CloudCodeTree LogoCloudCodeTree
AI NewsTutorialsAbout
CloudCodeTree Logo
CloudCodeTree
  • AI News
  • Tutorials
  • About
← Back to AI News
Two Anthropic Updates This Week: The September 1 Sonnet 5 Cost Cliff and the Open-Weights Clarification

Two Anthropic Updates This Week: The September 1 Sonnet 5 Cost Cliff and the Open-Weights Clarification

Chris Harper

2 min read

Aug 2, 2026 · 12:09 UTC

AI
News
LLM

Sonnet 5's intro pricing ends August 31 — a 50% rate step-up stacks with a ~30% tokenizer inflation effect; Anthropic also clarifies it has never sought to ban open-weight models.

Sonnet 5 cost cliff: 30 days to audit your pipelines

Claude Sonnet 5 moves from introductory pricing ($2/$10 per MTok input/output) to standard rates ($3/$15) on September 1 — a flat 50% increase. The compounding factor: Sonnet 5 uses a new tokenizer that produces roughly 30% more tokens from the same text as Sonnet 4.6. Per-token rates are unchanged, but 30% more tokens means 30% more spend at any rate. Combined, a typical workload that was $1,000/month at Sonnet 4.6 pricing may land at $1,850+ after September 1.

Three moves before August 31:

  1. Route simpler tasks to Haiku 4.5. Summarization, classification, and extraction don't need Sonnet-class reasoning. Haiku 4.5 costs roughly 7× less per token and uses the same new tokenizer — you get the full cost baseline without overpaying for capability.

  2. Enable prompt caching on repeated context. System prompts, RAG passages, and policy documents sent in every request qualify. Cache reads are ~90% cheaper than full input tokens, and that discount persists regardless of when the price hike lands.

  3. Contact Anthropic sales by August 15 if your monthly API spend exceeds ~$5K. Volume commitments negotiated before the rate change can lock in introductory-tier pricing.

Anthropic clarifies open-weights stance

A July 27 statement from CEO Dario Amodei clarified Anthropic has never advocated banning open-weight models. Its concerns target industrial-scale distillation of frontier models by state actors and chip exports to adversarial nations — not the open-source ecosystem. For engineers using Llama 4, Qwen 3, or Mistral models: no policy change is coming that affects your use of those models.

Why it matters: The cost cliff is real and the countdown is compressed — 30 days to plan and route before the hike lands.

Sources: Claude pricing — Anthropic · Sonnet 5 intro pricing deadline — FinOps LLM · Our position on open-weights models — Anthropic