Last week Anthropic released Opus 5.5; this week it immediately followed with Claude Sonnet 5.5 — the launches are coming faster and faster.
How long has it been since the major vendors last called on the industry to slow down, repeatedly warning that AI is about to destroy humanity? Less than a month. Destroy humanity at the start of the month; at the end of the month, humanity keeps marching on.
This Sonnet 5.5 is again positioned as a high cost-effectiveness, volume-driving productivity play, priced at $2 per million input tokens and $10 per million output tokens. But with fewer tokens required per task, the actual overall cost dropped by roughly 30%, and response speed improved by over 30%.
The capability gains are mainly in coding and Agent tasks. Terminal-Bench 4.0 score jumped from 10.3% on the previous generation to 70.6%, and CursorBench reached 55.5%, closing in on Opus 5.5 and GPT-6 Sol.
A routine iteration with no particularly notable highlights — just smarter and safer. I won't bother listing the biosecurity risk review, cyberattack defense, and sandbox escape prevention metrics.
Oh, right — "anti-distillation" has been further strengthened, and it is highly targeted.
All talk of the welfare of all humanity, yet the actions are remarkably honest. On one hand, pushing risk to the extreme, calling on the entire industry to slow the pace of R&D and shore up safety defenses; on the other, rolling out new models one after another.
Those who reached shore first are in a hurry to kick away the ladder.