Claude Sonnet 5's intro pricing ends Aug 31 — and a tokenizer catch
On Sep 1 the $2/M intro rate rises 50% while a new tokenizer adds ~30% more tokens per call — real cost climbs 20-35% over 4.6. Plus: GCC bars AI patches.
Copy markdown
The rate rises 50% on September 1
Sonnet 5's introductory $2/$10 per million tokens (in/out) holds only through Aug 31; on Sep 1 it reverts to standard $3/$15 — a 50% jump on the exact same traffic. Migration is otherwise a one-line change: swap claude-sonnet-4-6 for claude-sonnet-5.
The tokenizer catch that's not on the rate card
Sonnet 5 ships a new tokenizer that emits up to ~30-35% more tokens for identical text than 4.6. Stacked on the rate change, a job that ran ~$675/mo on 4.6 lands around $810-911/mo — 20-35% higher — despite the "same price" launch framing.
Blunt the bill: caching and the Batch API
Prompt caching and the Batch API are the best levers to claw back the tokenizer penalty — audit token-heavy prompts before Sep 1. For the price you also get 1M context, 128k output, adaptive thinking, and the first Sonnet-tier model with real-time cybersecurity safeguards.
Elsewhere: GCC won't take your AI-written patches
The GCC steering committee adopted a policy declining significant AI/LLM-generated contributions — test cases excepted — over unresolved copyright provenance. If you lean on Copilot or Claude for compiler patches, expect them rejected.