Edition #17. Anthropic made the news three times this week, each time for a different reason. On May 28 came Claude Opus 4.8: better coding scores at the same price. That same day it raised 65 billion at a 965 billion valuation. And IBM put up 5 billion for open-source security, triggered by Mythos, the same Anthropic model the IMF feared a 'systemic shock' from earlier this month. Let's dive in!
Claude Opus 4.8: same price, better code, fast mode three times cheaper
Anthropic calls 4.8 no leap, but a 4.7 that behaves better. Same price (five dollars per million input tokens, twenty-five for output), same one-million-token context window. On SWE-bench Pro it hits 69.2 percent, ten points above GPT-5.5, and it does that with fifteen percent fewer turns and thirty-five percent fewer output tokens than 4.7. New: a fast mode that runs 2.5 times faster and three times cheaper, and effort control that lets you set the reasoning level yourself, from normal to extra high.
The real gain is in the behavior. 4.7 too often reported a task was done when it wasn't. 4.8 flags its own mistakes before it delivers and is more honest about what it doesn't know. That saves me the half review I'd otherwise have to do myself, and it's exactly what makes a model usable for a process where money is on the line instead of a demo.
65 billion raised at a 965 billion valuation
Series H: 65 billion raised, 965 billion valuation, nearly a tripling since February (380 billion). Led by Altimeter and Sequoia among others, widely seen as the last private round before the IPO. But the number that matters leaked alongside it: run-rate revenue stands at 47 billion.
That 47 billion is the story, not the valuation. Not a promise about what AI might one day do, but revenue coming in now from companies that have put Claude to work. We build on it daily, and a vendor with this revenue doesn't fall over and doesn't have to double its prices to survive. Two years ago that was still a gamble.
IBM puts up 5 billion for open-source security, because of an Anthropic model
IBM and Red Hat launched Project Lightwell, five billion dollars for open-source security. The trigger, in the words of IBM boss Arvind Krishna: Mythos. That same model found nearly 3,900 high- to critical-severity vulnerabilities in open source in a preview, and is so good at finding and exploiting weak spots that Anthropic only releases it to a select group, under the name Project Glasswing.
Lightwell becomes a clearinghouse where companies report and fix vulnerabilities, backed by 20,000 IBM engineers plus AI. Bank of America, JPMorgan and Visa signed on first. The core: the same reasoning power that writes code also finds the holes. Open source sits in nearly every stack, including yours, and 'never had a problem' is no longer a defense when a model finds four thousand holes in a weekend.
What kept me busy in Claude Code this week: dynamic workflows
The feature I find most interesting shipped with Opus 4.8: dynamic workflows. Instead of doing a task itself, Claude writes an orchestration script that runs tens to hundreds of sub-agents in parallel and checks its own work. The example from the announcement: porting Bun from Zig to Rust, 750,000 lines, eleven days, 99.8 percent of tests passing.
Built for jobs too big for a single session: auditing a codebase, a migration across thousands of files, a security review. It costs far more tokens, so you deploy it deliberately. The pattern stays the same as the weeks before: the human shifts from operator to orchestrator.
Want to work smarter with AI?
Want to work smarter with AI, inside your business or just for yourself?
At AI Agent B.V. we help businesses and individuals with AI advisory, AI implementation and practical training. Not just stories about what AI might one day do, but concretely helping you use AI smarter today in your work, processes and daily operations. From strategy and tooling choices to training, implementation and custom AI agents.
Interested? Book a free half hour.
