Reading
Haiku 5.5 vs GPT-6 Luna: Anthropic's 90% Cut Has a Catch
Haiku 5.5 costs the same as GPT-6 Luna and beats it on Anthropic's own tests. But past 100K tokens the bargain disappears. Here's who really saves.
Prices last read 2026-10-13
Short answer
Not undercut. Matched. For short prompts Haiku 5.5 and GPT-6 Luna have the same list price. For long prompts, Luna is cheaper.Anthropic launched Claude Haiku 5.5 on October 7, 2026. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 cost $1 and $5. OpenAI's standard rate for GPT-6 Luna is the same $0.10 and $0.50 for inputs up to 272K tokens. Once a prompt goes past 100K tokens, Haiku 5.5 costs $0.50 and $2.50, five times Luna's short-context rate. Anthropic's own benchmark table puts Haiku 5.5 ahead of Luna on all six tests where both models have scores.Source: Introducing Claude Haiku 5.5 (Anthropic) · 2026-10-13; Pricing (Claude Platform docs) · 2026-10-13; API pricing (OpenAI) · 2026-10-13
What actually changed on October 7
Anthropic calls Haiku 5.5 its cheapest and fastest small model. It is pitched at high-volume work such as summaries, classification and database queries, and at running as a subagent under Opus 5.5 or Sonnet 5.5. It is also the first Haiku with an adjustable effort setting, from Low up to Max. The model ID is claude-haiku-5-5, and Anthropic says it is available on all its platforms, including AWS, Google Cloud and Microsoft Azure. The same announcement halved Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens. Anthropic says that makes most agentic work on Sonnet 5.5 about 20% cheaper.
Source: Introducing Claude Haiku 5.5 (Anthropic) · 2026-10-13; Anthropic releases Claude Haiku 5.5 small model and halves Sonnet 5.5 cache read prices (SiliconANGLE) · 2026-10-13
Why '90% cheaper' really means about 75%
Anthropic's own footnote explains the gap. The cut is 90% for requests up to 100K tokens and 50% above that, and Anthropic says about 90% of Haiku 4.5 requests fell under the threshold. Haiku 5.5 also uses an updated tokenizer that needs slightly more tokens per task. Taking both into account, Anthropic says Haiku 5.5 costs about 75% less than Haiku 4.5 on average. Anthropic's pricing docs say the newer tokenizer family produces roughly 30% more tokens for the same text, and that the exact increase depends on the content. One correction: SiliconANGLE's launch report described long prompts as getting a discount. In Anthropic's own pricing table, long prompts cost more.
Source: Introducing Claude Haiku 5.5 (Anthropic) · 2026-10-13; Pricing (Claude Platform docs) · 2026-10-13; Anthropic releases Claude Haiku 5.5 small model and halves Sonnet 5.5 cache read prices (SiliconANGLE) · 2026-10-13
a request over the threshold pays the higher prices even when part of its prompt is a cache hit
Claude Platform docs, Long context pricing (read 2026-10-13)Source: Pricing (Claude Platform docs) · 2026-10-13
This is the line agent builders should read twice. The docs say prompt length counts all input tokens, cache reads and writes included, and that each request is priced on its own. A long cached conversation that crosses 100K tokens pays the higher rate on the whole request, not just on the part above the threshold.
Source: Pricing (Claude Platform docs) · 2026-10-13
The Luna fight: timing, prices and benchmarks
Notebookcheck, citing OpenAI's October 7 release notes, reports that ChatGPT Free and Go users started getting GPT-6 Luna in the regular Chat tab on October 8. That was one day after Haiku 5.5 launched, and after Luna had been available only in the desktop app. On the API, OpenAI lists Luna at $0.10 input, $0.01 cached input and $0.50 output for inputs up to 272K tokens. Above 272K it costs $0.20 and $0.75. Anthropic's table lists Haiku 5.5 ahead of Luna on OSWorld 2.1 computer use (72.4% vs 48.9%), Terminal-Bench 4.0 (39.2% vs 16.4%), FrontierCode 1.1 (46.4% vs 42.4%), Chartography (46.4% vs 29.1%) and both Artificial Analysis knowledge-work tests. The page sends readers to the Haiku 5.5 system card for evaluation methods and has no footnote explaining how the Luna scores were obtained.
Source: ChatGPT Free and Go users start receiving GPT-6 Luna in Chat (Notebookcheck) · 2026-10-13; API pricing (OpenAI) · 2026-10-13; Introducing Claude Haiku 5.5 (Anthropic) · 2026-10-13
Our take: a calculated shot, with a 90% headline that oversells it
Launching at exactly Luna's price, the day before Luna reached ChatGPT's free users, is not a coincidence. It's a fight picked on purpose, and I like it. Small models have been the industry's boring afterthought, and Haiku 4.5 at $1/$5 was frankly overpriced next to OpenAI's budget tier. But the '90% cheaper' headline is marketing math. Anthropic's own average is 75%, and the 100K cliff punishes the long-context agent loops this model is supposed to serve. If your prompts are long, Luna is the cheaper model, full stop. As for the benchmarks, a vendor table where the vendor wins every row is a press release, not a verdict. Haiku 5.5 is probably underrated as a subagent and overrated as a Luna killer until independent labs run the numbers.
Who's affected and what to do
- Haiku 4.5 API users: check how long your prompts arePull token counts from your logs. If nearly all requests stay under 100K input tokens, cache included, you are in the 90%-off zone. Anything over 100K is billed at the higher Haiku 5.5 rate.
- Re-measure tokens, not just pricesThe newer tokenizer produces more tokens for the same text. Run a sample of real traffic through Haiku 5.5 and compare billed tokens before you trust any savings estimate.
- Test the effort settingHaiku 5.5 lets you trade cost against quality from Low to Max. Anthropic's charts plot accuracy against cost, so try a few levels on your own tasks rather than defaulting to the top one.
- Sonnet 5.5 users with heavy cachingCache reads are now $0.10 per million tokens, with no code change required. Agent setups that reuse long cached prefixes benefit most.
- Weighing Luna instead?Benchmark both models on your own evals. Anthropic's comparison is self-reported, and prices only match up to 100K tokens.
Source: Pricing (Claude Platform docs) · 2026-10-13; Introducing Claude Haiku 5.5 (Anthropic) · 2026-10-13; API pricing (OpenAI) · 2026-10-13
Claude app subscribers don't pay these token rates directly. In the same announcement, Anthropic said Max 5x subscribers get $100 in monthly API credits, Max 20x subscribers get $200, and Team plans get up to $500 pooled, rolling out that week. Those credits now go much further on Haiku 5.5.
Source: Introducing Claude Haiku 5.5 (Anthropic) · 2026-10-13
Optional: compare Claude subscription listings from sellers (a seller listing, not official access or an endorsement). That link carries our partner code Sk52g, which is how a purchase is attributed to this site. Promotion code UPD3S.
Sources (all read 2026-10-13)