Three weeks ago I wrote that the cheapest tier was the one place OpenAI won outright: GPT-6 Luna was ten times cheaper than Claude Haiku 4.5 on every axis, and Anthropic had nothing at that price.
On 7 October Anthropic released Claude Haiku 5.5, and that line is no longer true. Here is the bottom tier now, from both providers' own pricing pages:
| Claude Haiku 5.5 | GPT-6 Luna | Claude Haiku 4.5 (previous) | |
|---|---|---|---|
| Input | $0.10 | $0.10 | $1.00 |
| Cache read | $0.01 | $0.01 | $0.10 |
| Cache write | $0.125 | $0.125 | $1.25 |
| Output | $0.50 | $0.50 | $5.00 |
| Context window | 1M tokens | not published | 200K tokens |
| Batch discount | 50% | 50% | 50% |
Per million tokens, standard tier.
Identical, to the cent. And if you stop reading at this table, you will make an expensive mistake, because both rows have an asterisk — and the asterisks are not the same size.
The asterisk: Haiku 5.5's 100K step
Claude Haiku 5.5 is priced by prompt length. A request whose prompt is over 100,000 tokens pays a second price list: $0.50 input, $0.05 cache read, $2.50 output. That is 5× every rate.
Three details make this sharper than it looks:
- It reprices the whole request, not just the tokens past the line. A 100,001-token prompt is billed entirely at the higher rate.
- Cached tokens count towards the 100K. Anthropic's documentation is explicit: the prompt length includes cache reads and cache writes. A mostly-cached 120K prompt is still a 120K prompt.
- Every other current Claude model has no step at all. Fable 5.1, Opus 5.5 and Sonnet 5.5 bill one rate across the full 1M window. Haiku 5.5 is the exception, which makes it easy to miss if you already know Anthropic's pricing.
GPT-6 Luna has a step too, but later and gentler: over 272K input tokens, 2× input and cache, 1.5× output — the same rule OpenAI applies to Astra and GPT-6.1 Sol.
So the two models are the same price up to 100K tokens, Haiku 5.5 is 5× the price between 100K and 272K, and still about 2.5× above that. I built a long-context pricing calculator to show exactly where the lines cross for your prompt sizes.
Workload 1: classification — a genuine tie
Ten million classifications a month, 500 input tokens each (400 of them a cached system prompt), 20 output tokens:
| Monthly | |
|---|---|
| Claude Haiku 5.5 | $240 |
| GPT-6 Luna | $240 |
| Claude Haiku 4.5 | $2,400 |
This is where most cheap-tier traffic lives — routing, extraction, moderation, tagging — and for it the decision is no longer about price. Pick on accuracy on your own eval set, latency, and which SDK you already run.
Workload 2: RAG — fine at 80K, five times the price at 120K
A retrieval pipeline sending 80K tokens of documents per question with a 1,000-token answer costs $0.0085 per request on both models. At 100,000 requests a month that is $850 either way.
Let the retriever bring back a little more and land at 120K:
| Per request, 120K prompt | No cache | 90% cached |
|---|---|---|
| Claude Haiku 5.5 | $0.0625 | $0.0139 |
| GPT-6 Luna | $0.0125 | $0.0028 |
5× either way. Caching doesn't rescue you, because the cached tokens are repriced too.
The more dangerous version is the one where prompt sizes vary. Suppose 90% of your requests are 80K and 10% are 105K. On Luna, the average barely moves ($0.00875). On Haiku 5.5, that 10% raises the average request to $0.0132 — 55% more than if every request were 80K. Your dashboard shows a normal week of traffic and a bill half again bigger, and nothing in the code changed.
If you run RAG on Haiku 5.5, cap the retrieved context below 100K in code — not in a prompt instruction. A hard token budget on retrieval is cheaper than any amount of monitoring after the fact.
Workload 3: agent loops — where the step bites hardest
Agents resend their whole history every turn, so their prompts grow until they cross whatever line the pricing page draws. Using the presets in my agent loop cost calculator, with prompt caching on:
| Monthly, caching on | Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Voice call (25 short turns, 10K calls) | $55 | $55 |
| Support chat (8 turns, 20K chats) | $62 | $62 |
| Coding agent (40 turns, 1K sessions) | $555 | $137 |
| Research agent (30 turns, 500 sessions) | $332 | $90 |
Short conversations never reach 100K and the two models tie. The coding agent crosses 100K at turn 16 and the research agent at turn 9; from then on every turn is billed at 5×, and Haiku 5.5 ends up four times the price of Luna.
One more comparison worth sitting with: on that research agent, a full 30-turn session on Haiku 5.5 costs about what the old Haiku 4.5 spent on its first 18 turns before running out of its 200K window. Haiku 5.5's 1M window lets the session finish, but the 90% price cut is spent almost entirely on the long-context step. The new model is dramatically cheaper for short work and not cheaper for long work.
The tokenizer footnote
Per-token prices only compare cleanly when the tokens are the same size, and they aren't.
Anthropic's documentation says Claude models from 4.7 onwards — Haiku 5.5 included — use a newer tokenizer that produces roughly 30% more tokens for the same text than the older one Haiku 4.5 used. So the saving against Haiku 4.5 is real but smaller than the 90% in the price table — closer to 87% on the same text. Against Luna, nobody can tell you the ratio in advance: OpenAI and Anthropic tokenize differently, and the difference depends on your content (code, Urdu and JSON all behave differently from English prose).
The honest comparison is to run a sample of your real traffic through both APIs and read the usage fields. Equal prices per token are not equal prices per request.
What changed in the mid tier
The same week moved the mid tier back to a tie. Claude Sonnet 5.5 (released 28 September) kept Sonnet's $2 / $10 price, and its cache-read rate is $0.10 — half of Sonnet 5's, and the same as GPT-6.1 Sol's. Up to 272K tokens, Sonnet 5.5 and GPT-6.1 Sol now cost exactly the same on every line.
Above 272K, Sol takes OpenAI's long-context surcharge and Sonnet 5.5 doesn't, so long-context work tilts to Sonnet: on the research-agent preset, $1,281 a month against $1,530. That is the mirror image of the cheap tier, where the step belongs to Anthropic.
What I'd pick
High-volume, short-prompt work (classification, routing, extraction): it's a tie on price. Choose on your evals and latency, and keep both behind the same interface so you can switch when the next price cut lands.
RAG with large or variable context: GPT-6 Luna, unless you can hard-cap prompts under 100K tokens. If you can, Haiku 5.5's 1M window and identical price are equally good.
Long-running agents on a cheap model: Luna, by about 4×, until Anthropic moves Haiku's step. Or summarise and trim history so the prompt never crosses 100K — which is good practice anyway.
Anything where prompt length is unpredictable: price the distribution, not the average. A pricing step turns a smooth cost curve into a staircase, and averages hide staircases.
References: Anthropic pricing documentation (Haiku 5.5 prices, the 100K rule, the tokenizer note) · Anthropic models overview · OpenAI API pricing (GPT-6 Luna and the 272K long-context rates) · Claude Sonnet 5.5 launch coverage, VentureBeat. Prices checked against both providers' documentation on 17 October 2026. Worked examples are computed with the same formulas as the long-context pricing and agent loop calculators.