Claude Haiku 4.5 for cheap tasks: check the multiplier first

What the id actually is

There is one Haiku id in this catalog: stable-claude-haiku-4-5, released 2025-10-01. It carries a 200,000-token context and a 64,000-token output ceiling, and its capabilities are text and tools — it is the only Claude id here without thinking.

Its official rate is $1 per million input tokens and $5 per million output, the lowest official rate of any Claude model in the catalog. But it is a stable-* id, which means a 0.4× multiplier rather than the 0.1× that the pool-lane Claude ids carry, so it bills at $0.40 and $2.00 per million.

That multiplier is the whole story of this page. A model that is cheap upstream is not automatically the cheap option here, and the only way to know is to multiply before choosing.

Against the pool lane, it is not the cheap option

claude-sonnet-5 is $3/$15 official at 0.1×, which bills at $0.30 and $1.50 per million. Haiku bills at $0.40 and $2.00. So on the numbers the small model is 1.33× the price of Sonnet 5 on both input and output, while carrying a fifth of the context, half the output ceiling, and no thinking.

It is under claude-opus-5, which bills at $0.50 and $2.50 — Haiku is 0.8× that. And claude-sonnet-4-6 matches Sonnet 5 at $0.30 and $1.50 with the same 200,000-token context Haiku has.

None of that makes Haiku a mistake; it makes "small model equals cheap" a claim to check rather than assume. If your reason for reaching for it is cost and your workload can run on the pool lane, the arithmetic points elsewhere.

for MODEL in stable-claude-haiku-4-5 claude-sonnet-5; do
  echo "=== $MODEL ==="
  curl -sS https://token-share.app/v1/messages \
    -H "x-api-key: $TOKEN_SHARE_KEY" \
    -H "anthropic-version: 2023-06-01" \
    -H "content-type: application/json" \
    -d "{
      \"model\": \"$MODEL\",
      \"max_tokens\": 512,
      \"messages\": [{\"role\": \"user\", \"content\": \"Classify this ticket as bug, feature, or question. Reply with one word.\"}]
    }" | jq '.usage'
done

Against the other dedicated lanes, it is the low number

The comparison where Haiku wins on price is within the stable-* family, since every stable Claude lane carries the same 0.4× multiplier and only the official rate differs. stable-claude-sonnet-5 bills at $0.80/$4.00 — 2× Haiku. stable-claude-sonnet-4-6 is $1.20/$6.00, or 3×. stable-claude-opus-5 and stable-claude-opus-4-8 are $2.00/$10.00, or 5×. The Fable lanes, stable-claude-fable-5 and stable-claude-fable-5-1, are $4.00/$20.00 — 10× Haiku on both.

So the question that puts Haiku in a good light is not "small or large" but "which dedicated lane". If your requirement is a lane served on our own key rather than the shared account pool, Haiku is between 2× and 10× cheaper than the alternatives in that group.

A stable-* id does not participate in pool failover — that is what the 0.4× buys, a separate path to the same weights, and it is a different question from which model you want.

What the missing thinking capability rules out

stable-claude-haiku-4-5 lists text and tools; the other eleven Claude ids list thinking as well. On the eleven, a task that fails on the first attempt can be given a thinking budget and retried. On Haiku there is no such lever — if the output is wrong, the next move is a different model, not a different parameter.

That draws a usable line. Work whose answer is determined by the input — classification, extraction, reformatting, routing a request to the right handler, a short summary of material already in the prompt — has no reasoning chain for a budget to improve, so the absent capability costs nothing.

Work that fails in ways more deliberation would catch is where the missing lever is felt, and it is also where a per-call price advantage of a few tenths of a dollar per million tokens stops being the deciding number.

Common questions