I'd give Haiku 5.5 small, repeated jobs with an answer you can check quickly, such as sorting email or pulling one figure from a long document. Sonnet 5.5 stays my pick for scoped deliverables and Opus 5.5 for judgement.
The benchmark scores and customer quotes come from Anthropic, and VentureBeat reads its chart as showing a much lower Terminal-Bench score at the default medium effort. None of the sources says how Haiku 5.5 counts against usage limits on Free, Pro, Max or Team plans, or which effort level Claude.ai uses for it. Sources checked 8 October 2026.
Repeated runs on real jobs where Haiku 5.5 needs more correction than Sonnet 5.5 would narrow where I'd use it. Anthropic publishing how Haiku 5.5 counts against plan usage limits, or independent scores at the default medium effort, would change how much weight the price cut deserves.
I'd give Claude Haiku 5.5 the small jobs you repeat and can check at a glance, like sorting a pile of email into folders or pulling one figure out of a long document. Sonnet 5.5 stays my pick for scoped deliverables such as a deck or a review, and Opus 5.5 for work where you still have to decide what matters.
That adds a rung below the split in Sonnet 5.5 for scoped Claude work, Opus for judgement, published on 2 October 2026, without changing it. What follows draws on Anthropic's launch material and docs, plus VentureBeat's report, all checked 8 October 2026.
What Anthropic released on 7 October 2026
Anthropic pitches Haiku 5.5 at high-volume, cost-sensitive work such as summaries and classification, and as a helper that bigger models hand smaller tasks to. Its Haiku page says Free, Pro, Max, Team and Enterprise users can select it on Claude.ai, on the web and in the mobile apps, and that it's in Claude Code, which Anthropic's pricing page lists for paid plans only. Developers call it as claude-haiku-5-5 on the Claude API, Amazon Bedrock, Google Cloud or Microsoft Foundry.
In the API docs it gets a 1M-token context window and up to 128K tokens of output, up from 200K and 64K on Haiku 4.5. It's also the first Haiku with an effort setting, which trades answer quality against speed and cost, and the API default is medium.
Who gets the price cut
On the Claude API, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that, the rates are $0.50 and $2.50. Haiku 4.5 costs $1 and $5, and Sonnet 5.5 costs $2 and $10. That's a 90% cut on Haiku 4.5's prices for shorter prompts and 50% for longer ones. Anthropic puts the average saving at about 75% once two things are allowed for: 90% of Haiku 4.5 requests fell under the 100,000-token line, and the new tokeniser counts the same text as roughly 30% more tokens. VentureBeat's Carl Franzen points out that the lower rate matches OpenAI's GPT-6 Luna.
On a Pro or Max subscription the monthly price doesn't move, and Haiku 5.5 is one more option in the model picker. Anthropic's pricing page says how far your usage goes depends partly on the model you choose, but neither it nor the launch page says how Haiku 5.5 counts against the limits on Free, Pro, Max or Team plans. So I wouldn't switch to it hoping to make an allowance last longer. Max and Team subscribers do get something new: on 7 October 2026 Anthropic said it would add a monthly API credit ($100 on Max 5x, $200 on Max 20x, up to $500 pooled across a Team plan) that works on any model, Haiku 5.5 included. Enterprise is the plan where the per-token price could show up in the bill, because Anthropic's pricing page says Enterprise usage is billed at API rates.
If you already call Haiku 4.5 through the API, it hasn't been deprecated: Anthropic's deprecations page lists it as Active on the Claude API, with retirement not sooner than 15 October 2026, and promises at least 60 days' notice before a public model is retired. The migration guide covers the code changes, including errors on non-default temperature settings and on assistant prefill. Priority Tier isn't available on Haiku 5.5.
Where Anthropic's own numbers favour Sonnet
On Anthropic's Terminal-Bench 4.0 figures, a test of multi-step work at a command line, Haiku 5.5 scores 39.2% against 70.6% for Sonnet 5.5. Haiku 4.5 scored 0.0%, so it's a big step up that still leaves a wide gap. The gap is narrower on OSWorld 2.1 (offline subset), which measures operating a computer through long tasks: Anthropic reports 72.4% for Haiku 5.5 and 83.9% for Sonnet 5.5. Anthropic's launch page says "Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks". VentureBeat reads Anthropic's effort chart as placing the 39% Terminal-Bench score at maximum effort, with the default medium setting at about 20%.
That's why I'd keep Haiku 5.5 to jobs where a wrong answer is quick to spot. A misfiled email, or a figure that doesn't match the page it came from, shows up as soon as you look. A weak argument in a board paper can read well and still be wrong, and that's where Sonnet 5.5 or Opus 5.5 earns its place. Summarising your own meeting notes sits on Haiku's side, because you were there and can tell when something's missing.
Trying it on a job you already do
Choose a repeat job where you already know what a right answer looks like, then give the same input to Haiku 5.5 and Sonnet 5.5 from the Claude.ai model picker and note what you had to correct in each. If you use the API, record the effort setting too, since VentureBeat's reading suggests it moves results a long way. Next time that job comes round, keep Haiku 5.5 for it only if your fixes stay small.
Sources
Anthropic, Introducing Claude Haiku 5.5 (7 October 2026). Anthropic, Claude Haiku. Anthropic, Claude Haiku 5.5 overview and What's new in Claude Haiku 5.5. Anthropic, Claude Haiku 5.5 migration guide. Anthropic, Model deprecations. Anthropic, Claude plans and pricing. VentureBeat, Carl Franzen, Anthropic launches Claude Haiku 5.5 with 90% API price reduction, matching GPT-6 Luna (7 October 2026). cliffinkent, Sonnet 5.5 for scoped Claude work, Opus for judgement. Checked 8 October 2026.