GPT-5.6 Luna is cheaper than Claude Haiku 4.5 and scores twice as high. Here is what Haiku still has.
Luna costs $0.20 / $1.20 and scores 38. Haiku 4.5 costs $1 / $5 and scores 18. On the numbers it is not close, and Haiku's retirement window opens in October.
GTM Architect & Growth Operator · Now · 16 September 2026 · 6 min read
On this page
GPT-6 Luna has replaced GPT-5.6 Luna
OpenAI shipped GPT-6 Luna on 22 September 2026 at $0.10 / $0.50 per million tokens against GPT-5.6 Luna’s $0.20 / $1.20. The names are close enough to mix up: everything below is the 5.6 model. For the current one, read GPT-6 Luna against Claude Haiku 4.5.
Anthropic describes Haiku 4.5 as the fastest model with near-frontier intelligence. Inside Anthropic’s lineup, the first half is true.
Put it next to GPT-5.6 Luna and the numbers support neither half.
| Info | Claude Haiku 4.5 | GPT-5.6 Luna |
|---|---|---|
| Input / output per 1M | $1 / $5 | $0.20 / $1.20 |
| Cache read per 1M | $0.10 | $0.02 |
| Intelligence Index v4.3 | 18 (reasoning) | 38 |
| Output speed, tokens/s | 84.7 | 115.0 |
| Blended price per 1M | $0.77 | $0.17 |
| Context window | 200K | 1,050,000 |
| Max output | 64K | 128,000 |
| Knowledge cutoff | Feb 2025 (reliable) | 16 Feb 2026 |
| Retirement | Not sooner than 15 Oct 2026 | Not stated |
Cheaper, higher scoring, faster, longer context, newer knowledge. I went looking for the row where Haiku 4.5 comes out ahead, in any of the numbers either vendor publishes or Artificial Analysis measures, and did not find one.
Twenty points at a fifth of the price
On Intelligence Index v4.3, Luna scores 38. Haiku 4.5 scores 18 with reasoning on and an estimated 15 without.
Twenty points is not a small-model rounding error. It is wider than the gap between Sonnet 5 and Opus 5 in Anthropic’s own lineup, which I wrote up in Sonnet 5 against Opus 5. And Luna is the cheaper model: five times cheaper on input, about four times on output, $0.17 against $0.77 blended.
Speed usually decides the small tier. Luna wins that too, at 115 output tokens per second against Haiku 4.5’s 84.7 on Artificial Analysis’s measurement.
The context gap is structural
Haiku 4.5 has a 200K context window and a 64K output limit. Luna has 1,050,000 and 128,000.
You cannot engineer around that. A 400K-token document fits in Luna’s window and does not fit in Haiku’s at all. Luna does charge more for it: above 272K input tokens it bills the whole request at 2x input and 1.5x output, so $0.40 / $1.80. That is still cheaper than Haiku’s standard rate, on a prompt Haiku cannot take.
Key takeaway
Most small-model comparisons argue about price and speed. This one is decided before either comes up, by how much the model can read.
The retirement window
Anthropic’s models page lists Haiku 4.5’s retirement as not sooner than 15 October 2026 on Anthropic-operated platforms.
“Not sooner than” is a floor, not a scheduled shutdown. Nothing says Haiku 4.5 goes away on that date. But it is the earliest date Anthropic has committed to keeping it, a month from now, and Amazon Bedrock and Google Cloud set their own. If a production system depends on Haiku 4.5, the migration plan belongs on this quarter’s list, whichever model it points to.
What Haiku 4.5 still has
Not the numbers. What it has is Anthropic.
- An existing Claude integration. If your prompts, tools and evaluation harness are built on the Claude API, switching vendor for the small tier adds a second API, a second set of behaviours and a second bill.
- Commercial and data terms. If your agreements, data residency or procurement approvals cover Anthropic and not OpenAI, that decides it regardless of price.
- One vendor for routing. A router that escalates from a small Claude model to Opus 5 or Fable 5.1 keeps everything in one API.
Those are real reasons. They are reasons to stay, not reasons to choose Haiku 4.5 fresh.
Which one I would use
| Situation | Model | Why |
|---|---|---|
| New high-volume work, free to choose | GPT-5.6 Luna | Cheaper, twenty index points higher, faster, five times the context. |
| Documents longer than 200K tokens | GPT-5.6 Luna | Haiku 4.5 cannot take them. |
| Existing Claude integration, simple tasks | Haiku 4.5, for now | Staying in one API has a cost of its own. Plan for the retirement window. |
| Agreements cover Anthropic only | Claude | Consider Sonnet 5 for anything Haiku 4.5 struggles with; it scores 38 at $2 / $10. |
For the Claude-side alternative, I compared Haiku 4.5 against Sonnet 5. Luna against Anthropic’s mid tier is in Luna against Sonnet 5.
What would change my mind
A new small Claude model. Haiku 4.5’s reliable knowledge cutoff is February 2025, and a successor is the most likely thing to change this comparison.
A Luna price change. At $0.20 input, a move of a few cents changes the ratio a lot.
A cost-per-task comparison on simple, high-volume work. On very short tasks the index gap may not show up at all, and then the question is only price and speed, which Luna still wins on.
Questions people asked
Is GPT-5.6 Luna better than Claude Haiku 4.5?
On every published measure I can find, yes: 38 against 18 on the Intelligence Index v4.3, faster, a far larger context window, a newer knowledge cutoff, and cheaper.
Which is cheaper, Claude Haiku 4.5 or GPT-5.6 Luna?
Luna: $0.20 / $1.20 against $1 / $5, and $0.17 against $0.77 blended. Above 272K tokens Luna’s price rises, but Haiku 4.5 cannot take a prompt that long.
When is Claude Haiku 4.5 being retired?
Not sooner than 15 October 2026 on Anthropic-operated platforms, which is a floor rather than a scheduled date. Bedrock and Google Cloud set their own.
Should I use Claude Haiku 4.5 or GPT-5.6 Luna?
Luna for new work where you can choose the vendor. Haiku 4.5 only where staying inside Anthropic matters more than the numbers, with a plan for its retirement window.
If you are moving a high-volume pipeline off a small model, the contact page is the fastest route to me.
