💡 Sonnet 5's September Price Cliff Is Bigger Than the List Price Suggests

Claude Sonnet 5's introductory pricing — $2 / $10 per million input/output tokens — expires on August 31, 2026. Standard pricing of $3 / $15 takes over on September 1. That's a 50% hike on both sides of the ledger. But here's the part most teams aren't accounting for: Sonnet 5 ships with a new tokenizer that inflates token counts by roughly 27–42% for identical input compared to Claude 4.x models. The effective cost of an unchanged workload is already higher than Sonnet 4.6 — before the rate card changes.

What this means in practice

• A workload that costs $100/day on Sonnet 4.6 might cost ~$130–140/day on Sonnet 5 during the promo window, purely from tokenizer inflation.
• On September 1, that same workload hits standard pricing on top of inflated token counts — the combined effect can push September's run rate to roughly 1.5–1.9× the pre-migration July baseline.
• Budgeting off August invoices will underestimate September costs, because August still benefits from the promotional rate masking part of the inflation impact.

How to measure tokenizer inflation for your workload

# Compare token counts for the same prompt on both models
import anthropic
client = anthropic.Anthropic()

prompt = "Your representative production prompt here..."

for model in ["claude-sonnet-4-6-20250904", "claude-sonnet-5-20260630"]:
    r = client.messages.create(
        model=model, max_tokens=1,
        messages=[{"role": "user", "content": prompt}]
    )
    print(f"{model}: {r.usage.input_tokens} input tokens")

Run this across your ten highest-traffic prompt templates to get a realistic inflation multiplier for your workload specifically — generic estimates (27–42%) hide wide variance by content type.

💡 What 400,000 Claude Code Sessions Reveal About Expertise and Output

Anthropic published a new research report — How Claude Code is used in practice — based on a privacy-preserving analysis of approximately 400,000 interactive sessions from ~235,000 people between October 2025 and April 2026. The headline finding is nuanced but directly actionable: the more domain expertise a person brings to a session, the more work Claude does per instruction. Experts don't just use Claude Code more — they unlock a qualitatively different mode of collaboration.

Key findings

• Work split: In a typical session, people make most of the planning decisions (what to do) and Claude makes most of the execution decisions (how to do it). This division holds across occupation types.
• Expertise multiplier: Domain experts give fewer, higher-level instructions and Claude runs longer, more autonomous execution chains per instruction. Non-experts tend to micro-direct, which limits Claude's leverage.
• Occupation parity on success rate: On coding tasks, every major occupation achieves nearly the same success rate (verifiable outcomes like passing tests or committed work) as software engineers — but experts achieve more per session.
• Implication: The bottleneck for most teams isn't Claude's capability — it's the quality and specificity of the direction given to it.

This is the same logic behind the shift from "give me code that does X" to "here is the acceptance criterion, here are the constraints — go." The research gives that intuition an empirical backing at scale.

💡 Two Days' Notice: Claude Opus 4.1 Retires on August 5

Claude Opus 4.1 (claude-opus-4-1-20250805) is being retired from the Claude API on August 5, 2026 — the day after tomorrow. If any of your production workflows, CI pipelines, or API integrations are pinned to this model ID, they will break in 48 hours. Anthropic's recommended migration target is Claude Opus 4.8.

What to update

• Search your codebase for the string claude-opus-4-1 — it may appear in config files, environment variables, or hardcoded strings in client initialisation.
• If you're on Claude Code and have an explicit model override set, check your .claude/settings.json for a pinned model key.
• Replace with claude-opus-4-8-20260101 (or the model alias claude-opus-4-8 if your SDK version supports it).

# Find Opus 4.1 references in your project
grep -r "claude-opus-4-1" . --include="*.json" --include="*.env" \
  --include="*.py" --include="*.ts" --include="*.js"