Claude Sonnet 5 Is the New Claude Code Default: Recount Your Token Budgets

Claude Sonnet 5 (claude-sonnet-5) landed on June 30, 2026, and it is not a preview you have to opt into. On the Anthropic API the sonnet alias now resolves to it, and in Claude Code it is the account default for Pro, Team Standard, and Enterprise subscription seats. That means a lot of people are already running it without changing a line of config. The catch is that “drop-in upgrade” does not mean “same token math”, and if you budget cost or size max_tokens by hand, the numbers you tuned for Sonnet 4.6 are now wrong.

The tokenizer changed under you

Sonnet 5 ships a new tokenizer. The same input text produces roughly 30% more tokens than on Sonnet 4.6. Per-token pricing is unchanged, $3/$15 per million input/output tokens (with an introductory $2/$10 in effect through August 31, 2026), but 30% more tokens for identical text means the cost of an equivalent request goes up even though the rate card did not.

Three things you measure in tokens all shift:

• Per-request cost. Any estimate derived from a Sonnet 4.6 token count is now low.
• max_tokens budgets. An output limit sized close to your expected response can truncate on Sonnet 5, because the same reply spends more tokens.
• Context capacity in text terms. The window is 1M tokens, but each token covers less text on average, so the same window holds less prose.

Do not extrapolate. Recount against the model with the count-tokens endpoint before you trust any budget:

curl https://api.anthropic.com/v1/messages/count_tokens \
  --header "x-api-key: $ANTHROPIC_API_KEY" \
  --header "anthropic-version: 2023-06-01" \
  --header "content-type: application/json" \
  --data '{
    "model": "claude-sonnet-5",
    "messages": [{ "role": "user", "content": "Summarize this build log." }]
  }'

Run the same payload against claude-sonnet-4-6 and compare input_tokens. That delta is your budget correction.

Three API behaviors that now 400

If you call Sonnet 5 directly, three requests that worked on Sonnet 4.6 now return 400:

• Sampling parameters. Setting temperature, top_p, or top_k to a non-default value is rejected. Remove them and steer with the system prompt instead.
• Manual extended thinking. thinking: {"type": "enabled", "budget_tokens": N} is gone. Use adaptive thinking, which is on by default.
• Assistant prefill. Still unsupported, unchanged from Sonnet 4.6.

What it means inside Claude Code

Sonnet 5 requires Claude Code v2.1.197 or later; run claude update if sonnet still resolves to 4.6. On the Anthropic API it always runs with the native 1M context window, with no [1m] suffix and no usage credits, and sessions auto-compact at about 967K tokens. If you need a hard 200K cap for cost control, set CLAUDE_CODE_DISABLE_1M_CONTEXT=1. And if you want to control exactly when your team moves versions, pin the full ID rather than riding the alias:

{
  "model": "claude-sonnet-5"
}

The migration itself is genuinely a one-line model-ID swap. The work is in the budgets around it. Full details are in the Sonnet 5 release notes.