Haiku 5.5 and the 100K price step - what the fast slot would payConcept

Claude Haiku 5.5 (shipped 10-7) is a tenth of Haiku 4.5's price below 100K tokens and half above it, with a 1M window. Your fast slot is still pinned to 4.5; the fallback alias already moved.

Prices per million tokens, from Anthropic's pricing page: Haiku 5.5 costs $0.10 in and $0.50 out for prompts up to 100,000 tokens, $0.50 and $2.50 beyond, cache reads $0.01. Haiku 4.5 is $1 in, $5 out. Sonnet 5.5 is $2 / $10 and Sonnet 4.6 is $3 / $15. A price with a step means the cheap rate holds only while the prompt stays under the line.

Two things change under the hood. The newer tokenizer counts about 30% more tokens for the same text, so a $0.10 rate is closer to $0.13 in old-token terms - still far cheaper. And adaptive thinking is on by default at effort medium; the Haiku 4.5 way of passing budget_tokens returns a 400 error. JARBUS's _claude_client.py passes neither, so a swap is one line in models.env.

Claude Code 2.1.293 (10-7) made Haiku 5.5 the default haiku. CLAUDE_FALLBACK_MODELS=sonnet,haiku uses aliases, so the fallback chain already lands on 5.5 when a primary is capped. Only the pinned id is behind.

Worked example

See the slot, see who reads it, try one run on the new model without changing the file.

cat /Users/tom/Claude/_config/models.env
grep -n "FAST_MODEL\|DEFAULT_MODEL" /Users/tom/Claude/JARBUS/scripts/_claude_client.py
grep -n "budget_tokens\|thinking\|tool_choice" /Users/tom/Claude/JARBUS/scripts/_claude_client.py || echo "none set - safe to swap"
cd /Users/tom/Claude/JARBUS/scripts && CLAUDE_FAST_MODEL=claude-haiku-5-5 python3 -c "from models import model; print(model('CLAUDE_FAST_MODEL'))"
claude --version

• models.env: five slots. The fast slot reads claude-haiku-4-5.
• _claude_client.py line 70: FAST_MODEL comes from JARBUS_FAST_MODEL, else the shared setting. Callers opt in by passing model=FAST_MODEL.
• No thinking or tool-choice parameters anywhere in the client, so Haiku 5.5's stricter rules cannot break a call.
• The env var wins over the file for one process: the resolver prints claude-haiku-5-5 while models.env is untouched.
• 2.1.296 tonight; 2.1.293 is where 'haiku' started meaning 5.5.

Gotchas

• The step is at 100,000 prompt tokens, counted with the new tokenizer; a long media transcript can cross it.
• Default effort is medium on Haiku 5.5; for rubric fills that is fine, for anything judgment-heavy set it explicitly.
• Caches are model-scoped: switching models forfeits cache reuse for that slot's prompts until they are re-cached.