Opus 4.8: the benchmarks crept up, but the fast mode is the actual upgrade

Anthropic dropped Opus 4.8 on 28 May 2026, about six weeks after 4.7. Same six-week heartbeat, same polished blog post. Let's separate the marketing from the upgrade.

Anthropic's own number: agentic coding 64.3%→69.2%. That's a five-point move on Anthropic's chosen metric — genuinely useful, not a revolution, and (as always) a figure the vendor picked and you should treat as marketing-adjacent. More interesting is the reliability claim: ~4× less likely to let flaws in its own code pass unremarked. If that holds up in real agent runs, it's worth more than any leaderboard.

At the bottom, the post teased 'Mythos-class' models 'in the coming weeks.' Read that as: 4.8 is the appetiser. Days later, Fable 5 — the first public Mythos model — showed up, only to get yanked offline by a US directive. So the honest summary of Opus 4.8 is that it was a solid, incremental release whose most important line was a pre-announcement of the model that actually made the news.