Anthropic makes Sonnet 5's $2/$10 pricing permanent
Anthropic scrapped the planned end of Claude Sonnet 5's introductory rate, locking $2 per million input and $10 per million output tokens past August 31.
Anthropic said on Monday that the introductory pricing on Claude Sonnet 5 will not expire. The model launched in June at $2 per million input tokens and $10 per million output tokens, with that rate advertised as running only through 31 August 2026. Those numbers are now the standing price.
What changes
The published plan had Sonnet 5 reverting to $3 per million input and $15 per million output tokens once the promotional window closed — the level Anthropic's mid-tier models have historically occupied. Making the lower figure permanent is a one-third cut to input cost and a one-third cut to output cost against the price developers had been told to budget for from September.
Anthropic framed it as a simple removal of the deadline rather than a new promotion, which matters for procurement: teams that had modelled a step-up in unit costs six weeks out no longer need contingency for it.
Context
The decision lands in a market where per-token prices are being pushed down from both directions. Open-weight releases from Chinese labs and Meta have made capable models available at hosting cost, and independent evaluations have repeatedly shown cheaper models capturing a large share of frontier performance on coding and agentic tasks. At the same time, agentic workloads consume far more tokens than chat — long tool-calling runs can burn millions of output tokens per task — so the mid-tier price point, not the flagship one, increasingly determines what an AI product costs to operate.
Why it matters
Sonnet-class models are the default workhorse behind a large share of production coding agents and enterprise pipelines, including Anthropic's own Claude Code. Freezing that rate removes a scheduled cost increase from thousands of budgets and signals that Anthropic sees no room to raise mid-tier prices while open-weight competition holds. For anyone forecasting inference spend, the mid-tier floor just got harder to move upward.