field note
Opus 4.8 shipped on the 28th.You probably hit update, watched it work, and moved on. I did too, for about an hour.Then I actually read what shipped with it. And I realized most people are about to overpay for this model for no reason at all.The model is a small step. The stuff bolted on next to it is the actual release. Three controls. All off by default. All sitting there while you never touch them.Let me show you what I mean.The 30-Second VersionSame price as 4.7. $5 in, $25 out per million tokens. They didn’t raise it.1M context window, on by default now.SWE-bench Verified crept from 87.6 to 88.6. SWE-bench Pro jumped harder, 64.3 to 69.2. Terminal-bench went 66 to 74.And the number that actually matters: it’s 4x less likely to hand you broken code without flagging it. First Claude to score a flat 0% on confidently reporting busted results as if they worked.The benchmarks aren’t the story though.claude-opus-4-8# $5 / $25 per 1M tokens (standard) · $10 / $50 (fast, 2.5x speed) The story is three things: a control over how hard it thinks, a fast mode that got 3x cheaper, and an agent-fleet feature everyone’s losing their mind over.Let’s go.The Di
< read by a human · updated as things change >
browse hackathons