Claude Sonnet 5's Launch Pricing Ends September 1, and the Real Story Is What It Reverts To
Claude Sonnet 5 launched with introductory pricing of $2 per million input tokens and $10 per million output tokens. On September 1, 2026, that reverts to $3 and $15, a uniform 50% increase across both directions. If you built a cost model around the launch price for a high-volume workload, this is the week to requalify it.
The more interesting fact isn't the size of the increase, it's what the reverted price equals: exactly what Sonnet 4.6 already charges. Sonnet 5 isn't settling into a premium tier above its predecessor, it's returning to price parity with it. That reframes the whole "launch pricing" period as a customer-acquisition discount on a model that was always going to cost the same as the one before it, not a permanently cheaper successor.
The direction matters as much as the number
The same week Anthropic's reversion was reported, OpenAI moved the opposite direction: GPT-5.6's cheapest tier, Luna, dropped 80% to $0.20 and $1.20 per million tokens, while the mid tier, Terra, fell 20%. Two frontier labs adjusted pricing in the same window, one up, one down, which is itself worth noticing. It suggests pricing at the frontier isn't converging toward a single market rate, it's diverging by tier and by vendor, and the ladder you're actually paying against depends heavily on which specific model and tier you picked, not a single moving "AI pricing" number.
What to actually do about it
If a meaningful share of your workload runs on Sonnet, re-run your unit economics before September 1 using $3/$15, not the launch rate. If your workload is high-volume and cost-sensitive, this is also a reasonable moment to check whether a cheaper tier from a competing provider, GPT-5.6 Luna's new pricing is a real option, now clears your quality bar for the simpler slice of your requests. Pricing at the frontier is moving fast enough in both directions that a cost assumption from three months ago isn't safe to still be running on.