GPT-6 Luna lists at $0.100/1M / $0.500/1M per 1M tokens. Claude Haiku 4.5 lists at $1.00/1M / $5.00/1M. Luna undercuts Haiku on stickers in this catalog. Cache reads, context windows, and Exact vs Approx counts still decide real bills. Rates checked 23 Sep 2026.
Compare high-volume turns
Paste a typical chat or routing turn into the cost calculator twice. Claude stays Approx in-browser. Confirm shipping budgets with Anthropic count_tokens when the spend is contractual.
Sticker rates: Luna is cheaper
| Meter | GPT-6 Luna | Claude Haiku 4.5 |
|---|---|---|
| Input / 1M | $0.100/1M | $1.00/1M |
| Output / 1M | $0.500/1M | $5.00/1M |
| Cache read / 1M | $0.010/1M | $0.100/1M |
| Cache write / 1M | $0.125/1M | $1.25/1M |
| Batch input / output | $0.050/1M / $0.250/1M | $0.500/1M / $2.50/1M |
| Context window | about 1.05M (272K cliff) | 200K |
| Tokenizer in TokenCalculator | Exact (o200k) | Approx (confirm with count_tokens) |
Where the invoice diverges
1. Stickers
Luna is an order of magnitude cheaper on input and 10x cheaper on output versus Haiku 4.5 in this catalog. On equal token counts, Luna wins the meter for high-volume chat and routing.
2. Cache and context
Luna cache reads are $0.01 per 1M. Haiku cache reads are $0.10 per 1M. Luna also offers a 1.05M window with a 272K pricing cliff. Haiku stays at 200K with flat list rates across that window. Fat RAG dumps favor careful Luna cliff design or a different tier.
3. Quality and tooling
Haiku can still win if Claude quality, Claude Code, or Anthropic tooling clears your evals and Luna does not. Rate cards rank price. Workloads rank finished-task cost.
Worked numbers (planning only)
Example A, uncached: 10K input + 2K output. Luna about $0.002. Haiku about $0.02.
Example B, cache heavy: 50K cached + 2K fresh + 1K output. Luna about $0.0005 cache + $0.0002 fresh + $0.0005 output about $0.0012. Haiku about $0.005 cache + $0.002 fresh + $0.005 output = $0.012.
Workload cheatsheet
- High-volume chat and routing: lean Luna on the meter unless Claude quality wins.
- Anthropic-standardized stacks: keep Haiku when switching vendors costs more than stickers save.
- Prompts past ~272K on Luna: model the long-context band; Haiku caps at 200K context.
- Move to Sol or Opus 5.5 when either budget model fails coding or agent evals.
Common mistakes
- Picking Haiku on brand loyalty without pricing equal prompts
- Budgeting Haiku with OpenAI Exact token counts
- Ignoring Luna Fast mode 2x on interactive paths
- Comparing GPT-5.6 Luna to Haiku instead of GPT-6 Luna
Frequently asked questions
Is GPT-6 Luna cheaper than Claude Haiku 4.5?
Yes on list rates in this catalog. Luna is $0.100/1M / $0.500/1M. Haiku is $1.00/1M / $5.00/1M.
GPT-6 Luna vs Claude Haiku 4.5 for high volume?
Luna usually wins the meter on equal tokens. Pick Haiku when Claude quality or tooling is required.
Are browser token counts Exact for both?
Luna is Exact (o200k). Haiku is Approx in TokenCalculator. Confirm Claude with count_tokens.
Does Luna have a long-context surcharge Haiku lacks?
Luna reprices the full request above 272K input. Haiku publishes flat rates inside its 200K window and does not offer Luna-scale context.
How does GPT-6 Luna compare to GPT-5.6 Luna?
GPT-6 Luna halves input versus GPT-5.6 Luna ($0.200/1M) and cuts output from $1.20/1M to $0.500/1M.
When should I use Sol instead?
When Luna fails coding or multi-step agent evals. Sol is the usual OpenAI step up.
Are these rates live?
No. Curated catalog rates last verified 23 Sep 2026. Confirm on OpenAI and Anthropic pricing pages.
What should I open first?
Paste one production turn into the cost calculator on Luna and Haiku 4.5, then open the compare pair URL.
Next steps
- GPT-6 Luna pricing
- GPT-6 Sol pricing
- Claude API pricing hub
- Cheapest LLM API hub
- How we verify rates
- Official Anthropic pricing
Related Sep 2026 pricing guides
Frontier rate cards, cost-per-task math, and long-context cliffs. Cross-link these when you compare Sol, Luna, Opus 5.5, Grok 4.7, or Astra.