TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Anthropic introduced Claude Haiku 5.5 on October 7, 2026, positioning it as a fast, lower-cost model. Its listed input and output rates match OpenAI GPT-6 Luna up to 100,000 tokens, but rise above that threshold; independent benchmark comparisons and broader performance claims are not supplied in the report.
Anthropic announced Claude Haiku 5.5 on October 7, presenting it as a fast, lower-cost model and setting its listed API rates at $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens. The release also brings a higher price tier for longer requests and arrives with separate changes to Sonnet 5.5 cache pricing and API credits for some subscribers.
According to Simon Willison’s October 7 report, Haiku 5.5’s rates up to 100,000 tokens match those of OpenAI’s GPT-6 Luna. Above that request length, Haiku 5.5’s rates rise fivefold, to $0.50 per million input tokens and $2.50 per million output tokens. Willison reported that Luna’s rates rise above 272,000 tokens to $0.20 and $0.75 per million, respectively. The two models’ prices therefore depend on request length as well as token counts.
Willison also found that Haiku 5.5’s tokenizer used about 1.25 times as many tokens as Haiku 4.5 for the same long prompt in his Token Counter tool. That observation suggests listed rates alone may not capture the full cost of a given workload. It is an individual tool-based comparison, not a published measurement across all prompts or uses.
In a hands-on test, Willison asked Haiku 5.5 to generate an SVG of a pelican riding a bicycle. He said the low-effort response took seven seconds and cost 0.0936 cents; a maximum-effort response took five minutes and nine seconds and cost 3.3826 cents. The model does not let users disable reasoning, he reported, and defaults to medium effort. These are results from one prompt and do not establish typical response times, costs or quality.
How Token Length Changes the Price
The release gives developers and organizations another lower-priced model option, but the advertised rate is not a single price for every request. For workloads that stay below 100,000 tokens, Haiku 5.5 matches Luna’s listed rates. Longer requests cross into a substantially higher Haiku price tier, while Willison’s reported tokenizer result suggests some prompts may consume more billable tokens than they did with Haiku 4.5.
Those details matter for teams estimating API bills or choosing a model for long documents and other large prompts. A fair cost comparison needs to account for input and output usage, request length, and the number of tokens each model assigns to the same material. Willison wrote that Haiku 5.5 reported higher benchmark scores than Luna in his comparison, but the supplied report does not identify the benchmarks or provide results that would support a broader performance conclusion.
Anthropic also said it would halve the price of Sonnet 5.5 cache reads and add monthly API credits to certain subscriptions. Together, the changes affect both pay-as-you-go use and the value subscribers may get from using Anthropic’s API. The credits are not a general discount for every customer: eligibility and amounts vary by plan, and unused credits do not carry forward.
As an affiliate, we earn on qualifying purchases.
Haiku 4.5’s Pricing Baseline
Haiku 4.5, the preceding model discussed in Willison’s report, had been released almost a year earlier. Its listed API prices were $1 per million input tokens and $5 per million output tokens. Willison described those rates as relatively expensive at the time and said they were ten times Luna’s then-listed rates. Haiku 5.5’s lower rates up to 100,000 tokens mark a change in the price comparison, though the longer-request tier and tokenizer differences complicate a direct comparison.
Willison’s pelican test offers a limited illustration of the change between versions. He reported that Haiku 4.5 produced a weaker bicycle illustration at a cost of 0.7583 cents, and noted that the older model did not support reasoning-effort levels. The comparison is based on a single creative task; it should not be read as a systematic evaluation of the two models.
The report also describes Anthropic’s subscription changes. Max subscribers were set to receive monthly API credits of $100 for Max 5x and $200 for Max 20x; Team subscribers were to receive up to $500 pooled across users. Subscribers must select the API organization that receives the credits in billing settings. Willison said the amounts match the subscription prices, and that the credits expire monthly. Anthropic also added an option to disable automatic API balance reloads, so requests stop when the balance runs out.
“The new Haiku exactly matches the price of GPT-6 Luna—$0.10/$0.50—up to 100,000 tokens.”
— Simon Willison, in his October 7 report
As an affiliate, we earn on qualifying purchases.
Questions About Cost and Performance
The supplied report does not include Anthropic’s full model announcement, a model card, independent benchmark tables or broad testing results. It is not clear from this material how Haiku 5.5 performs across different tasks, how its response speed varies under normal use, or how the tokenizer comparison changes across languages and prompt types.
It is also unclear how the models’ billing handles requests that approach or cross the 100,000-token threshold in practice; the report gives the stated rates but no billing examples for mixed-length usage. The exact availability and eligibility details for subscriber credits beyond the plans described are not provided. The announcement’s claims and one reviewer’s tests should not be treated as a substitute for independent, workload-specific cost and quality comparisons.
large document AI processing tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Watch Rollout and Independent Tests
Anthropic said the monthly API credits for Max and Team subscribers would roll out during the week of October 7. Eligible customers can select the API organization that receives the credit through billing settings; credits expire at the end of each monthly period. Subscribers planning to use them can also turn off automatic reloads, according to Willison’s account, to stop API requests when the balance is exhausted.
For developers weighing Haiku 5.5 against alternatives, the next useful evidence will be testing against their own prompts and usage patterns, including tokenization, request lengths above 100,000 tokens, and reasoning settings. Further independent evaluations or fuller documentation from Anthropic could clarify model capabilities, availability and how the new rates apply across different workloads.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Anthropic announce?
Anthropic announced Claude Haiku 5.5 on October 7, 2026. The announcement was also accompanied by a reported reduction in Sonnet 5.5 cache-read prices and monthly API credits for eligible Max and Team subscribers.
How much does Haiku 5.5 cost?
As reported by Simon Willison, requests up to 100,000 tokens cost $0.10 per million input tokens and $0.50 per million output tokens. Above 100,000 tokens, the listed rates increase to $0.50 and $2.50 per million, respectively.
Is Haiku 5.5 cheaper than GPT-6 Luna?
The report says their listed rates match up to 100,000 tokens. Above that point, Haiku 5.5’s reported rates are higher; Luna’s rates increase after 272,000 tokens to $0.20 per million input tokens and $0.75 per million output tokens. Tokenizer differences may also affect total cost.
What API credits did Anthropic announce?
The reported monthly credits are $100 for Max 5x, $200 for Max 20x, and up to $500 pooled across Team users. They do not roll over, and subscribers must choose the API organization that receives them in billing settings.
Has Haiku 5.5’s performance been independently established?
The supplied report contains a reviewer’s tests and benchmark comparison but no benchmark tables or broad independent evaluation. Its results should be treated as limited observations, not proof of performance across tasks.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
