Tech Stack
Anthropic released Claude Sonnet 5 on June 30, and the launch pricing tells you exactly what they're doing: $2 per million input tokens and $10 per million output through August 31, then $3/$15 after. That's a fifth of Opus 4.8's $5/$25 — for a model that, on some benchmarks, is functionally Opus's equal.
If you run production agents, this release isn't news. It's a repricing of your entire stack.
The benchmark that made me look twice
On GDPval-AA v2 — the knowledge-work benchmark — Sonnet 5 scores 1,618. Opus 4.8 scores 1,615. Statistically, that's a tie, at 20% of the price (until August 31).
The full picture is more nuanced, and the nuance is where your routing decisions live:
- -Agentic coding: 63.2% vs Opus 4.8's 69.2% (and up from Sonnet 4.6's 58.1%). Opus keeps a real 6-point lead where it matters most for engineers.
- -Terminal-Bench 2.1: 80.4% — remarkable for a mid-tier model. Terminal-heavy agent work is now firmly Sonnet territory.
- -Computer use (OSWorld-Verified): 81.2% — browser and desktop automation agents just got dramatically cheaper to run.
Add the platform specs — 1M token context, 128K output (300K via the batch API beta), adaptive thinking always on — and Sonnet 5 has every structural feature that used to justify reaching for the big model.
What I'm changing in my own routing
I run a routing ladder across everything I build — the same escalation pattern I've used since running Haiku-to-Sonnet routing for a construction SaaS sales team. Here's the post-Sonnet-5 version:
- -High-volume classification, summaries, extraction — Haiku, unchanged. Nothing below Sonnet got repriced.
- -Agent loops: tool calling, terminal work, browser automation, knowledge work — Sonnet 5, and this is the big migration. This tier used to split between Sonnet 4.6 (when budget won) and Opus (when quality won). At near-Opus quality for a fifth the price, the split is gone.
- -Hard engineering: multi-file refactors, long agentic coding sessions — still Opus 4.8. The 6-point agentic-coding gap is real, and in code, quality failures compound.
- -Architecture decisions and unsticking hard problems — Fable 5, sparingly.
The practical effect: the expensive middle of my token bill — agent orchestration, research loops, document processing — just dropped by roughly 60-80%.
The Aug 31 detail everyone will miss
The $2/$10 pricing is introductory; it becomes $3/$15 on September 1. Two months of the strongest price-performance in the market is a window, and it's exactly the window to run your migration experiments: A/B your agent workloads on Sonnet 5 now, measure quality deltas on your actual tasks (not benchmarks), and lock your new routing before the price normalizes. Even at $3/$15, most of the migration math still holds — but the free experimentation window doesn't.
The bigger pattern
Every provider is converging on the same shape: a cheap tier, a workhorse tier that keeps absorbing the flagship's abilities, and a flagship that exists for the 10% of tasks that genuinely need it. The winners in this market aren't the teams using the best model — they're the teams whose routing tables are never more than a release behind.
Sonnet 5 made every routing table in production stale on June 30. Mine included. Update yours.
Want to build something like this?
I architect and deploy end-to-end AI systems — from MVP to revenue.
Let's Talk