OpenAI's budget GPT-6 tier, released September 22, 2026 at $0.10/$0.50 per million tokens - a tenth of Claude Haiku 4.5's input price, half of GPT-5.6 Luna. Context is the same 1.05M tokens as its bigger siblings. The number nobody expected is DeepSWE v1.1 at 66.6%: 2.2 points behind Sol for about a twentieth of the cost. The catch lands on anything needing judgment - 20.7% on AutomationBench and a 7.6% factual error rate. Luna is a worker, not a planner.
Parameters
Undisclosed
Context Window
1.05M
License
Proprietary
Release Date
2026-09-22
API Pricing
Input Price (per 1M tokens)
$0.1
Output Price (per 1M tokens)
$0.5
Billing Mode: standard
Strengths
- •66.6% on DeepSWE v1.1 at $0.22 per task - 2.2 points behind Sol for about 5% of the cost
- •$0.10/$0.50 enters territory usually left to self-hosted open weights
- •Same 1.05M-token context as the top tiers; free and Go users get it in the desktop app
Weaknesses
- •20.7% on AutomationBench and 52.7% on OSWorld 2.0 - planning and computer use fall off a cliff
- •7.6% factual error rate, nearly double GPT-6 Astra's 3.9%
- •Above 272K input tokens a 2x input / 1.5x output multiplier applies
Use Cases
- •Routing, classification, extraction and bulk summarization
- •First-pass code generation reviewed by a human or a stronger model
- •Long-document pipelines where volume, not judgment, decides the bill
Deep Analysis
Input / Output Price
$0.10 / $0.50 per 1M tokens
Cached read $0.01; a tenth of Claude Haiku 4.5's input rate
Positioning
GPT-5.6 Sol parity at high reasoning effort
OpenAI-reported, at roughly 1% of the older model's cost
Context Window
1.05M tokens
128K max output; 272K input threshold triggers 2x input and 1.5x output
Safety Movement
Refusal rollback 77% -> 42%
The largest improvement of the three September 22 launches
Availability
Free and Go tiers on desktop
Also on the OpenAI API, ChatGPT Work and Codex for paid plans
Release
Sep 22, 2026
Knowledge cutoff May 18, 2026
Strengths
- ・Cheapest entry into the GPT-6 family by an order of magnitude, at $0.10 input and $0.50 output per million tokens
- ・Cuts refusal rollback from 77% to 42%, the largest safety movement announced across the three launches
- ・Reaches GPT-5.6 Sol level at higher reasoning effort according to OpenAI, at about a hundredth of the cost
- ・Reaches free and Go users on desktop, making it the widest-reach model of the day
Weaknesses
- ・No Artificial Analysis Intelligence Index score has been published for Luna, so its ceiling is unmeasured independently
- ・Lowest capability tier of the GPT-6 family; it is an efficiency model, not a reasoning flagship
- ・No system card at launch, and reasoning cannot be fully switched off
- ・Carries the same 272K-token surcharge as Sol, which undercuts the headline rate on long documents
Competitor Comparison
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| GPT-6 Sol | AA 47.5 (max) | N/A | N/A | $2/$10 |
| Claude Haiku 4.5 | N/A | N/A | N/A | $1/$5 |
| GPT-5.6 Sol (promo) | AA 59 | N/A | N/A | $4/$20 |
GPT-6 Luna is the third release of September 22 and the one that moves the floor rather than the ceiling. At $0.10 per million input tokens and $0.50 per million output it costs a tenth of Claude Haiku 4.5 on input, a price bracket usually occupied by self-hosted open weights rather than a frontier lab's API. OpenAI says it matches GPT-5.6 Sol at higher reasoning effort at roughly 1% of that model's cost, and reports its refusal rollback falling from 77% to 42% — the biggest safety improvement of the three models announced that day. It carries the same 1.05M-token window and the same 272K surcharge as Sol, and no Independent Intelligence Index score has been published for it.
Sources
Analysis generated: 2026-09-23