Anthropic releases Claude Haiku 5.5 with 90% price cut

By Splitfeed EditorialLast updated Oct 7, 2026, 19:30 UTC

On October 7, 2026, Anthropic released Claude Haiku 5.5, the third model in its Claude 5.5 family, available immediately through its API and on AWS. Token prices dropped by up to 90 percent compared to the previous Haiku model, landing at a level VentureBeat reported as matching GPT-6 Luna pricing. The model's performance on the OSWorld computer use benchmark jumped from 15.7 percent, the score of its predecessor, to 72.4 percent. One complication noted by The Decoder is that Claude Haiku 5.5 uses a new tokenizer that consumes more tokens per task, which partially offsets the advertised savings when comparing real-world costs against competing models.

The release lands at a commercially sensitive moment for Anthropic, which Reuters and others reported is expanding its model lineup ahead of a planned IPO. That context gives the pricing decision a strategic dimension beyond simple competition: a dramatically cheaper small model widens Anthropic's developer base, locks in usage patterns, and generates the kind of volume metrics that matter to investors. The tokenizer detail complicates the headline number, and whether buyers actually save 90 percent in practice depends heavily on the types of tasks they run, making direct cost comparisons with GPT-6 Luna genuinely uncertain rather than settled by the announcement alone.

Pro-AI voices treated the release as a clear win, focusing on Haiku 5.5's role in multi-model pipelines. Rowan at gymbrowan.bsky.social called it a near-perfect subagent for coding, classification, and repetitive tasks alongside larger models like Opus and Sonnet 5.5. Eris at eriskii.net described it as effectively free under 100,000 tokens of context. Anthropic's own account framed it as the cheapest, fastest, and most capable small model it has released, and highlighted pairings with Opus 5.5 or Sonnet 5.5 for high-volume workloads like summaries and database queries. The anti-AI camp produced no public reactions to this story; that camp would typically raise concerns about the energy and infrastructure costs hidden behind aggressive price cuts, and whether subsidized pricing is sustainable or reflects actual efficiency gains. The middle ground, represented by Zubiqo at zubiqo.bsky.social, acknowledged that a cheaper model is routine but argued the more significant move is using subsidized context pricing to entrench Haiku 5.5 inside multi-step agent workflows before competitors can respond.

The argument will sharpen once developers run head-to-head cost benchmarks that account for the new tokenizer, and if Anthropic files IPO documentation that reveals the unit economics behind the pricing strategy. Those numbers would clarify whether the 90 percent figure holds at scale or whether the tokenizer offset is large enough to make the gap with GPT-6 Luna narrower than the headline suggests.