Anthropic’s smallest model just got roughly ten times cheaper per token, with a catch buried in the price table. Claude Haiku 5.5, announced on October 7, 2026, costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Its predecessor, Haiku 4.5, was $1 and $5. Anthropic calls the new model “the cheapest, fastest, and most capable small model we’ve ever released.” The numbers support the first claim more clearly than the third.
Why 75% and 90% both appear
Anthropic’s launch materials say Haiku 5.5 costs about 75% less to run than Haiku 4.5 on average, while the per-token list price is 90% lower under the 100,000-token line and 50% lower above it. These figures are not in conflict. A developer write-up on BuildMVPFast explains that Haiku 5.5 uses a newer tokenizer that produces roughly 30% more tokens for the same text, which pulls the blended saving down from 90% toward 75%. Images can cost more too: that write-up reports a 2,000 by 1,500 pixel image uses about 2.5 times as many visual tokens as on Haiku 4.5. The takeaway is that the headline price cut is real, but a bill will not fall by exactly the ratio on the pricing page.
Cache reads cost $0.01 per million tokens at the lower tier. Anthropic also cut Sonnet 5.5’s cache reads to $0.10 per million in the same announcement, which helps teams that resend the same long context repeatedly.
The 100,000-token cliff
Haiku 5.5 introduces a pricing step that other current Claude models do not have. Once a request’s prompt passes 100,000 tokens, the whole request is billed at $0.50 input and $2.50 output per million. According to the BuildMVPFast analysis of Anthropic’s docs, prompt length counts all input tokens, including cache reads and writes, so a request with 95,000 cached tokens plus 6,000 new ones crosses the line. Agent loops, where context grows each turn, are the most exposed.
The same write-up gives a worked example: a single turn with a 120,000-token prompt and 2,000 output tokens costs about $0.065 on Haiku 5.5, $0.13 on Haiku 4.5, and $0.26 on Sonnet 5.5. Trimmed to a 90,000-token prompt, the Haiku 5.5 cost drops to about a cent. Even over the cliff the model is roughly half the price of its predecessor, but the drop from compaction or tighter retrieval is far larger than the drop from switching models.
What it can and cannot do
Anthropic’s benchmark table shows large jumps over Haiku 4.5. On Terminal-Bench 4.0 it scores 39.2% against 0.0%, on the offline OSWorld 2.1 subset 72.4% against 15.7%, and on Humanity’s Last Exam without tools 45.9% against 10.2%. These are the company’s own results. Anthropic also says Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding, with Sonnet 5.5 at 70.6% on Terminal-Bench 4.0, and points Haiku at high-volume, cost-sensitive work: summaries, compaction, database queries, classification, live customer support and subagent tasks.
Independent tests are mixed. In one developer’s code-search trial reported by BuildMVPFast, Haiku found 44 of 48 changed files versus 45 for Opus, at $0.44 versus $6.17. In a fact-checking trial from the same write-up, Haiku wrongly flagged 30 correct sentences as errors while Opus flagged none. Those are single-developer anecdotes, not benchmarks, but they illustrate the usual pattern: small models are excellent at narrow retrieval and extraction and weaker where false positives are costly. The write-up also notes that a Reddit claim of $24.96 for Haiku 5.5 versus $1.96 for OpenAI’s GPT-6 Luna was doubted by commenters because the test shared no prompts, so cross-vendor cost comparisons should be run on your own workload. The write-up lists Luna at the same $0.10 and $0.50 rates.
Migration is not a drop-in
Teams on Haiku 4.5 should expect some code changes. According to the BuildMVPFast summary, budget_tokens thinking settings, assistant prefill, and the temperature, top_p and top_k parameters either return errors or must be removed. Thinking tokens count toward the maximum output setting, an older computer-use tool version fails, and priority-tier arrangements on Haiku 4.5 do not carry over. Haiku 5.5 is also the first Haiku-class model with an adjustable effort setting, and it is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.
What This Means for Philippine Founders
At $0.10 and $0.50 per million tokens, tasks that were too expensive to automate a month ago may now pencil out: tagging and summarizing customer chats, extracting fields from invoices and receipts, drafting first-pass replies for a support team. For a startup billing in pesos, per-conversation cost is what matters, so the sensible first step is to log token counts on a real week of traffic and calculate cost per resolved ticket under both the under-100,000 and over-100,000 token tiers. Keep prompts short, cache the stable parts, and watch for the tokenizer difference when estimating.
The model’s weaker spots deserve equal attention. How well Haiku 5.5 handles Filipino, Taglish and regional-language text was not covered in the sources reviewed here, so test it on actual customer messages before replacing a human review step. A reasonable setup is Haiku for the high-volume first pass, a larger model for ambiguous cases, and a person for anything with legal, financial or reputational weight. That keeps the savings without betting the product on a small model’s judgment.
Share this article