Claude Sonnet 5 launched June 30 at $2 per million input tokens and $10 per million output tokens, with a clearly stated caveat: introductory pricing through August 31, with an increase to $3/$15 landing September 1.

On August 10, Anthropic reversed that. The pricing page now carries a note confirming the $2/$10 rate is permanent and the September increase will not happen.

Twenty-two days into a 63-day introductory window is a fast reversal. Something changed.

Source spread

• Anthropic — Claude platform pricing — [builder]. Primary source. The note at the Sonnet 5 row is the official announcement.
• Datafloq — Anthropic confirms September price rise — [skeptic]. Pre-freeze coverage of the expected September increase; useful for understanding what teams were budgeting for.
• Tech Journal — Sonnet 5 pricing now permanent — [builder]. Covers the announcement and builder implications.
• Blur Brah Lab — Anthropic cancels September hike — [builder]. Quick-take coverage; accurate on the core fact.

Pros & cons

What's straightforwardly good:

• The 50% increase disappears. Teams already running Sonnet 5 in production don't need to model a rate change or plan a migration. That's work that didn't happen.
• $2/$10 at Sonnet 5 quality is a genuinely competitive rate. DeepSeek V4 Pro 0813 is cheaper at $0.44/$0.87/M, but that's a different capability tier for most production use cases.
• The pricing is clean and durable now. A model you can plan around is worth something beyond the per-token rate.

What's worth watching:

• Anthropic held $3/$15 for Sonnet 4.6 while dropping Sonnet 5 to $2/$10 permanently. That's a two-tier structure where the better model is also the cheaper one. That raises a reasonable question: what exactly is Sonnet 4.6 still for?
• The ~30% tokenizer overhead on Sonnet 5 vs 4.6 still applies. The $2/$10 rate card isn't the whole story. On typical text workloads, the real effective cost is closer to $2.60/MTok equivalent — still better than $3, but worth accounting for in cost models.
• This reversal was made with 40 days still left in the introductory window. If Anthropic can reverse an announced pricing decision on 22 days of data, that's useful information about how stable their rate commitments are.

Why this happened when it did

Three things were moving in July and early August that could explain a pricing reversal before the introductory window closed.

Competition got cheaper fast. GLM-5.3 launched API-access at $18/month. DeepSeek V4 Pro 0813 went GA at $0.44/M. Grok 4.6 repriced. A September move to $3/$15 would have put Sonnet 5 in a market that had gotten materially more competitive since launch pricing was set.

Adoption likely ran ahead of projections. July usage data showed Anthropic holding enterprise share. If Sonnet 5 usage grew faster than expected at the $2/$10 rate, the revenue math at that price might simply be better than a lower-volume rate at $3/$15.

Product-line spacing. Sonnet 5 at $2/$10 sits 60% cheaper than Opus 5 at $5/$25. Move Sonnet 5 to $3/$15 and that gap shrinks. The current spacing creates a clean decision: Haiku 4.5 for high-volume cheap work, Sonnet 5 for production-quality mid-tier, Opus 5 for the hardest reasoning tasks. Raise Sonnet 5's price and the positioning gets muddier.

I don't know which of these drove the decision. Probably more than one.

Further reading

• Anthropic — Claude platform pricing — the official source; the note on the Sonnet 5 row is the announcement
• TechCrunch — Anthropic launches Claude Sonnet 5 — June 30 launch context, including the original introductory pricing framing
• Datafloq — Prior expectations for the September increase — what teams were planning against before the reversal