Haiku 5.5 Costs $0.10 In. Muse Spark's Budget Crown Is Gone

Claude Haiku 5.5 lists at $0.10 per million input tokens, which is 12.5 times cheaper than Muse Spark 1.3 at $1.25. Muse Spark 1.3 is the cheapest API card helloai tracks. On output the gap is smaller but still large: $0.50 against $4.25, about 8.5 times cheaper. Anthropic launched the model on Oct 7, 2026, and says it costs about 75% less to run on average than Haiku 4.5. The budget agentic slot used to have an obvious pick. Now the pick depends on how much capability a team will give up for price.

Every Haiku 5.5 figure here comes from press reports and aggregators, not from Anthropic's pricing page: Reuters via The Star, Forklog, Android Authority and AI Weekly. Those reports give it 1M context, 128K max output, the API id claude-haiku-5-5, and availability on AWS, Google Cloud and Azure. It is also the first Haiku with adjustable effort levels. Haiku 5.5 is not tracked on helloai and has no verified text-board row, so there is no Elo or index score for it here. Muse Spark 1.3 has 1M context and an Arena Elo of 1494, so this compares a measured model with one that so far has only a price.

List price is not task cost, and three things narrow the gap. First, the $0.10/$0.50 rate only applies to prompts under 100,000 tokens. Above that the price rises to $0.50/$2.50, and the input advantage over Muse Spark shrinks to 2.5 times. Agent loops that build up tool output and history cross that line quickly. Second, Forklog notes that the per-token price is about 90% lower than Haiku 4.5's $1/$5, but a new tokenizer cuts the real per-task saving to about 75%. Third, effort levels let callers spend more output tokens on harder steps, so the bill depends on how the model is configured as well as what it lists at.

The counterargument is capability. Haiku has always been Anthropic's speed-and-cost tier, and nothing verified yet shows how close 5.5 comes to a model rated 1494. Until a board row appears, teams have to run their own evals on their own workloads. Anthropic's price ladder in this cycle gives a reference point: Sonnet 5.5 at $2/$10 and Opus 5.5 at $4/$20, covered in our Oct 12 piece. Opus input costs 40 times Haiku input, which shows how differently Anthropic expects each tier to be used.

The bigger change will be in routing defaults. At $0.10 input, the cheapest reasonable choice is to send every request to Haiku 5.5 first and escalate to Sonnet, Muse Spark or Opus only when it fails a check. If a verified Elo lands near Muse Spark's, the budget default changes outright. If it lands well below, Haiku 5.5 still becomes the triage layer that decides which calls deserve a more expensive model, and that layer may end up handling most of a team's traffic.