Claude Haiku 5.5: Token Prices Down 90%… But What Does It Actually Cost?
Small models are the workhorses you run cheap and often. This one's price tag is anything but ordinary, though. The short version: per-token prices really did drop to a tenth. Whether your bill drops to a tenth is another story, because three conditions in the fine print decide that. Here's what's real, what's conditional, and how it stacks up against its closest rival. Ribbit.
The short version
Price — $0.10 in · $0.50 out, a tenth of its predecessor (for prompts up to 100K tokens)
Performance — Tops GPT-6 Luna on every row of Anthropic's table; 43 on the independent AA Intelligence Index
The catch — A new tokenizer counts about 30% more tokens, and past 100K the price jumps 5x
Haiku 5.5 at a glance
- Anthropic's small model, released October 7, 2026. Built for high-volume repeat work like summaries, classification and database queries, and for serving as a subagent under bigger models.
- Context grew from 200K to 1M tokens, and a single answer can now run to 128K tokens, up from 64K. It reads text and images; its knowledge cutoff is June 2026.
- The first Haiku with an effort setting: five levels, from low to max, with medium as the default. You choose how long it thinks before answering.
- Also the first Haiku with built-in safeguards for high-risk cybersecurity requests. Defensive work is allowed, penetration testing is blocked, and a verification program can widen the scope.
- Available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Sources: AI Times · MarkTechPost
Side by side with its peers
Pricing (official price pages, checked Oct 15, 2026; per 1M tokens, USD)
| Model | Input | Output |
|---|---|---|
| Claude Sonnet 5.5 (bigger sibling) | $2 | $10 |
| Claude Haiku 4.5 (predecessor) | $1 | $5 |
| GPT-6 Luna | $0.10 | $0.50 |
| Claude Haiku 5.5 | $0.10 | $0.50 |
- One catch: past 100K prompt tokens, Haiku 5.5 jumps to $0.50 / $2.50 — five times more. Luna's pricier tier only starts around 272K tokens, so long requests can actually cost less on Luna.
- Cache reads are the cheapest in class at $0.01, and batch processing takes another 50% off.
Sources: Claude pricing · OpenAI pricing
Benchmarks (Anthropic's published table)
- GDPval-AA v2.1, a knowledge-work evaluation: 1620. Haiku 4.5 scored 735, GPT-6 Luna 1437, and the bigger Sonnet 5.5 leads at 1840.
- OSWorld 2.1, computer use: 72.4%, against 15.7% for its predecessor and 48.9% for Luna. Operating a desktop is where the gap opens widest.
- Terminal-Bench 4.0, command-line work: 39.2%. The predecessor scored 0.0%, so it went from solving none to solving about four in ten. Luna posts 16.4%, Sonnet 5.5 70.6%.
- Humanity's Last Exam: 45.9% without tools and 57.4% with tools, up from 10.2% and 18.7%.
- AA-Briefcase v1.1, document work: 1578, ahead of Luna's 1336. Chart reading lands at 46.4%, more than seven times the predecessor's 6.4%.
The cold water
- Three numbers, kept separate. The 90% cut is a per-token rate for prompts up to 100K tokens (input $1 to $0.10, output $5 to $0.50). Past 100K tokens, the cut is 50%.
- The new tokenizer counts the same text as roughly 30% more tokens (Anthropic's developer documentation). Cheaper tokens, more of them — the saving on a real bill narrows accordingly.
- "Around 75% less" is the average running cost stated in Anthropic's launch post. Its pricing table (90% lower up to 100K tokens, 50% lower above) and the real request mix (about 90% of Haiku 4.5 requests were under 100K tokens) sit behind it; the figure does not follow from the rate cut alone.
- The 100K boundary is a cliff. A 99K-token prompt costs about a cent to send; at 101K tokens it costs about five cents. Five times the price, right past the line.
- The flashy coding score is a max-effort number. Terminal-Bench 39.2% appears to be at the highest effort setting; at the default medium, analysis puts it around 20%. For a model meant to run cheap, peak performance costs thinking time.
- The benchmark table is Anthropic's own. Who measured Luna's scores, and under what conditions, isn't stated. The independent Artificial Analysis index does put Haiku 5.5 (43) ahead of Luna (38) — but the row-by-row numbers aren't independently verified.
Sources: Unite.ai · MarkTechPost · AI Times (AAII 43)
What Haiku 5.5 is for
- High-volume repetition: summaries, conversation compaction, classification, database lookups. Work that runs thousands of times a day, where the token price is the bill.
- Subagent work: Opus or Sonnet plans and directs, Haiku handles the grunt work. A Korean usage guide sums the pattern up as "Opus plans, Haiku processes."
- Real-time replies: it's the fastest model in Anthropic's lineup, aimed at live support, voice agents and in-app assistants where a slow answer is a failed answer.
- Computer use: the OSWorld score points at browser automation and screen-driven tasks — territory where a small model has rarely challenged the big ones.
Sources: Anthropic announcement · AI Matters guide
Already out in the field
- Asana reported task-completion latency down by over 30%, with inference up to 2.5x faster per agent turn.
- AlphaSense saw document Q&A accuracy rise from 0.76 to 0.84, a statistically significant improvement.
- Box reported scores 11 points higher than Haiku 4.5 at about half the latency.
- HubSpot was quoted at 92.8%, averaged over three runs of CRM evaluations. Cognition also took part in early testing.
- Worth noting: all of these are customer quotes inside Anthropic's announcement, not figures the companies published on their own channels.
Sources: AI Weekly · Moneycontrol
Announced the same day
- Sonnet 5.5 cache reads were halved, from $0.20 to $0.10 per million tokens. Anthropic says that cuts Sonnet's cost on most agent work by about 20%.
- Claude Max and Team subscribers now get monthly platform API credits: $100 on Max 5x, $200 on Max 20x, and up to $500 shared across a Team plan — usable on any model.
- The Claude Python and TypeScript SDKs gained beta computer-use and browser-use features, a natural pairing for Haiku 5.5's speed and price.
Source: Sol News
Puzz's one-liner
The tenth-price token rates are real, and so is the performance jump — on paper. Your bill, though, gets recalculated by the tokenizer and the 100K cliff. For high-volume short requests, it's the strongest small model going; for giant documents, check the cliff first. Ribbit.