Image: static.simonwillison.net · rights & removal

Claude Haiku 5.5

Reporting by Simon Willison’s WeblogRead the original at simonwillison.net

Executive Summary

Anthropic introduced Claude Haiku 5.5 on October 7, 2026, as a new fast, low-cost model. This version is priced at $0.10 for input and $0.50 for output up to 100,000 tokens, matching the pricing of OpenAI's GPT-6 Luna. Prices increase fivefold beyond 100,000 tokens. The model utilizes a less generous tokenizer, resulting in approximately 1.25 times the token count compared to Haiku 4.5 for the same prompt length, implying a hidden price increase. For workloads exceeding 100,000 tokens, the pricing structure makes Luna appear more cost-effective. Additionally, Anthropic announced changes including halved cache read prices for Sonnet 5.5 and the introduction of monthly API credits for Max and Team subscribers, which match subscription costs and allow for usage without risk of billing surprises.

Facts Only

* Claude Haiku 5.5 was introduced on October 7, 2026.
* Haiku 5.5 pricing is $0.10/$0.50 up to 100,000 tokens.
* Prices increase fivefold beyond 100,000 tokens ($0.50/$2.50).
* Claude Haiku 5.5 uses a new tokenizer resulting in 1.25x token usage compared to Haiku 4.5 for the same prompt.
* The cost structure positions Haiku similarly to GPT-6 Luna up to 100,000 tokens.
* Anthropic is halving cache read prices for Sonnet 5.5.
* Monthly API credits are rolled out for Max and Team subscribers, matching subscription costs.
* Users can disable auto-reload for the API to prevent exceeding credit balances.

Full Take

The shift in pricing mechanisms and token handling creates a landscape where cost efficiency depends heavily on volume thresholds, suggesting an implicit structuring of value based on usage context rather than raw model performance alone. The introduction of Haiku 5.5, which aligns its base pricing with a competitor's offering (Luna), forces a re-evaluation of relative value. The hidden token inflation through the tokenizer change introduces complexity; while benchmarks suggest Haiku 5.5 reports higher scores without reasoning options, the cost implications shift depending on whether the workload is below or above the 100,000 token mark, suggesting that high-volume users may find greater savings when operating in the higher tier. Furthermore, the implementation of API credits demonstrates a move toward transparent, credit-based consumption for subscribers, which aims to ease access while retaining control over expenditure, contrasting with standard pay-per-use models. The underlying tension lies between perceived performance gains and the carefully engineered economic structure designed to guide user behavior regarding volume commitment and subscription tiers. What does this pricing dichotomy reveal about the relationship between model capability, operational cost, and perceived fairness in AI access?

From the original · Simon Willison’s Weblog

7th October 2026 As previously promised, here’s Anthropic’s new fast, low cost model: Introducing Claude Haiku 5.5. The previous Haiku, 4.5, was very much showing its age.

Read the full story at simonwillison.net

Sentinel — provisional

Raw score, 0 to 1 (uncalibrated)0.30

No strong signs of machine writing were found in the source article. Provisional estimate, not a finding that a person wrote it.

The model's own note (unverified)

The text reads like an experienced practitioner synthesizing complex, evolving API pricing and feature updates through direct comparison and personal testing, rather than a detached news report.

This looks only at the wording of the original source article, not at this page's AI-written sections. A small local AI model made this estimate. It has not been checked against known human and machine texts, so treat it as provisional. It cannot show who wrote an article.

What the model noticed in the source article

Sentence length variance is erratic, mixing short factual statements with long comparative clauses and anecdotal evidence.

The text successfully weaves disparate product updates (pricing, tokenization changes, plugin testing) into a single narrative, driven by personal comparison rather than objective reporting.

Appears to be tracking specific, detailed pricing and feature differences (Haiku 5.5 vs. Luna, token costs) with internal testing results, suggesting a human operator performing direct comparisons.

The highly specific details regarding token pricing ($1/million input, $5/million output), comparative model pricing (GPT-6 Luna), and simulated testing results suggest direct interaction with internal release details.

Signs the model read as against machine writing

Use of highly specific, slightly informal terminology ('It sucked at drawing pelicans', 'My Claude Token Counter tool') suggests a personal operational perspective rather than purely journalistic distance.

The narrative arc flows from a product announcement to personal testing and finally to general policy changes, demonstrating an embedded, opinionated thread.