Sonnet 5.5 claims 30% faster. My inbox noticed something different
Everyone reported the benchmark. Nobody reported what the speed does to the shape of a workday. I ran my normal Monday twice — once per model.
Sonnet 5.5 launched with numbers: 30% faster output, up to 30% lower cost per task, Terminal-Bench at 70.6% against Sonnet 5's 10.3%. Fine. Benchmarks don't pay my invoices, so I ran a deliberate experiment — one full Monday of my actual work twice, once on each model, same tasks, my own timing.
The tasks were unglamorous: fourteen client emails, a project brief rewrite, some contract question-mark paragraphs, a table of survey results summarized, and the weekly newsletter draft. Exactly the 'well-scoped everyday work' Anthropic claims Sonnet 5.5 is built for. On Sonnet 5, the day's AI portions took about 2 hours. On 5.5, about 80 minutes.
But the interesting finding wasn't minutes. With the old model, my pattern was batch-and-wait — queue several prompts, switch to something else, come back. With the faster one, I stopped batching. Ask, get, move on, ask again. The workflow changed shape because latency crossed under my attention-switching threshold. That's worth more than the raw 30%, and it's why speed improvements compound oddly.
Quality held up in the places I could measure. The email drafts needed the same number of edits, the brief was arguably tighter, and the contract answers flagged the same clauses. The Terminal-Bench jump suggests real capability gains on agentic coding, which I can't personally verify in a day of email triage — but nothing got worse in my lane.
The cost cut is the quiet revolution here. My monthly bill at the same workload dropped noticeably, and for teams running thousands of tasks a day, 30% is headcount math. For solo me, it's the difference between the subscription being obviously worth it versus me shrugging at renewal. Cheap and fast is how AI stops being a line item you think about.