Claude Sonnet 5 Took Over Half My Backlog in One Afternoon

Two Anthropic stories landed within forty eight hours and both matter, but let's start with the one that ate my Tuesday.

The Experiment

Sonnet 5 shipped June 30 at $2 per million input tokens, $10 out, with a promise: performance approaching Opus 4.8 while actually using tools like browsers and terminals without hand holding. Anthropic calls it the most agentic Sonnet yet, capable of making plans and running autonomously at a level that recently required pricier models.

Skeptical but curious, I dumped eleven stale GitHub issues from my school website project into Claude Code and walked away to make coffee.

When I came back, six of them were done. Real fixes, sensible commit messages, one even added a test.

What Six Fixed Issues Actually Looked Like

• Broken aria-label on the staff directory search (screen readers announced nothing useful)
• Calendar widget treating submission timestamps as UTC when the school runs on Eastern time
• Duplicate fetch firing twice in a useEffect with an unstable dependency array
• Mobile nav z-index war with the sticky header
• Dead CSS from a redesign nobody cleaned up
• Alt text pass on every image that had none

None of these were hard. That's the point. They were the chores that lose priority battles forever, and an afternoon of unattended computing deleted them.

The remaining five needed judgment calls it correctly refused to guess on, like anything touching student data handling. Honestly, knowing where the line is might be the most impressive part.

Why the Pricing Changes Everything

At this cost I can leave it running on background chores the way people leave roomba vacuums wandering the house. My contribution graph looks fitter than I am right now. And the effort controls mean you dial between quick answers and deep autonomous runs depending on the task, with rate limits raised across Chat, Cowork, Claude Code and the API to match.

Story Two: Fable 5 Comes Back

Fable 5 and Mythos 5 returned July 1 after being pulled offline mid month over government cybersecurity concerns. They're back with stricter safeguards, and Mythos 5 access stays limited to vetted partners through Project Glasswing while a broader trusted access program gets built.

Watching a frontier model get recalled and redeployed like a car with a faulty airbag was not on my 2026 bingo card.

But here we are, and the precedent matters for everyone building on top of these APIs.

Takeaways

• Cheap capable agents change what's worth delegating
• Refusals on ambiguous high-stakes work are a feature, not a limitation
• Model availability is now a thing your architecture should survive

Cheers, Yassen