Anthropic releases Claude Haiku 5.5 with 90% API price cut

By Splitfeed EditorialLast updated Oct 8, 2026, 09:50 UTC

On October 7, 2026, Anthropic released Claude Haiku 5.5, the third model in its Claude 5.5 family, through its API and on AWS. The company announced token prices cut by up to 90 percent compared to its predecessor, which Anthropic described as averaging around 75 percent less to run than Claude Haiku 4.5. The model's OSWorld computer use benchmark score reached 72.4 percent, a substantial jump from 15.7 percent for the prior version. The launch timing appears connected to an October 15 pricing deadline, with one piece of coverage framing that date as forcing a floor on API pricing amid competitive pressure from OpenAI's GPT-6 Luna, which the new Haiku pricing now matches.

The release matters beyond a routine price cut because it lands as Anthropic prepares for a planned IPO, making the competitive and financial signals here unusually loaded. The headline reduction is complicated by a new tokenizer that ships alongside the model and consumes more tokens per task than its predecessor, partially eroding the stated savings. How much of that 90 percent figure survives real workloads depends on the nature of those tasks, and that question sits at the center of genuine dispute. The benchmark leap on OSWorld is striking on its face, but the gap between benchmark performance and production reliability in agentic workflows is a standing point of contention in the field.

The pro-AI camp is enthusiastic, focusing on the model's fit for high-volume, cost-sensitive work. Accounts including gymbrowan.bsky.social and Anthropic's own bot framed Haiku 5.5 as a natural subagent for larger models like Opus 5.5 and Sonnet 5.5, suited to tasks like summaries, classification, and database queries. Eris at eriskii.net called it superior to Luna and noted that under 100,000 tokens of context it is effectively free. The anti-AI camp has not published reactions to this specific story, though that camp would typically scrutinize the labor and infrastructure costs embedded in aggressively subsidized pricing, and question what price wars signal about the sustainability of AI business models. The middle ground, represented here by Zubiqo, took a more structural view, arguing that the real strategic move is not cheaper inference in isolation but using subsidized context to lock in multi-step agent workflows before competitors can.

The figure to watch is how the new tokenizer affects effective costs across common workloads, which independent benchmarks and developer reports in the weeks following the October 7 launch should begin to surface. The October 15 deadline cited in coverage may also clarify whether further pricing moves from Anthropic or its rivals follow directly from this release.