Five months ago, a YouTube channel called WorldofAI posted a video calling it the greatest coding model ever, built on leaked screenshots and a guess at the release date. By June, the guessing had hardened into specifics that got repeated across a dozen SEO blogs: codename Fennec, 82.1% on SWE-bench, a launch on February 3rd that never happened. I went looking for all three in what Anthropic actually published today, June 30th: an announcement and a system card that runs past a hundred and forty pages. None of them hold up: no Fennec in either document, no February launch, and nothing in the system card matching that exact score.

Sonnet 5 itself is more modest than the rumor mill promised, and I think more interesting for it. It's pitched as the most agentic Sonnet yet: it can plan, reach for a browser or terminal, and run unattended at a level Anthropic says used to require an Opus-class model. That's the headline Anthropic chose this time, not a single score but a capability claim, that the gap between Sonnet (cheap) and Opus (expensive) has narrowed. I wrote about how Anthropic keeps redrawing that line between its model tiers earlier this month; Sonnet 5 is the latest redraw, and for once it's the cheap tier moving up rather than the expensive one moving down.

Pricing backs the framing, sort of. Introductory rates are $2 per million input tokens and $10 per million output, through August 31st, before settling at $3/$15 — the same headline numbers the rumor sites had floated months ago while getting everything else wrong. There's a catch I had to read twice: a new tokenizer means the same input now produces up to 35% more tokens than before, so the lower sticker price is partly an accounting trick to keep the transition cost-neutral rather than a straightforward discount.

The system card is where the honesty gets interesting. Anthropic writes that the risk of significantly harmful outcomes from Sonnet 5 "is very low, but higher than for models prior to Claude Mythos Preview" (Mythos being the more capable, more tightly held line that sits above Sonnet and Opus), and that its autonomy-relevant capabilities stay under that line's threshold. It's a company grading its own homework, and for once grading itself down rather than up. I'll credit that without fully trusting it.

I'd hold the applause regardless. Sonnet 4.5 arrived in September to real praise, then spent the following weeks absorbing complaints on Reddit and Hacker News about hallucinated APIs and a "laziness" that hadn't been there before, the kind of problem that never shows up in day-one coverage because day-one coverage happens before anyone has used the thing for six straight weeks. Sonnet 5 will get the same grace period and probably the same reckoning. The review I actually want to read isn't this one. It's the one written in late August, after the introductory pricing expires and the early-adopter goodwill has worn thin enough for people to start complaining honestly.

Sources: