By Claude Sonnet 5’s introductory pricing ends August 31. On September 1, input costs jump from $2 to $3 per million tokens and output costs from $10 to $15 — a clean 50% increase on both. But the real bill shock isn’t the price hike. It’s the new tokenizer Anthropic shipped with Sonnet 5, which turns the same input text into roughly 30% more tokens than Sonnet 4.6 did. The two effects compound. A workload that cost $1,000 a month on Sonnet 4.6 could hit $1,950 after September 1 — without touching a single line of code.
The Tokenizer Is the Hidden Multiplier
When Anthropic launched Sonnet 5, the headline was “cost-neutral.” That was technically true at the per-token level but misleading in practice. The new tokenizer doesn’t just swap in for the old one — it produces more tokens for the same text. Average overhead is around 30%, and some workloads see up to 35%.
You’ve been absorbing this cost since you migrated to Sonnet 5. On September 1, the price per token also goes up. Both hit at once. This is why re-running your Sonnet 4.6 cost benchmarks against Sonnet 5 — as the model, not just a name swap — matters right now.
The Numbers
| Model | Input (per 1M) | Output (per 1M) | Best for |
|---|---|---|---|
| Sonnet 5 (now) | $2 | $10 | Complex reasoning, code |
| Sonnet 5 (Sept 1+) | $3 | $15 | Complex reasoning, code |
| Haiku 4.5 | $1 | $5 | Fast, high-volume, simple tasks |
| Sonnet 4.6 | $3 | $15 | Legacy (still available) |
The worked math: a $1,000/month Sonnet 4.6 bill becomes $1,300 on Sonnet 5 at intro pricing (tokenizer overhead alone), then $1,950 after September 1 when the price increase lands on top. That’s a 95% effective cost increase from a nominal “50% price increase.”
Three Steps to Audit Your Bill Before September 1
1. Count Tokens on Your Actual Prompts
The token counting API is free. Don’t estimate — run it against your production prompts.
import anthropic
client = anthropic.Anthropic()
response = client.messages.count_tokens(
model="claude-sonnet-5",
messages=[{"role": "user", "content": your_production_prompt}]
)
print(f"Input tokens: {response.input_tokens}")Run this against your top five most expensive prompt templates. Cost follows a power law — fix the biggest ones first.
2. Check Your Cache Hit Rate
In early 2026, Anthropic quietly dropped the prompt cache TTL from 60 minutes to 5 minutes. Workloads that relied on the hour-long window saw 30–60% cost spikes overnight. If you haven’t checked your cache hit rate since then, check it now via the Anthropic Analytics API.
Cache reads cost 0.1× the standard input rate — a 90% discount. But caching only activates for blocks of at least 1,024 tokens. A 600-token system prompt won’t cache at all.
3. Route High-Frequency Shallow Tasks to Haiku 4.5
Haiku 4.5 is $1/$5 per million tokens — roughly 3× cheaper than post-September Sonnet 5. For routing, classification, simple summarization, and boilerplate generation, it performs well enough. Reserve Sonnet 5 for code generation, complex reasoning, and agentic workflows where the capability gap actually matters.
If You Can’t Optimize in Time
The Message Batches API cuts costs by 50% for async workloads — if your use case tolerates latency, this is the fastest lever to pull. It supports claude-sonnet-5 and requires no model changes, just switching the API endpoint.
Alternatively, stay on Sonnet 4.6. It’s still available and priced at parity with Sonnet 5 post-September 1. You lose the expanded context window and improved agentic performance — but you get a known cost baseline and no tokenizer overhead.
25 Days Left
This is how the model pricing game works in 2026: introductory pricing hooks you during migration, then standard pricing locks in once you’re dependent. It happened with GPT-5.6 Luna, Grok 4.6, and now Sonnet 5. Build token cost monitoring into your pipeline. The counting API is free. The only reason not to audit is assuming your old benchmarks still apply — and with a new tokenizer, they don’t.
The deadline is August 31.
