OpenAI cut GPT-5.6 Luna’s API price by 80% on July 30 — and it barely made a ripple. At $0.20 per million input tokens, Luna is now the cheapest capable model from a Western AI lab, undercutting Claude Haiku 4.5 by more than four times. If you’re running classification, summarization, or routing workloads on older OpenAI models, the math just changed. Whether you’ve noticed yet is a different question.
What Actually Changed #
The GPT-5.6 family launched on July 9 with three tiers. Luna is the fast, cheap tier. Terra is the middle. Sol is the flagship. OpenAI dropped Luna’s price by 80% and Terra’s by 20%, effective July 30:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Change |
|---|---|---|---|
gpt-5.6-sol |
$5.00 | $30.00 | — |
gpt-5.6-terra |
$2.00 | $12.00 | –20% |
gpt-5.6-luna |
$0.20 | $1.20 | –80% |
In real terms: a workload consuming 50 million input tokens per month dropped from $50 to $10. At 100 million tokens, you’re saving $80 a month on input alone — nearly $1,000 a year for doing nothing except updating a model string.
The Trap You Need to Avoid #
This is where developers are going to lose money: the default gpt-5.6
alias routes to Sol, not Luna. If you swap to the GPT-5.6 family without specifying the tier, you’re paying $5/M instead of $0.20/M.
model = "gpt-5.6"
model = "gpt-5.6-luna"
Luna is available through the OpenAI API, ChatGPT Work, Codex CLI, GitHub Copilot, and Amazon Bedrock. It is not the default model in the standard ChatGPT consumer interface — you need API access or ChatGPT Work to use it programmatically.
What Luna Is Actually Good At #
Luna is not a toy. On the Artificial Analysis Intelligence Index, Luna scores 51 against Claude Haiku 4.5’s 30. It generates 148 tokens per second versus Haiku’s 99. It accepts a 1,050,000-token context window. For a model at this price point, those are strong numbers.
- Classification and routing (intent detection, category labeling)
- Extraction and transformation (structured output from unstructured text)
- Summarization (documents, transcripts, long-form content)
- Real-time assistants where latency matters more than depth
- Large batch jobs where per-token cost is the primary constraint
Skip Luna for anything requiring sustained multi-step reasoning, complex code generation over large codebases, or long-horizon agentic work. That’s what Sol is for. OpenAI’s own Codex CLI switched its auto-review feature to Luna — which tells you something about where the quality ceiling sits for code review tasks, but also its limits for generating production code.
The Competitive Pressure Behind the Cut #
OpenAI announced this price cut alongside a milestone: one billion weekly active ChatGPT users and two million businesses, reached less than four years after launch. The timing is deliberate. But the more important context is competitive pressure.
Chinese models are cheap. DeepSeek Flash runs around $0.14 per million input tokens. Qwen 3.6-35B sits at $0.15. Luna at $0.20 is now in the same ballpark — exactly where OpenAI needs to be to compete on price. VentureBeat called it directly: this is an AI price war, and Luna just became OpenAI’s frontline model for cost-sensitive workloads.
For comparison, Claude Haiku 4.5 — Anthropic’s cheapest tier — costs $1.00 per million input tokens. That’s five times Luna’s new price. Anthropic has quality arguments (Haiku outperforms Luna on coding and agent tasks per independent benchmarks), but for pure volume workloads, the cost gap is hard to justify without a clear quality advantage on your specific tasks.
Should You Switch? #
Run these checks before migrating any production workload to Luna:
Identify your Luna-eligible workloads— classification, extraction, summarization, routing. Anything currently on GPT-4o mini or GPT-4.1 mini is worth evaluating first.Test output quality on your actual tasks— benchmark scores don’t always transfer to production prompts. Run a sample before committing.** Use the explicit model ID**—gpt-5.6-luna
, notgpt-5.6
.Keep Sol or Terra for reasoning-heavy paths— route by task complexity, not by blanket model replacement.** Check Bedrock availability**if you’re on AWS — Luna is listed asopenai.gpt-5.6-luna
on Amazon Bedrock with the same pricing.
The 80% price cut is significant on its own. But the real gain is building a routing layer that sends cheap tasks to Luna and complex tasks to Sol or Terra. That’s where savings compound across a system — and where most teams haven’t looked yet.