Claude Haiku 5.5 costs 90% less per token, but a new tokenizer eats part of the saving

Anthropic released Claude Haiku 5.5 on 7 October 2026 and called it “the cheapest, fastest, and most capable small model we’ve ever released,” in its launch post. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, against $1 and $5 for Haiku 4.5.

That is a 90% cut on paper. The fine print, which MIXED News read closely a day later, shows why Anthropic’s own average is lower.

Three numbers for the same discount

The launch post uses two figures. One is “90% lower” for requests up to 100,000 tokens. The other is about 75% less to run on average. MIXED News points to Anthropic’s footnote for the gap: 90% of Haiku 4.5 requests fell under the 100,000-token mark, and the new tokenizer uses more tokens per task.

The tokenizer is the part to understand. The launch post says Haiku 5.5 uses “slightly more” tokens than Haiku 4.5. Anthropic’s developer documentation, as MIXED News reports it, says the same text produces approximately 30% more tokens, depending on content, and tells developers to recount their prompts and revisit max_tokens values and cost estimates.

A rough calculation shows what that does. If a prompt that cost $1 on Haiku 4.5 now counts as 1.3 times as many tokens at one tenth of the price, the bill is about $0.13, not $0.10. That is still a large saving, but it is closer to 87% than 90%, and that is before the second catch.

The 100,000-token cliff

Above 100,000 prompt tokens, the rate rises to $0.50 input and $2.50 output per million tokens. Cache reads go from $0.01 to $0.05. The higher rates apply to the whole request, not only the tokens past the threshold.

For many workloads this does not matter. Summaries, classification, database queries and compaction of short conversations sit well under the line. But anyone sending long documents or full agent transcripts to a small model should check where their requests land. A request of 110,000 tokens costs five times as much per token as one of 99,000.

Anthropic is clear about what the model is for: high-volume, cost-sensitive work such as summaries, compaction, database queries and classification, and as a subagent next to Opus 5.5 and Sonnet 5.5. It says Sonnet 5.5 or Opus 5.5 remain the better choice for complex agentic coding.

What else changed

Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, from Low through Max, with medium as the default. MIXED News lists a 1 million token context window, up from 200,000, and a maximum output of 128,000 tokens, up from 64,000. The knowledge cutoff is June 2026, and Anthropic has committed to keep the model available until at least 7 October 2027.

On Anthropic’s own benchmarks the gains over Haiku 4.5 are large:

• Terminal-Bench 4.0: 39.2%, up from 0.0%.
• Humanity’s Last Exam without tools: 45.9%, up from 10.2%.
• OSWorld 2.1, offline subset: 72.4%, up from 15.7%.
• GDPval-AA v2.1: 1620 Elo, up from 735.

The company also cites customers. Asana reported more than a 30% latency reduction, Box reported a score 11 points higher at about half the latency, and AlphaSense reported 0.84 against 0.76 on 400 queries. These are numbers selected by the vendor and by partners who chose to share them, so treat them as leads for your own tests. Anthropic notes that Sonnet 5.5 scores higher on every benchmark listed.

There are limits worth knowing. Cybersecurity safeguards are stricter than on Haiku 4.5 and still block penetration testing work. Organizations that need broader cyber or biology access are directed to Anthropic’s verification programs.

The rest of the price list

Alongside the launch, Anthropic halved the cache-read price of Sonnet 5.5 to $0.10 per million tokens, which it says makes most agentic work about 20% cheaper. Sonnet 5.5 stays at $2 input and $10 output. Opus 5.5 is $4 and $20, per MIXED News.

Subscribers got something too: monthly API credits of $100 for Max, $200 for Ultra, and up to $500 pooled for Team, with the rollout starting that week.

What to do with it

The practical move is to recount, not assume. Take a sample of your real prompts, run them through the Haiku 5.5 tokenizer, and compute cost at both the under-100K and over-100K rates. Then compare against Sonnet 5.5 with the new cache price, because for cache-heavy agent loops that cheaper cache read may close much of the gap.

For short, repetitive jobs, Haiku 5.5 looks like a clear saving even after the tokenizer effect. For long-context work the answer depends on where your prompts sit relative to the 100,000-token line, and that is a number only you can measure.

Not yet known: how the tokenizer difference varies across languages and code, since Anthropic says only that the increase depends on content.