OpenAI launches GPT-6 Sol and Luna at half the token price of GPT-5.6, with roughly flat intelligence scores per independent analysis
OpenAI released two new API/ChatGPT models, GPT-6 Sol and GPT-6 Luna, priced 50% lower than their GPT-5.6 predecessors (Sol: $4→$2 input, $20→$10 output per million tokens; Luna: $0.20→$0.10 input, $1.20→$0.50 output). Terra was discontinued. Prompt caching was improved to a 90% discount on cached input tokens, with a new caching dashboard/diagnostics tool, and reasoning effort/tool availability can now be changed without invalidating cache. Models are available as gpt-6-sol and gpt-6-luna via API, rolling out in ChatGPT Work/Codex for paid tiers, with limited Luna access for Free/Go users on desktop. Independent analysis (Artificial Analysis) found intelligence scores essentially unchanged from GPT-5.6 (Intelligence Index: Sol 47→48, Luna flat at 37; coding agent index: Sol +2, Luna -2) with regressions on GDPval-AA v2.1 knowledge-work benchmarks (Sol -~100 Elo, Luna -~75 Elo), while per-task cost roughly halved.
Entities: OpenAI, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Sol, GPT-5.6 Luna, Terra
0 primary
What happened
OpenAI released two new models, GPT-6 Sol and GPT-6 Luna, at half the API price of their GPT-5.6 predecessors (Sol: $4 to $2 input, $20 to $10 output per million tokens; Luna: $0.20 to $0.10 input, $1.20 to $0.50 output). Terra was discontinued. Cached input tokens now get a 90% discount with a new diagnostics dashboard, and prompt caching survives changes to reasoning effort or tool settings. Independent testing by Artificial Analysis found intelligence scores essentially flat versus GPT-5.6 (Sol's Intelligence Index moved 47 to 48, Luna stayed at 37), with both models regressing on GDPval-AA v2.1 knowledge-work benchmarks.
Why it matters
This is a cost change, not a capability change, and that matters most to anyone with existing OpenAI API spend: the same task now costs roughly half as much to run, which directly shifts cost-per-task math for developers and enterprises with production workloads. It also sharpens the pricing fight with Anthropic, since OpenAI is explicitly framing this as undercutting Claude Opus 5 and Fable 5.1 on cost-per-task. For anyone choosing a model based on capability rather than budget, there is little reason to switch, since GDPval knowledge-work performance actually got worse.
What is noise
OpenAI's benchmark selection is doing a lot of work here: it highlights cost-per-task wins on OSWorld 2.0, AutomationBench, FrontierCode 1.1 and DeepSWE v1.1 while quietly omitting GDPval and Terminal-Bench 4.0, where the new models regress. It also compares against Claude Opus 5 rather than the newer, cheaper Opus 5.5, which flatters the price comparison. Treat this as a well-evidenced price cut dressed up with selective benchmarking as a performance story.
Watch next
- 01Anthropic's pricing response, particularly whether Opus 5.5 or Fable 5.1 gets a matching cut within the next month
- 02Independent benchmark trackers (Artificial Analysis, LMSYS) updating GDPval and Terminal-Bench 4.0 scores for Sol and Luna as more labs replicate the regression
- 03API usage and revenue signals from OpenAI or third-party trackers showing whether the price cut actually shifts developer volume away from Anthropic or Google
Coverage
1 storyMore economics signals
Full feed →- Cloudflare mandates AI companies to separate web crawlers for search and training1 Jul 202690
- NYT-led publishers file summary judgment brief citing internal OpenAI/Microsoft messages calling AI training "astonishing theft" and admitting chatbots substitute for journalism18 Sept 202686
- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Anthropic launches Claude Opus 5.5, matching Fable 5.1 performance at ~40% lower operating cost with faster output and less "Claudish" writing22 Sept 202679