Opus, Sonnet, Haiku: Stop Optimizing the Wrong Number | by Nitish Agarwal | Jun, 2026 | MediumSitemapOpen in appSign up<br>Sign in
Medium Logo
Get app<br>Write
Search
Sign up<br>Sign in
Member-only story
Opus, Sonnet, Haiku: Stop Optimizing the Wrong Number
Nitish Agarwal
7 min read·<br>Just now
Listen
Share
Press enter or click to view image in full size
Photo by Justin Morgan on UnsplashThere’s a piece of napkin math making the rounds in agent-architecture circles: Opus is ~19x the price of Haiku, so never let your premium model touch a loop it doesn’t have to. It’s a great line. It’s also running on 2025 prices.<br>Here’s the actual rate card as of mid-2026. Haiku 4.5 is $1 in / $5 out per million tokens. Sonnet 4.6 is $3 / $15. Opus 4.8 is $5 / $25 the same sticker as 4.7 and 4.6 before it. Opus used to live at $15 / $75. Anthropic cut the flagship by roughly 3x and left it there.<br>So the real spread today is about 5x between Opus and Haiku, and 1.7x between Opus and Sonnet. The “19x tax” is dead. If your whole case for a tiered architecture was dodging a price premium, that case just got a lot weaker.<br>And yet tiered agent systems a frontier model orchestrating cheaper workers still routinely cut production bills by 40–60%. The reason has almost nothing to do with the per-token rate. It has to do with a number nobody puts on a slide.<br>the number nobody quotes<br>When Anthropic published the engineering breakdown of their multi-agent Research system, they ran an ablation on the BrowseComp benchmark to figure out what actually drives quality. Three factors explained 95% of the performance variance. Token usage accounted for…
Written by Nitish Agarwal<br>34 followers<br>·39 following
Help
Status
About
Careers
Press
Blog
Privacy
Rules
Terms
Text to speech