OpenAI Slashes API Costs by 50% with GPT-6 Sol and Luna Models

OpenAI Launches GPT-6 Sol and Luna with Dramatic Cost Reductions

OpenAI released GPT-6 Sol and GPT-6 Luna on Tuesday, delivering a 50% cost reduction across its mid-tier and budget AI model offerings1

1

5

5

. The new AI models represent OpenAI's latest move in an intensifying competitive AI landscape, arriving just 90 minutes after Anthropic unveiled Claude Opus 5.54

4

. GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from GPT-5.6 Sol's $4 and $20 respectively3

3

. GPT-6 Luna delivers even steeper savings at $0.10 per million input tokens and $0.50 per million output tokens, representing a 50% reduction on input and 58.3% on output compared to its predecessor5

5

. These are permanent prices, not promotional rates, according to an OpenAI spokesperson5

5

.

Doubled Accuracy and Improved Performance Benchmarks

Beyond cost reduction, OpenAI claims GPT-6 Sol makes approximately half as many factual errors as GPT-5.6 Sol2

2

. If the previous model generated errors 20% of the time, GPT-6 Sol would reduce that to roughly 10% on identical queries2

2

. GPT-6 Luna at higher effort levels now matches GPT-5.6 Sol's performance at approximately one-hundredth of the cost2

2

. The models were trained using similar methods as GPT-6 Astra, released in early September, though Astra remains OpenAI's flagship for the most complex tasks1

1

. OpenAI's benchmark comparisons directly target Anthropic Claude Opus 5.5 across multiple domains. On AutomationBench, which tests professional workflows across 47 tools, GPT-6 Sol scored 33.2% at $0.27 per task, beating Claude Opus 5's 26.9% at 11.1 times the cost4

4

. In coding tests using DeepSWE 1.1, GPT-6 Sol achieved 68.8% compared to Claude Fable 5's 69.9%, but at roughly 80% lower cost per task4

4

.

Infrastructure Improvements Drive Price War Escalation

OpenAI attributes the dramatic cost reduction to improvements in caching and inference that allow more efficient model serving at scale3

3

. The company raised default cache hit rates with cached input tokens now carrying a 90% discount4

4

. Developers can now set explicit breakpoints to define where cached prompt prefixes end and adjust reasoning effort or tools without losing cached context4

4

. GitHub reported these caching improvements cut the share of prompt tokens requiring fresh processing by more than half across billions of requests over recent months4

4

. Ara Kharazian, lead economist at Ramp, characterized the simultaneous releases as a price war driving down AI pricing and potentially limiting profitability for both companies4

4

. The pressure stems partly from Chinese open-weight models from Alibaba, DeepSeek and others offering competitive performance at lower price points4

4

.

Enterprise Use Cases and Competitive Positioning

GPT-6 Sol targets complex work including coding, debugging, feature development and data analysis that developers and knowledge workers perform repeatedly5

5

. GPT-6 Luna focuses on high-volume tasks with clear objectives like summarization, document extraction and answering straightforward questions5

5

. Both models rolled out to ChatGPT Work and Codex, covering Plus, Pro, Business, Enterprise and Edu accounts, while free and Go users receive Luna access in the desktop app4

4

. At $2/$10 per million tokens, GPT-6 Sol matches Claude Sonnet 5's permanent pricing but costs 50% less than the new Claude Opus 5.5's $4/$20 rates5

5

. Google Gemini 3.8 Flash remains more aggressive at $0.75/$3.75 under introductory pricing through December 31, increasing to $1.50/$7.50 on January 1, 20275

5

. The rapid improvement timeline proves striking—GPT-5.6 Sol and Luna launched in July, meaning OpenAI doubled factual accuracy and dramatically reduced costs in under three months2

2

.