Anthropic launches Claude Haiku 5.5 at $0.10 per million tokens, with effort controls and a 100k cliff

For GCC organisations running AI at volume, this shifts the economics of repetitive enterprise workloads: Arabic and English support triage, KYC and invoice extraction, document compaction, retrieval pre-filters, and the fan-out layer of multi-agent architectures. At $0.10 and $0.50 per million tokens, a Haiku-class call is cheap enough to run multiple times per document while still costing less than a single Sonnet pass. The Decoder notes the launch pricing matches OpenAI's GPT-6 Luna and frames it as a small-model price war. For enterprise buyers, the practical consequence is that cost per completed task, rather than nominal cost per model, is now the primary metric.

The capability jump makes this price cut practical rather than cosmetic. Haiku 5.5 scores 72.4% on the OSWorld 2.1 offline subset compared to 15.7% for Haiku 4.5, 1,620 on GDPval-AA v2.1 against 735, and 39.2% on Terminal-Bench 4.0 against 0%. Early deployments cited in Anthropic's announcement illustrate the intended deployment pattern: AlphaSense runs its Ask in Document feature at roughly 8 million calls weekly and measured 0.84 against 0.76 for Haiku 4.5 across 400 queries; HubSpot reports 92.8% on its CRM evaluation suite; Cognition placed Haiku 5.5 in the sidekick slot of Devin Fusion alongside Opus 5.5 as lead, reporting a FrontierCode score of 66.2. The architectural boundary remains clear: Sonnet 5.5 still scores 70.6% on Terminal-Bench 4.0, and Anthropic positions Haiku specifically for compaction, summarisation and subagent duties rather than complex autonomous coding.

Three operational traps determine whether teams realise the headline savings. First, token density: the Anthropic model documentation states the new tokeniser produces roughly 30% more tokens for the same English text compared to Haiku 4.5, meaning 1M tokens equates to about 555k words versus roughly 750k previously. Existing budget models carry an immediate 30% error before applying unit prices. Second, the 100k threshold: the $0.10 and $0.50 tier covers around 90% of historic Haiku requests, but any request exceeding 100k input tokens is billed at five times the baseline across input, output and cache under Anthropic's pricing documentation. Prompts hovering near 110k represent an expensive configuration defect. Third, data residency pricing: regional endpoints on Bedrock and Google Cloud Vertex carry a 10% premium over global pools, and pinning inference_geo to us on the Claude API applies a 1.1x multiplier across token classes, which affects UAE regulated deployments.

Our recommended approach treats this as an intelligent routing problem rather than a direct model swap. Place a gateway such as FastLLM Proxy in front of providers, configuring per-route model pinning, hard budgets and token limits. Triage, parsing and subagent extraction can land on claude-haiku-5-5, while primary agent loops remain on Sonnet 5.5 or Opus 5.5. Enable prompt caching on every system prompt and static context prefix, taking advantage of $0.01 per million cache reads. Set effort explicitly per route (the API default is medium) following the official effort parameter guide and run an evaluation sweep on internal traffic, particularly Arabic datasets, before migration. A representative subagent configuration using the Anthropic Python SDK looks like this:

import anthropic

client = anthropic.Anthropic()

Subagent extraction route: pinned model, low effort, cached system prompt

resp = client.messages.create( model="claude-haiku-5-5", # Bedrock: anthropic.claude-haiku-5-5 max_tokens=2048, output_config={"effort": "low"}, # default is medium; sweep low and medium on your evals system=[{ "type": "text", "text": EXTRACTION_SYSTEM_PROMPT, "cache_control": {"type": "ephemeral"}, # cache reads at $0.01/MTok }], messages=[{"role": "user", "content": doc_text[:90_000]}], # keep strictly under the 100k tier )