// head_to_head
Claude Haiku 4.5 vs GPT-5.6 Luna
Published LiveBench scores across all seven categories, live list pricing, context windows, and the measured cost of a point of capability — for both models, side by side.
GPT-5.6 Luna wins outright — it scores higher and costs less.
GPT-5.6 Luna leads by 20.0 points overall while listing 4.4× cheaper per blended million tokens. There is no trade-off to reason about here; the only reason to pick the other is a constraint this table does not show, like an existing contract, a region, or a provider you are already on. The two cost measures disagree here, which is worth knowing: GPT-5.6 Luna has the lower sticker price, but Claude Haiku 4.5 earns each point of capability for less — $0.3893 against $0.4693 — because per-token rates do not predict how many tokens a model actually spends on a task.
anthropic
Claude Haiku 4.5
- Blended / 1M
- $2.00
- Context
- 200K
- Released
- Oct 15, 2025
- Overall score
- 46.0
openai
GPT-5.6 Luna
- Blended / 1M
- $0.450
- Context
- 1.1M
- Released
- Jul 9, 2026
- Overall score
- 66.1
Specs and pricing
| Metric | Claude Haiku 4.5 | GPT-5.6 Luna |
|---|---|---|
| LiveBench overall Mean of the seven LiveBench category scores, 0–100. Higher is better. | 46.0 | 66.1win |
| Cost per point Measured benchmark spend divided by overall score — dollars per point of capability. | $0.3893win | $0.4693 |
| Blended price / 1M 3:1 input:output mix, the usual shape of production traffic. | $2.00 | $0.450win |
| Input price / 1M | $1.00 | $0.200win |
| Output price / 1M | $5.00 | $1.20win |
| Cached input / 1M Price of an input token served from the prompt cache, where the provider publishes one. | $0.100 | $0.020win |
| Context window | 200K | 1.1Mwin |
| Max output tokens | 64K | 128Kwin |
Benchmarks by category
An overall score averages away the thing you probably care about. These are the seven LiveBench categories scored separately — Claude Haiku 4.5 on top, GPT-5.6 Luna below, both out of 100.
What each one costs to run
Per-token prices are hard to feel. These are monthly list costs for both models across five workload shapes, using each provider's published cached-input rate where there is one.
| Workload | Claude Haiku 4.5 | GPT-5.6 Luna |
|---|---|---|
| Support chatbot 1.2K in / 400 out × 200K requests | $575.20/mo | $131.04/mo |
| RAG assistant 8K in / 600 out × 100K requests | $740.00/mo | $160.00/mo |
| Coding agent 40K in / 4K out × 20K requests | $696.00/mo | $155.20/mo |
| Document extraction 20K in / 1.5K out × 50K requests | $1,330/mo | $281.00/mo |
| Bulk classification 500 in / 20 out × 5M requests | $2,550/mo | $530.00/mo |
Which should you pick?
You are running this at volume
Claude Haiku 4.5
Lowest measured cost per point of capability at $0.3893 per point — the gap compounds with every request.
Quality matters more than the bill
GPT-5.6 Luna
Highest overall LiveBench score of the two at 66.1.
The workload is coding or agentic work
GPT-5.6 Luna
Leads on coding — 73.6 against 48.1.
You need to fit large documents in one call
GPT-5.6 Luna
Wider context window — 1.1M against 200K.
Claude Haiku 4.5 vs GPT-5.6 Luna FAQ
Which is better, Claude Haiku 4.5 or GPT-5.6 Luna?
GPT-5.6 Luna wins outright — it scores higher and costs less. GPT-5.6 Luna leads by 20.0 points overall while listing 4.4× cheaper per blended million tokens. There is no trade-off to reason about here; the only reason to pick the other is a constraint this table does not show, like an existing contract, a region, or a provider you are already on. The two cost measures disagree here, which is worth knowing: GPT-5.6 Luna has the lower sticker price, but Claude Haiku 4.5 earns each point of capability for less — $0.3893 against $0.4693 — because per-token rates do not predict how many tokens a model actually spends on a task.
Is Claude Haiku 4.5 cheaper than GPT-5.6 Luna?
GPT-5.6 Luna is cheaper. On a 3:1 input:output blend, Claude Haiku 4.5 lists at $2.00 per million tokens and GPT-5.6 Luna at $0.450 — GPT-5.6 Luna is 4.4× cheaper. Input and output are priced separately — Claude Haiku 4.5 charges $1.00 in and $5.00 out, GPT-5.6 Luna charges $0.200 and $1.20 — so the model that looks cheaper flips depending on how output-heavy your workload is.
Claude Haiku 4.5 vs GPT-5.6 Luna: which scores higher on benchmarks?
Claude Haiku 4.5 scores 46.0 and GPT-5.6 Luna scores 66.1 overall on LiveBench, the mean of its seven categories. That is a 20.0-point lead for GPT-5.6 Luna. Category scores differ from the overall figure — a model can lead on reasoning and trail on coding, which the per-category table above breaks out.
Which gives better value for money, Claude Haiku 4.5 or GPT-5.6 Luna?
Claude Haiku 4.5. Cost per point divides the measured dollars LiveBench spent running the benchmark by the score it earned, so it captures something token pricing misses: a reasoning model can emit many times more tokens than its per-token rate suggests. Claude Haiku 4.5 works out at $0.3893 per point and GPT-5.6 Luna at $0.4693.
Does Claude Haiku 4.5 or GPT-5.6 Luna have a bigger context window?
GPT-5.6 Luna has the larger context window: 200K for Claude Haiku 4.5 against 1.1M for GPT-5.6 Luna. Note that a window you can fill is not a window you should fill — retrieval quality usually degrades well before the limit, and you pay for every token you put in it.
Do Claude Haiku 4.5 and GPT-5.6 Luna support prompt caching?
Both publish a cached-input rate: $0.100 per million for Claude Haiku 4.5 and $0.020 for GPT-5.6 Luna, against full input rates of $1.00 and $0.200. On a workload with a long stable prefix — a system prompt, a tool schema, a retrieved corpus — that changes the economics more than the headline price does.
Related comparisons
How these numbers are produced
- Price — provider list price from OpenRouter, refreshed every 15 minutes. “Blended” is a 3:1 input:output mix.
- Scores — LiveBench release 2026-09-02, using their own category map. Each model shows its strongest published run. A blank means “not evaluated”, never “bad”.
- Cost per point — the measured dollars LiveBench spent on the run, divided by the score it earned.
- “Win” — awarded only past a threshold: one full point on a benchmark score, 10% on a price, 25% on a context window. Anything tighter reports as a tie, because effort settings alone move a LiveBench score by more than that.
Published benchmarks rank models on someone else's tasks. Before committing, see LLM & agent evaluation for building an eval on your own.