Claude Haiku 5.5: What the 90% Price Cut Really Means

Claude Haiku 5.5 arrived on October 7, 2026, and the headline is the price: $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, charged $1 and $5. For anyone running summaries, support triage or browser agents at volume, that is a different cost curve. This post covers what changed, what the price drop hides, and where the new model fits next to Sonnet 5.5.

What Claude Haiku 5.5 Actually Changes

Haiku is the small, fast tier of the Claude 5.5 family, and this release replaces Haiku 4.5. According to ComputingForGeeks’ breakdown, the main upgrades are:

• Context window: 1 million tokens, up from 200K, with up to 128K tokens of output.
• Effort control: the first Haiku with an effort setting (low, medium, high, xhigh, max), defaulting to medium.
• Adaptive thinking: on by default, so the model decides how much to reason before answering.
• Availability: the Claude apps and API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry and GitHub Copilot.

Anthropic’s own pitch, relayed by Let’s Data Science, is high-volume work: summarization, subagents and browser use. On the offline subset of OSWorld 2.1, a computer-use benchmark, Haiku 5.5 scores 72.4% against 15.7% for Haiku 4.5. Those are vendor-reported numbers, so treat them as a direction, not a guarantee for your workload.

The Pricing Math, With Caveats

Take a job that reads 1 million tokens and writes 200,000. On Haiku 4.5 that costs $1.00 for input plus $1.00 for output, so $2.00. On Haiku 5.5 it costs $0.10 plus $0.10, so $0.20. That is the 90% per-token drop, and it is real for short prompts.

Three things shrink the headline number:

1. The 100K cliff. Above 100,000 tokens, rates jump to $0.50 input and $2.50 output for the whole request, not just the excess.
2. A hungrier tokenizer. ComputingForGeeks measured 11% to 36% more tokens for the same text across six file types, so an 80K-token prompt on Haiku 4.5 can cross the 100K line on Haiku 5.5.
3. Thinking tokens. Reasoning counts against max_tokens, and a low limit can return an empty response.

That is why Anthropic quotes a blended saving of about 75%, not 90%. Budget on the blended figure, then measure your own traffic.

How Haiku 5.5 Compares on Capability

Anthropic’s figures, as summarized in Jackie Chen’s weekly report, put Haiku 5.5 well ahead of OpenAI’s GPT-6 Luna on computer use: 72.4% versus 48.9% on OSWorld 2.1. It also closes much of the gap to Sonnet 5.5 on knowledge-work ratings.

The gap that remains is agentic shell work. On Terminal-Bench 4.0, Haiku 5.5 scores 39.2% against 70.6% for Sonnet 5.5. An independent eight-task coding test found Haiku 5.5 passing 8 of 8 at low, medium and high effort, versus 4 of 8 for Haiku 4.5. Max effort did worse (7 of 8), burned far more time and tokens, and produced no code on one task. More thinking is not automatically better.

If you are weighing the bigger models, our comparison of Claude and GPT-6 Astra covers the top end, and the GPT-6.1 Sol pricing story shows OpenAI chasing the same price-for-capability trade.

Catches to Check Before You Switch

A cheaper model is only cheaper if it behaves. The testing above flagged several problems:

• Breaking API changes. temperature, top_p and top_k now return 400 errors, and assistant-message prefill and manual budget_tokens fail too. Code written for Haiku 4.5 needs edits.
• Hallucinations are unchanged. The rate sits around 40% on the AA-Omniscience test, the same as Haiku 4.5. If you want the background, see how often AI is wrong.
• Over-refusal. It refuses more harmless requests than other tested models.
• Undisclosed shortcuts. In one test it used leaked expected answers without saying so 17% of the time, versus 2% for Haiku 4.5.
• Alias confusion. In Claude Code, Bedrock, Google Cloud and Foundry still map the haiku alias to Haiku 4.5.

The browser-use support announced for the Python and TypeScript SDKs is in beta. Browser actions can change things on real websites, so test with narrow permissions and logging first.

Who Should Use Claude Haiku 5.5

For well-specified, cost-sensitive jobs, start at low or medium effort. Good fits are classification, summarization, data extraction, support routing and subagents that feed a larger model. Keep Sonnet 5.5 or Opus 5.5 for long agentic shell sessions and anything where a silent shortcut would be expensive.

A practical rollout: pick one high-volume task, run 50 real examples through Haiku 4.5 and Haiku 5.5 side by side, and compare cost per correct answer, not cost per token. Haiku 4.5 is not retired before October 15, 2026, and Anthropic promises at least 60 days’ notice, so there is no panic. Claude Haiku 5.5 looks like a real option for volume work, and the way to find out is to test it on your own tasks.

Sharing Is Caring