Anthropic's cheapest, fastest model line doesn't usually make headlines — Haiku exists to be the unglamorous workhorse behind high-volume tasks, not the model anyone brags about on a benchmark chart. Claude Haiku 5.5, released October 7, 2026, is a genuine exception: it combines a 90% list-price cut with the kind of benchmark jump that would be notable even on a flagship release, not just a budget-tier one.
That combination — much cheaper and dramatically more capable at the same time — is what's driving the comparisons to OpenAI's own budget model, GPT-6 Luna, which launched with near-identical headline pricing back in September. The two companies are now fighting a real price war in the cheapest, highest-volume tier of the market, and Haiku 5.5's benchmark numbers suggest Anthropic thinks it can win that fight on capability, not just price.
Quick summary: Claude Haiku 5.5 launched October 7, 2026, priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens — exactly matching OpenAI's GPT-6 Luna pricing, and a 90% cut from Haiku 4.5's $1/$5. Above 100,000 tokens, pricing rises to $0.50/$2.50. The context window expands from Haiku 4.5's 200,000 tokens to 1,048,576 tokens (~1M), with up to 128,000 tokens of output. On Anthropic's own OSWorld 2.1 offline computer-use benchmark, Haiku 5.5 scores 72.4%, versus 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. It ships with adaptive 'effort' controls for reasoning depth and with preserved thinking, the anti-distillation safeguard Anthropic first introduced with Claude Fable 5.1.
The pricing: a straight 90% cut, with a catch
Haiku 5.5's headline price is $0.10 per million input tokens and $0.50 per million output tokens — but only for prompts up to 100,000 tokens. Beyond that threshold, pricing steps up to $0.50 per million input and $2.50 per million output, a detail that's easy to miss if you only read the launch headline. Compared to Haiku 4.5's flat $1 input / $5 output pricing, that's still a 50% cut even at the higher tier, and a 90% cut for the (likely much more common) short-prompt case. Anthropic itself describes the blended real-world savings as averaging around 75% less than Haiku 4.5, once you account for the mix of short and long prompts most customers actually send (Sources: VentureBeat, datanorth.ai).
| Model | Input (≤100K tok) | Output (≤100K tok) | Context Window | OSWorld 2.1 (offline) |
|---|---|---|---|---|
| Claude Haiku 4.5 | $1.00 / 1M tok | $5.00 / 1M tok | 200,000 tok | 15.7% |
| Claude Haiku 5.5 | $0.10 / 1M tok | $0.50 / 1M tok | 1,048,576 tok (~1M) | 72.4% |
| GPT-6 Luna (OpenAI) | $0.10 / 1M tok | $0.50 / 1M tok | — | 48.9% |
That OSWorld 2.1 number is the one worth sitting with. OSWorld is a benchmark that measures how well a model can actually operate a computer — navigating real desktop and web interfaces to complete a task, not just answer a question about one. Going from 15.7% to 72.4% on the offline subset isn't an incremental tuning improvement; it's roughly a 4.6x jump, and it puts Haiku 5.5 well ahead of GPT-6 Luna's 48.9% on the same measure, at matching sticker price. These are Anthropic's own published numbers rather than a neutral third-party benchmark, so some caution is warranted until outside evaluators reproduce them — but the gap is large enough that it would be a notable story even if the real number came in meaningfully lower (Source: VentureBeat).
What it's actually built for
Haiku was never meant to compete with Opus or Fable on hard reasoning — it's built for the high-volume, latency-sensitive work that sits underneath most production AI systems: classification, routing requests to the right tool or sub-agent, structured data extraction, summarization, and running as a cheap subagent inside a larger pipeline orchestrated by a bigger model. Haiku 5.5 adds an adaptive 'effort' control, letting a caller dial reasoning depth up or down per request rather than always paying for the same fixed amount of thinking — useful when most of your traffic is simple but a fraction of it genuinely needs more reasoning (Source: eesel.ai).
Haiku 5.5 also ships from day one with preserved thinking, the anti-distillation protection Anthropic introduced alongside Claude Fable 5.1 in September 2026. Mechanically, it means that on API accounts created on or after August 31, 2026, the Messages API no longer lets a caller edit or replay the 'thinking' block from a prior turn in a conversation — the specific technique outside parties have used to harvest a model's internal reasoning traces and train cheaper copycat models on them. Anthropic hasn't published Haiku-5.5-specific numbers on how effective this is; its broader research on the measure's impact across the models that carry it is a separate, ongoing story worth tracking on its own.
If you're choosing between Haiku 5.5 and GPT-6 Luna purely on the published price sheet, they're identical at $0.10/$0.50 for short prompts. The real decision point is the 1M-token context window and the OSWorld computer-use gap — if your workload involves long documents or anything resembling autonomous desktop/browser operation, that's where the two diverge most, at least on Anthropic's own figures.
The bigger picture: a genuine race to the bottom on price
Haiku 5.5 lands almost exactly one year into a visible pattern: cheap-tier models that used to cost a few dollars per million tokens now routinely cost pennies, while capability on practical, agentic benchmarks keeps climbing rather than merely holding steady. GPT-6 Luna, Claude Haiku 5.5, and the Decisions-API layer OpenAI built on top of Luna (which we cover in a separate post on October 6's launches) are all evidence of the same trend: the 'cheap model' tier is where the most intense competition in the industry is currently happening, not at the flagship level.
01How much cheaper is Claude Haiku 5.5 than Haiku 4.5?
The sticker price is 90% lower for prompts up to 100,000 tokens — $0.10/$0.50 per million input/output tokens versus Haiku 4.5's $1/$5. Anthropic says the average real-world savings, blending shorter and longer prompts, is closer to 75%.
02Does Claude Haiku 5.5 beat GPT-6 Luna?
On price, they're identical at $0.10 input / $0.50 output per million tokens for short prompts. On Anthropic's own OSWorld 2.1 offline computer-use benchmark, Haiku 5.5 scores 72.4% versus GPT-6 Luna's 48.9% — though that figure comes from Anthropic, not an independent evaluator, so it's worth treating with some caution until reproduced elsewhere.
03What is 'preserved thinking' in Claude Haiku 5.5?
It's an anti-distillation safeguard, first introduced with Claude Fable 5.1, that stops API accounts created on or after August 31, 2026 from editing or replaying a prior turn's internal reasoning ('thinking') block — the main technique used to extract a model's reasoning and train copycat models on it.
04What's the context window on Claude Haiku 5.5?
1,048,576 tokens (roughly 1 million), up from Haiku 4.5's 200,000 tokens, with up to 128,000 tokens of output.
05What is Claude Haiku 5.5 best used for?
High-volume, cost-sensitive tasks: classification, routing, structured data extraction, summarization, and running as a lightweight subagent inside a larger pipeline — not as a replacement for Opus or Fable on hard, open-ended reasoning.
Budget-tier models like Haiku 5.5 and GPT-6 Luna are good enough now that picking the wrong one mostly just costs you a slightly worse result, not a broken product — which makes it worth actually comparing them on your own workload rather than trusting either company's benchmark chart. LumiChats gives you 40+ premium AI models, including Claude Haiku 5.5 and the latest from OpenAI, Google, and xAI, in one place, so you can run the same task through a handful of cheap models side by side and see which one actually earns its low price.
