OpenAI launches GPT-5.6 Sol as AI model race speeds up

OpenAI has released GPT-5.6 Sol to general users, adding another high-end AI model for developers, startups and companies weighing which tools are worth paying for. For investors watching the AI race, the launch is a fresh sign that model makers are competing on both performance and cost, not just brand recognition.

The company launched Sol alongside two smaller models, Terra and Luna, after a two-week preview that Decrypt reported was limited by the U.S. Department of Commerce to about 20 trusted partners.

OpenAI is also changing how it names its models. Sol, Terra and Luna replace the company’s usual number-only approach with named tiers. According to OpenAI, Sol is the top model, Terra is designed for everyday use and matches GPT-5.5 at half the price, and Luna is the lower-cost option.

How the pricing works

OpenAI priced Sol at $5 per million input tokens and $30 per million output tokens. Luna costs $1 per million input tokens and $6 per million output tokens, according to the company.

A token is the small unit of text or data an AI model processes. In an application programming interface, or API, companies are usually billed based on how many tokens they send into a model and how many tokens the model sends back. That makes token pricing one of the clearest ways to compare the operating cost of AI tools.

Decrypt reported that rival pricing includes $10 per million input tokens and $50 per million output tokens for Anthropic’s Claude Fable 5, $2 and $12 for Google’s Gemini 3.1 Pro, and $15 and $75 for xAI’s Grok 4.5. Chinese competitors listed in the report were cheaper: DeepSeek’s V4 Pro at $1.74 and $3.48, and Xiaomi’s MiMo v2.5 Pro at $1 and $5.

That puts Sol below the listed prices for several premium U.S. models, while still above some Chinese alternatives on output cost.

What the benchmark numbers show

OpenAI is shipping Sol with two new controls: a maximum reasoning effort setting that lets the model spend more time on a task, and an ultra mode that sends work to subagents, or smaller helper systems that split up parts of a job.

On Terminal-Bench 2.1, a test focused on command-line tasks that measures planning, tool use and repeated problem-solving, Sol in ultra mode scored 91.9%, according to the benchmark figures cited by Decrypt. Standard Sol scored 88.8%.

Those results placed Sol ahead of Anthropic’s Claude Mythos 5 at 88.0%, Claude Fable 5 at 84.3% and Claude Opus 4.8 at 78.9%. Google’s Gemini 3.1 Pro Preview scored 70.7% on the same comparison, according to the reported figures.

OpenAI also emphasized cybersecurity testing. On ExploitBench, which measures how well a model can find and use software vulnerabilities, Sol matched Anthropic’s restricted Mythos Preview while using about one-third of the tokens, according to OpenAI. The company said Sol does not cross its internal “Cyber Critical” threshold, a risk category for more dangerous cyber capabilities.

Early testers weigh in

The release follows a busy stretch for AI launches. Decrypt reported that Sol arrived one day after xAI’s Grok 4.5 and hours after Meta’s Muse Spark 1.1, while Google’s Gemini 3 from November 2025 was the oldest major flagship model in the comparison.

Early access users have started posting reactions. Theo, the CEO of AI platform T3 Chat and a developer-focused YouTube creator, said on X that Sol was “world leading in computer use” and addressed issues he had with GPT-5.5.

For retail investors, the key takeaway is the shape of the competition: model labs are trying to win users with better task performance, lower token costs and specialized settings for more complex work. OpenAI’s release gives the market another data point in a fast-moving AI spending race, while the benchmark claims will now face broader testing from developers outside the preview group.

This story draws on original reporting from Decrypt.