Claude Haiku 5.5 is Anthropic's cheapest model yet, at $0.10 per million input tokens
Haiku 5.5 replaces Haiku 4.5 with up to 90% lower pricing, a 1M-token context window, and adaptive thinking.
On 7 October 2026, Anthropic released Claude Haiku 5.5, the third model in the Claude 5.5 family following Opus 5.5 and Sonnet 5.5. It replaces Haiku 4.5 as the entry-level option in the Claude line-up, and the headline number is hard to ignore: for prompts under 100,000 tokens, input costs $0.10 per million tokens. That is a 90% reduction from Haiku 4.5’s $1.00 per million.
The model is available immediately on the Anthropic API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
What it costs
Pricing is tiered by prompt length, which is new for the Haiku line.
| Haiku 5.5 (under 100K tokens) | Haiku 5.5 (over 100K tokens) | Haiku 4.5 | |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache reads | $0.01 | $0.05 | $0.10 |
| Cache writes (5 min) | $0.125 | $0.625 | $1.25 |
| Batch input | $0.05 | $0.25 | $0.50 |
| Batch output | $0.25 | $1.25 | $2.50 |
Anthropic says roughly 90% of requests to the previous Haiku model were under 100,000 tokens, so the lower tier applies to the vast majority of real-world usage. The average saving across short and long prompts works out to about 75% lower than Haiku 4.5.
One caveat worth knowing: Haiku 5.5 uses a new tokeniser that is roughly 25% less efficient than the one in Haiku 4.5. The same prompt that consumed 1,000 tokens before will consume around 1,250 tokens now. That partially offsets the price cut for longer or more complex inputs, so if you are migrating a high-volume pipeline, benchmark your actual token counts before assuming the full headline saving.
For context on competitive positioning: Haiku 5.5 matches GPT-6 Luna’s pricing exactly at $0.10/$0.50 for short prompts. Luna’s price increase kicks in at 272,000 tokens and rises only to $0.20/$0.75, whereas Haiku’s jumps to $0.50/$2.50 above 100,000 tokens. For very long-context work, Luna is currently cheaper at scale.
What’s new technically
Beyond pricing, Haiku 5.5 ships with a substantially expanded context window. The limit moves from 200,000 tokens in Haiku 4.5 to 1 million tokens, and maximum output grows from 64K to 128K tokens.
The bigger capability addition is adaptive thinking with effort controls, a first for any Haiku model. Thinking is on by default at medium effort. You can dial it up or down across five levels: low, medium, high, xhigh, and max. Lower effort means faster and cheaper responses; higher effort gives the model more processing budget for harder tasks. This is the same general mechanism available in Sonnet 5.5 and Opus 5.5, now brought down to the cheapest tier.
The model ID is claude-haiku-5-5, knowledge cutoff is June 2026, and it sits within Anthropic’s standard Claude family rate limits.
How it performs
Anthropic’s benchmarks show large jumps over the previous generation. On OSWorld 2.1 (a test of computer-use tasks), Haiku 5.5 scores 72.4% on the offline subset against 15.7% for Haiku 4.5. On Terminal-Bench 4.0, which covers multi-step command-line work, it reaches 39.2% compared with 0.0% for its predecessor. On Humanity’s Last Exam, it hits 57.4% with tools against 18.7% for Haiku 4.5.
Independent benchmarking platform Artificial Analysis ranks it first among small-class models on its Intelligence Index, ahead of GLM-5.3 Flash, Gemini 3.8 Flash, and GPT-6 Luna.
Early customer results from Anthropic’s launch partners are consistent with those numbers. HubSpot reported a 92.8% score on its CRM task evaluation suite, the highest it has seen, with Haiku 5.5 completing a stale-record audit faster than any other model tested and at the lowest false positive rate. Asana saw over 30% lower latency for task completions and up to 2.5x faster inference per agent turn across its AI Teammates evaluation covering bug triage, project set-up, and portfolio search. Box found Haiku 5.5 scored 11 points higher than Haiku 4.5 at roughly half the latency.
What this means for you
If you are running high-volume, cost-sensitive workloads, this is a meaningful update. Summarisation, classification, database queries, document compaction, and live customer support are all use cases Anthropic explicitly positions Haiku 5.5 for. At $0.10 per million input tokens, tasks that were borderline economical on Haiku 4.5 become straightforward.
For teams building agentic systems, Anthropic suggests using Haiku 5.5 as a subagent alongside Sonnet 5.5 or Opus 5.5, handling the high-frequency, lower-complexity steps while larger models take on orchestration or reasoning-heavy tasks. The effort controls let you tune each request individually rather than committing the whole pipeline to one level of compute.
For SaaS builders, the 1M-token context window opens up whole-document or large-codebase processing at a price point that was not available before in the Claude family at this tier.
The safety picture has also improved. The system card reports that indirect-injection attack success (a common prompt injection vector) fell from 83.2% on Haiku 4.5 to 7.1% on Haiku 5.5 in testing with up to 15 attacker attempts. The model does not cross Anthropic’s CB-2 or Autonomy-2 thresholds under its Responsible Scaling Policy.
Also shipping alongside this release
Anthropic is halving the price of cache reads on Claude Sonnet 5.5 at the same time, making Sonnet 5.5 roughly 20% cheaper on most agentic workloads. Claude Max and Team subscribers on the Claude platform are also receiving a new monthly API credit allocation. Neither change requires any action; existing API integrations pick up the lower cache-read pricing automatically.