OpenAI Releases GPT-5.6 Sol: Here’s How It Stacks Up Against Other AI Models - Decrypt

In brief

OpenAI has released GPT-5.6 Sol, alongside the cheaper Terra and Luna, ending a two-week preview the U.S. Department of Commerce kept boxed in.

Sol in ultra mode tops Terminal-Bench 2.1 at 91.9% and matches Anthropic's restricted Mythos Preview on ExploitBench while burning roughly a third of the tokens.

The launch lands one day after Grok 4.5 and hours after Meta's Muse Spark 1.1, leaving Google's November 2025 Gemini 3 as the oldest frontier flagship still standing.

OpenAI's GPT-5.6 Sol is now public. The company released its new flagship model to general users today, launching alongside two smaller siblings, Terra and Luna, after the U.S. Department of Commerce kept a preview restricted to about 20 trusted partners for two weeks.The naming strategy is new for OpenAI. This is the first time the company gives names to its models instead of simply using numbers. Sol, Terra, and Luna mark capability tiers that can move on their own cadence. Sol is the flagship, Terra the everyday model OpenAI says matches GPT-5.5 at half the price, and Luna the cheaper option.Pricing runs $5 and $30 per million input and output tokens for Sol, dropping to $1 and $6 for Luna. (Tokens, for those not in the know, are the smallest unit of information a model can handle. And companies typically price their models on a per token basis for API services) Two new knobs ship with it: a max reasoning effort that lets Sol think longer, and an ultra mode that farms work out to subagents.For context, Anthropic charges $10/$50 for Claude Fable 5, Google charges $2/$12 for Gemini 3.1 Pro, xAI charges $15/$75 for Grok 4.5.On the Chinese side: DeepSeek charges $1.74/$3.48 for V4 Pro, and Xiaomi charges just $1/$5 for MiMo v2.5 Pro—placing Sol between the premium U.S. frontier models and China's low-cost challengers.What the benchmarks showOn Terminal-Bench 2.1—a test of command-line workflows that reward planning, tool use, and iteration, scored as the share of tasks a model completes—Sol in its ultra configuration hit 91.9%, with standard Sol at 88.8%.That puts both 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 trailed the chart at 70.7%.OpenAI leaned hardest on cyber. On ExploitBench, which measures how well a model finds and weaponizes software vulnerabilities, Sol matched the restricted Mythos Preview while spending roughly a third of the tokens. OpenAI says Sol still doesn't cross the "Cyber Critical" line in its own risk framework.The testers already have opinionsEarly access was loud. Theo, a well-known developer, AI youtuber and CEO of the AI platform T3 Chat, called Sol "world leading in computer use" and said it fixed the complaints he had with GPT-5.5.