OpenAI has launched new versions of its flagship language models, GPT-6 Sol and GPT-6 Luna, introducing substantial price reductions for developers and enterprises. With input and output rates slashed by at least half compared to the previous GPT-5.6 models, OpenAI is aiming to attract customers who increasingly evaluate models based on total task completion costs, rather than benchmark performance alone.
Major price reductions for GPT-6 Sol and Luna
GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20, respectively, for GPT-5.6 Sol. For GPT-6 Luna, the charges have dropped to $0.10 for million input tokens and $0.50 for million output tokens, compared to $0.20 and $1.20 previously. For Sol, both input and output costs are now exactly half of prior rates, while Luna’s input cost is halved and output reduced by over 58%.
OpenAI attributes these reduced rates to enhanced inference mechanisms and better caching. The company noted in its launch communication that the previous GPT-5.6 pricing was promotional, whereas GPT-6’s new rates are intended to be permanent. The GPT-6 Astra model continues as the most premium tier, set at $10 per million input tokens and $50 for output.
| Model | Input ($/M tokens) | Output ($/M tokens) | Old Input | Old Output |
|---|---|---|---|---|
| GPT-6 Sol | $2 | $10 | $4 | $20 |
| GPT-6 Luna | $0.10 | $0.50 | $0.20 | $1.20 |
| GPT-6 Astra | $10 | $50 | N/A | N/A |
Mini dictionary: OpenAI, an artificial intelligence research firm based in the US, develops and licenses advanced AI models widely used in generative applications.
Recent adjustments also affected older models, with OpenAI reducing GPT-5.6 Luna’s price by 80% and GPT-5.6 Terra’s by 20%. This aggressive pricing signals that AI model providers are competing more heavily on cost than ever before.
AI providers focus on cost per completed task
OpenAI’s pricing strategy is shifting toward economic value over technical benchmarks. Business process testing at AutomationBench shows GPT-6 Sol achieved a completion score of 32.0% at a cost of $0.34 per task, while Anthropic’s Claude Opus 5.5 scored 40.0% at $1.28 per task. Despite Opus 5.5’s higher performance score, its price per task is nearly four times that of GPT-6 Sol.
GPT-6 Sol has outperformed previous Opus 5 models in the latest Agents’ exam, providing a similar level of output with a 60% lower cost per task.
Further benchmarking by Artificial Analysis shows GPT-6 Sol holds an Intelligence Index score of 48, compared to Xiaomi’s MiMo-V2.6-Pro at 46. However, the Xiaomi model remains significantly more affordable than OpenAI’s offering.
Mini dictionary: AutomationBench is a benchmarking tool that evaluates AI model performance across various business applications by measuring both task completion rate and cost efficiency.
Anthropic responds, market moves intensify
Anthropic, a leading AI competitor, launched Claude Opus 5.5 on the same day at $4 per million input tokens and $20 for output, marking a 20% reduction from the previous Opus 5. Anthropic claims this update will cut user spending by 40%, as the latest model requires fewer tokens per task.
At $2 per million input, GPT-6 Sol undercuts Claude Opus 5.5’s $4 rate and aligns with Claude Sonnet 5. The highest-priced models, GPT-6 Astra and Claude Fable 5.1, both charge $10 per million input tokens and $50 for output, according to VentureBeat.
OpenRouter has noted that DeepSeek, another AI provider, doubled its token share platform-wide to 18% by June. Chinese AI models collectively overtook their US counterparts in token share the same month. Xiaomi now offers its MiMo models under an MIT license, letting firms self-host rather than pay per token.
Mini dictionary: Anthropic, founded by former OpenAI team members, specializes in building advanced, safe large language models, most notably the Claude product line.
| Provider | Model | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|---|
| OpenAI | GPT-6 Sol | $2 | $10 |
| Anthropic | Claude Opus 5.5 | $4 | $20 |
| OpenAI | GPT-6 Astra | $10 | $50 |
| Anthropic | Claude Fable 5.1 | $10 | $50 |
Cheaper tokens, rising total costs
A report from Gartner projects that inference costs per agentic workflow could increase fivefold by 2028, warning that even as token prices fall, total AI spending may rise due to heightened workflow complexity.
Each successive generation of AI models will require more, and often pricier, tokens, resulting in higher overall costs.
A University of Oxford study found that quality-adjusted inference prices have dropped far faster than traditional indexes suggest. Nonetheless, as advanced models consume more tokens per task, the average buyer’s costs per completed job have stopped declining.
Lower token costs may not guarantee reduced overall AI expenses. Instead, cheaper pricing could simply encourage expanded usage and greater demand for services.
