Claude Haiku 5.5 was announced on 7 October 2026. To understand a price change, start with the billable units and a workload you can describe. This worksheet uses prices checked on 10 October and a deliberately small, synthetic token budget. It does not rank model quality or predict an application’s invoice.
The workload and billing scope
Assume 20 independent text requests, each billed for 2,500 input tokens and 200 output tokens: 50,000 input tokens and 4,000 output tokens overall. These are assumed billable counts, not measurements of the same text on different tokenizers. Each request is short enough for the short-context rates shown below.
The comparison uses direct provider APIs, ordinary on-demand processing, and global processing rather than a regional processing surcharge. OpenAI uses Standard mode. There is no prompt caching, Batch discount, negotiated agreement, paid tool use, storage or additional service in this calculation. Prices are in USD before taxes and currency conversion. Check the relevant cloud tariff separately if you use Bedrock, Vertex AI or another managed service.
Published prices and calculated totals
| Model | Input USD per million | Output USD per million | Input cost | Output cost | Total USD |
|---|---|---|---|---|---|
| Claude Haiku 4.5 | 1.00 | 5.00 | 0.050 | 0.020 | 0.070 |
| Claude Haiku 5.5 | 0.10 | 0.50 | 0.005 | 0.002 | 0.007 |
| GPT-6.1 Sol, Standard | 2.00 | 10.00 | 0.100 | 0.040 | 0.140 |
Source rates: Anthropic’s official pricing table and OpenAI’s official API pricing. Haiku 5.5 has higher rates when an individual prompt exceeds 100,000 tokens. This example remains below that threshold. Sol’s Standard short-context row also applies to the assumed requests.
Input cost = billable input tokens / 1,000,000 × input rate
Output cost = billable output tokens / 1,000,000 × output rate
Example Haiku 5.5 = 50,000 / 1,000,000 × 0.10
+ 4,000 / 1,000,000 × 0.50
= 0.007 USDFor this synthetic budget, the Haiku 4.5 and 5.5 totals differ by 0.063 USD. That arithmetic does not establish equal quality, equal tokenization or the number of retries needed to complete a real task. Sol serves a different model position; including it gives another tariff example rather than a claim that the models are interchangeable.
Why the actual invoice can differ
| Variable | What to record | Why it matters |
|---|---|---|
| Billable token counts | Actual usage fields, including billable reasoning where applicable | The same visible text can generate different counts |
| Repeated attempts | Calls that completed, failed or were retried | A per-call estimate is not a completed-task estimate |
| Prompt size | Full input on each request, including accumulated history | A larger request can cross a context-price threshold |
| Processing choice | Provider, mode and global or regional endpoint | Another tariff can apply |
| Tools and storage | Searches, hosted execution and retained files used | Separate charges may sit outside token prices |
Do not assume every input token is a cache hit
This baseline deliberately excludes caching. If you add it, distinguish new input, cache writes and cache reads in the usage record. A repeated document is not evidence that the provider billed every repetition as a hit. Duration, matching prefixes and supported configuration matter; confirm the current provider documentation before adding a discount to a spreadsheet.
A worksheet for your own decision
- Choose a task with a written acceptance criterion, not just a cheaper model name.
- Record the provider, model ID, date of the tariff, processing mode and location.
- Collect billable usage for representative requests through an authorized workflow.
- Calculate token charges and add any separate services.
- Compare accepted results and retries, then review latency and operational constraints.
Use the weekly infrastructure review for the dated release context and the archive to follow later changes. Recalculate this worksheet when the tariff or the workload changes, and keep the assumptions beside the total.