Anthropic ships Sonnet 5 and advertises that it's worse at hacking on purpose

A launch that markets a capability the model lost

Anthropic released Claude Sonnet 5 as an upgrade to Sonnet 4.6, positioned as "the most agentic Sonnet model yet" — able to plan, drive browsers and terminals, and run autonomously at levels that recently required bigger models. Introductory pricing is $2 per million input tokens and $10 per million output through August 31, rising to $3/$15 after. Effort is now a dial from low to extra-high, trading cost for depth.

The genuinely odd part is what the launch chooses to advertise: a regression. Anthropic states the model "shows substantially poorer performance than models such as Opus 4.8 and Mythos 5" on dangerous cyber skills like developing software exploits. On a Firefox exploit evaluation, "neither of the Sonnet models could successfully develop a working exploit (both scored 0.0%)." The system card goes further — on CyberGym vulnerability discovery, Sonnet 5 is called less capable than the older Sonnet 4.6, and with default safety mitigations on, it scores a flat 0. A weakened offensive-security profile is being sold as a safety win.

The thread does the price-performance math and comes up empty

The reaction wasn't about safety framing — it was arithmetic. Multiple readers ran Anthropic's own cost-per-task charts and reached the same verdict: Sonnet 5 is dominated by its bigger sibling across most of the curve.

"Opus 4.8 beats Sonnet 5 on the pareto frontier in several of their graphs (Agentic Search, Agentic Computer Use)… if you're doing something hard, just use a bigger model." — andai, 15k karma

A jewelry-business owner and longtime HN engineer put the operational takeaway bluntly: there's a hard ceiling on when Sonnet 5 makes sense.

"The cost per task chart is telling me that I should never use Sonnet 5 above medium effort level - Opus always performs better for a given cost." — an engineer who runs his own product on these models

The cyber-safety marketing drew open mockery. One voice with maxed karma reduced the whole positioning to a slogan:

"American AI company status: We are now bragging about how bad our models are unironically." — jchw, 29k karma

There were dissents from people actually shipping with it. A Salesforce AI engineer called it a solid workhorse update and said he'd been using Sonnet over Opus for most coding for a while, spending far less for near-equal output with a little task decomposition. And a buried footnote turned out to matter for real bills:

"Sonnet 5 is an upgrade to Sonnet 4.6, but it uses an updated tokenizer… can map to more tokens: roughly 1.0–1.35× depending on the content type. AKA expect higher costs on Sonnet 5 vs Sonnet 4.6 for the same tasks." — an engineer who read the release footnotes

So the effective price cut from launch pricing is partly clawed back by a tokenizer that counts more tokens for the same input. The clearest read of the day came from a commenter who noticed the release's own comparison language keeps pointing away from the product being launched: it repeatedly explains that Sonnet 5 is less capable than Opus 4.8 and Mythos 5 — an unusual thing for a headline model to keep saying about itself.