I do not want a room watching a loading bar, speed is sometimes worth more than the lowest price.
What I make of it
A model that decides per turn whether it needs to think, that fits how I think about these tools. Not more power for its own sake. Just less waste, and a tool that adapts to the work instead of the other way around.
That is exactly the kind of improvement you do not see in a demo, and do see in your bill and your day. To me that counts for more than any benchmark number.
This piece also appeared in Dutch: Wat er nieuw is in Claude Opus 4.8.