Claude Sonnet 5 changes thinking defaults and token counts

What breaks

A model-ID-only upgrade from Sonnet 4.6 can cause request errors and materially change token budgets, context calculations, and cost estimates.

• Manual extended thinking with budget_tokens returns an HTTP 400 error on Claude Sonnet 5.
• Non-default temperature, top_p, or top_k values return an HTTP 400 error.
• The same text produces approximately 30% more tokens than with the previous tokenizer, depending on content and workload.
• Adaptive thinking is enabled by default.

Affected identifiers and fields

Search your code, environment variables, gateway configuration, and stored request templates for these exact values.

• claude-sonnet-5
• claude-sonnet-4-6

• thinking
• budget_tokens
• adaptive thinking
• temperature
• top_p
• top_k
• tokenizer
• token counting

Migration target

Use Claude Sonnet 5 with adaptive thinking, omit manual thinking budgets and sampling fields, and remeasure token usage.

• claude-sonnet-5

What to do

Remove manual thinking budgets and non-default sampling parameters, rely on adaptive thinking, and use Anthropic's token-counting API to recalibrate limits and cost controls.

Before and after

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=4096,
    thinking={"type": "enabled", "budget_tokens": 2048},
    messages=[{"role": "user", "content": "Analyze this dataset."}],
)

response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Analyze this dataset."}],
)

Official source receipts

These provider pages support the dates, identifiers, symptoms, and migration direction recorded above.