Published: October 9, 2026 | Category: AI | By Mahesh
Disclosure: Anthropic makes Claude, the assistant that helped draft this article. Every figure below comes from the cited sources.
Anthropic released Claude Haiku 5.5 on 7 October 2026 and priced it at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens [1]. The previous Haiku, version 4.5, charged $1 and $5 [1]. That is a 90 percent cut on the headline rate, and Let's Data Science cites VentureBeat reporting that the price matches OpenAI's GPT-6 Luna [3].
What Changed
Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, which lets developers trade cost against intelligence per task [1]. Anthropic's charts show five levels: Low, Med, High, Xhigh and Max [1]. The model is available on Amazon Web Services, Google Cloud and Microsoft Azure under the ID claude-haiku-5-5 [1].
Prompts above 100,000 tokens cost $0.50 input and $2.50 output, still half the old Haiku rate [1]. Anthropic also halved the cache-read price of Claude Sonnet 5.5, from $0.20 to $0.10 per million tokens [1]. Cache reads on Haiku 5.5 itself cost $0.01 per million tokens at the lower tier [1].
The Benchmark Table
Anthropic's own comparison puts Haiku 5.5 at 72.4 percent on the offline subset of OSWorld 2.1, a computer-use test, against 15.7 percent for Haiku 4.5 and 48.9 percent for GPT-6 Luna [1]. On Terminal-Bench 4.0 it scores 39.2 percent, Luna scores 16.4 percent and Haiku 4.5 scores 0 [1]. On Humanity's Last Exam it reaches 45.9 percent without tools and 57.4 percent with them [1].
The same table shows where it stops. Sonnet 5.5 scores 83.9 percent on OSWorld 2.1 and 70.6 percent on Terminal-Bench 4.0, so the cheap model still trails the mid-tier one on agentic work by a wide margin [1]. All of these results are vendor-reported. Decrypt, which tested the model briefly, found it responded almost instantly but gave a wrong answer to a simple logic question [2]. One question proves little, but it is a useful reminder that benchmark averages hide individual failures.
How Much Cheaper, Really
Reports differ on the saving. Let's Data Science says Anthropic puts the lower price tier at about 90 percent of requests sent to the previous Haiku [3]. AI Weekly reports that Anthropic claims an average cost about 75 percent below Haiku 4.5 [4]. Decrypt attributes the same 75 percent figure to Anthropic but says the calculation rests on its own assumptions about prompt length and tokenization [2]. The anthropic.com page we retrieved lists prices but did not show a 75 percent figure. Treat 75 percent as an estimate and 90 percent as the list-price cut for short prompts.
"cheapest, fastest and most capable small model it has made"Anthropic's description of Haiku 5.5, as reported by Decrypt, 7 October 2026 [2]
What Customers Report
Anthropic's post quotes several companies, and their results are self-reported. HubSpot says Haiku 5.5 scored 92.8 percent averaged over three runs on its CRM suite [1]. Asana reports over 30 percent lower latency and up to 2.5 times faster inference per agent turn [1]. AlphaSense measured 0.84 against 0.76 for Haiku 4.5 on 400 queries [1].
Rogo said the model was accurate enough to pull segment revenue from 10-K filings when used as a subagent [1]. That subagent role is the pattern to notice. A small model that handles bulk steps while a larger one plans is where the price cut matters most.
Safeguards and Limits
Anthropic says Haiku 5.5 shows far fewer misaligned behaviors than Haiku 4.5 [1]. Its cybersecurity safeguards are stricter than the old Haiku's but looser than Sonnet 5.5's, allowing more defensive work while still blocking penetration testing [1]. Organizations that need broader biology or cyber work can apply to the Life Sciences Verification Program or the Cyber Verification Program [1].
What It Means for Indian Teams
This section is Depth Grid's own analysis. The sources do not discuss India. Prices are quoted in dollars, so Indian startups and IT services firms pay in a currency that moves against the rupee, and the saving will vary with the exchange rate. The practical gain is for high-volume work such as document extraction, support triage and CRM clean-up, where token bills grow fastest. A team running such a pipeline on Haiku 4.5 should re-test on Haiku 5.5 with the effort setting at Low and Med before assuming the headline saving applies.
What to Watch Next
Independent evaluations are the first thing to look for, since every score above comes from Anthropic. Next is whether OpenAI answers on Luna's price or capability. Finally, watch real bills: the 100,000-token threshold means long-context jobs pay five times the short-prompt rate, which can erase much of the saving for document-heavy work [1].
Sources
- Anthropic. "Introducing Claude Haiku 5.5." 7 October 2026. anthropic.com
- Jose Antonio Lanz, Decrypt. "Anthropic launches Haiku 5.5, cheapest, fastest Claude model." 7 October 2026. decrypt.co
- Let's Data Science. "Anthropic Releases Claude Haiku 5.5 With Lower Prices and Effort Controls." October 2026. letsdatascience.com
- AI Weekly. "Anthropic releases Claude Haiku 5.5, cutting prices by 75%." 7 October 2026. aiweekly.co
Read More on Depth Grid
- An AI Agent That Fixes Your Laptop Instead of Just Telling You How
- Meta's New AI Agent Wants to Book Your Travel and Bills
- OpenAI Shelved Its Next Model Over Lying. Then Shipped Agents on the Old One Anyway.
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.