Claude Opus 4.8: “a modest however tangible enchancment”
twenty eighth Might 2026
Anthropic shipped Claude Opus 4.8 at present. My favorite factor about it’s this observe within the launch announcement:
Customers will discover Opus 4.8 to be a modest however tangible enchancment on its predecessor. There’s nonetheless extra to be executed: we’re engaged on growing and releasing fashions that present most of the identical capabilities as Opus at a decrease value.
It’s so refreshing to see an AI lab actually describe a launch as a minor incremental enchancment over the earlier mannequin!
Honesty appears to be a theme. Right here’s my different favourite observe from that announcement:
Probably the most outstanding enhancements in Opus 4.8 is its honesty. We prepare all our fashions to be honest—for occasion, to keep away from making claims that they’ll’t assist. However a normal downside with AI fashions is that they generally leap to conclusions, confidently claiming to have made progress of their work regardless of the proof being skinny. Early testers report that Opus 4.8 is extra prone to flag uncertainties about its work and fewer prone to make unsupported claims. That is borne out in our evaluations, which present that Opus 4.8 is round 4 occasions much less probably than its predecessor to permit flaws in code it has written to move unremarked.
That linked system card consists of the next:
Claude Opus 4.8 had the bottom incorrect-rate of the six fashions on each benchmark—essentially the most direct measure of factual hallucination. It achieved this primarily by abstaining on questions on which it was unsure moderately than by answering extra questions accurately.
Mannequin traits
Not a lot has modified since 4.7.
It’s priced the identical as Opus 4.5/4.6/4.7—$5/million enter and $25 per million output. “Quick mode” is twice that worth, which is a major discount from their earlier fashions—quick mode on 4.6/4.7 stays at $30/$150. Word that quick mode is just obtainable to organizations which can be a part of the analysis preview, “Contact your account supervisor to request entry”.
Each the dependable data cutoff and the coaching knowledge cutoff are January 2026, the identical as for 4.7.
The context window continues to be 1,000,000 tokens, and the max output is 128,000 tokens.
The What’s new in Claude Opus 4.8 doc has a few of the extra attention-grabbing particulars. These caught my eye:
Mid-conversation system messages. Claude Opus 4.8 accepts function: “system” messages instantly after a person flip within the messages array (topic to placement guidelines). This allows you to append up to date directions later in a long-running dialog with out restating the total system immediate, which preserves immediate cache hits on the sooner turns and reduces enter value on agentic loops.
See additionally this replace to the Anthropic Python SDK. With the ability to steer the system immediate mid-conversation sounds actually highly effective. I used to be apprehensive this could be incompatible with the abstraction supplied by my very own LLM library, which expects a single system immediate per dialog… but it surely seems my latest redesign ought to deal with that simply high-quality.
Decrease immediate cache minimal. The minimal cacheable immediate size on Claude Opus 4.8 is 1,024 tokens, decrease than on Claude Opus 4.7.
I checked and 4.7’s minimal was 4,096.
And a few pelicans
Listed below are pelicans driving bicycles for all 5 considering ranges, low, medium, excessive, xhigh, and max:
This time I ran them utilizing the LLM CLI, exported the logs to Markdown after which had Claude Opus 4.8 construct me an HTML device that might render that Markdown with the svg fenced code blocks displayed as SVGs on the web page.
(I later had GPT-5.5 xhigh in Codex replace that code to take away any XSS holes. I’m positive Claude might have executed that if I’d requested, however GPT-5.5 is my code safety blanket in the meanwhile.)
The max one was clearly the most effective, but it surely did take 25 enter, 17,167 output tokens for a complete value of 43 cents!
