Summary
- OpenAI has introduced GPT-5.6 Sol, along with its more affordable counterparts, Terra and Luna, concluding a two-week exclusive preview by the U.S. Department of Commerce.
- In ultra mode, Sol achieved a score of 91.9% on Terminal-Bench 2.1 and matched Anthropic's restricted Mythos Preview on ExploitBench, using approximately one-third of the tokens.
- The release comes just a day after Grok 4.5 and hours after Meta's Muse Spark 1.1, leaving Google's Gemini 3 from November 2025 as the oldest flagship model still available.
OpenAI has made its latest flagship model, GPT-5.6 Sol, publicly accessible. This marks the first time OpenAI has released models with unique names—Sol, Terra, and Luna—rather than just numerical identifiers, following a two-week period where only about 20 partners could preview the model.
The new naming convention represents different capability tiers: Sol is the premium model, Terra serves as the everyday option priced to match GPT-5.5 at a lower cost, and Luna is the most economical choice.
Pricing details indicate Sol costs $5 and $30 per million tokens for input and output, respectively, while Luna is priced at $1 and $6. (Tokens represent the smallest data unit a model can process, with companies typically charging on a per-token basis for their API services.) Two new features have been introduced: a maximum reasoning effort that allows Sol to think for longer periods, and an ultra mode that delegates tasks to subagents.
For comparison, Anthropic's Claude Fable 5 is priced at $10/$50, Google's Gemini 3.1 Pro at $2/$12, and xAI's Grok 4.5 at $15/$75.
On the Chinese market, DeepSeek's V4 Pro is available at $1.74/$3.48, while Xiaomi's MiMo v2.5 Pro is offered at $1/$5, positioning Sol between U.S. premium models and lower-cost Chinese competitors.
Benchmark Performance
In the Terminal-Bench 2.1 tests, which assess command-line workflows emphasizing planning, tool usage, and iteration, Sol in its ultra mode scored 91.9%, while the standard version achieved 88.8%.
Both configurations outperformed Anthropic's Claude Mythos 5 at 88.0%, Claude Fable 5 at 84.3%, and Claude Opus 4.8 at 78.9%, with Google's Gemini 3.1 Pro Preview lagging at 70.7%.
OpenAI's focus on cybersecurity was evident in the ExploitBench results, where Sol matched the restricted Mythos Preview while utilizing about one-third of the tokens. OpenAI notes that Sol still does not meet the threshold for "Cyber Critical" in its risk assessment framework.
Feedback from Testers
Apparently, I can now discuss GPT-5.6?
It's an excellent model. While it may not be as "intelligent" as Fable, it's highly capable and resolves all the concerns I had with GPT-5.5.
It's incredibly persistent, running for an entire day without needing a /goal. It understands subagents well…
— Theo - t3.gg (@theo) July 8, 2026
Dan Shipper, whose team at Every tested Sol for a month, provided a succinct comparison: "GPT-5.6 is like a Porsche, while Fable is akin to a warp drive." He believes Sol is ideal for everyday tasks, whereas Fable is more suited for extraordinary needs.
GPT-5.6 is like a Porsche, Fable is like a warp drive.
We've been testing internally at @every for about a month. GPT-5.6 offers the best combination of power, speed, and performance for daily knowledge tasks and coding.
Fable is a different type of tool. If you need to get…
— Dan Shipper 📧 (@danshipper) July 8, 2026
Researcher Daichi Konno, who tested Sol early on, concluded that it significantly surpasses GPT-5.5 and is competitive with Fable 5, although he noted that Anthropic still leads in writing tasks. He highlighted that Sol's safeguards did not activate during life-science queries, suggesting it could become a preferred tool for biology researchers.
Having interacted with GPT-5.6-Sol, it's evident that its performance is clearly superior to GPT-5.5, and it is comparable to Claude Fable 5!
1. In writing tasks, Fable has the edge.
2. The safeguards did not trigger on life-science topics.The second point is crucial, as GPT-5.6-Sol may become the first choice for bio-researchers! https://t.co/pQmciCQH3q
— Daichi Konno / 紺野 大地 (@_daichikonno) July 9, 2026
Speculative insights suggest that GPT-5.6 may be the last in the 5.x series, with a larger GPT-6 expected to launch within a month. This timeline aligns with claims from a roadmap-tracking account, "Synthwave," about upcoming releases from Anthropic and DeepSeek.
The timing of Sol's launch is noteworthy as it coincides with Anthropic's Fable 5 transitioning out of subscription plans after its global reintroduction on July 1, which limited its usage credits after the weekly allowance expired on July 7.
The competitive landscape is intensifying. SpaceXAI released Grok 4.5 recently at a significantly lower price, which Elon Musk described as "comparable to Opus 4.7 but much faster." Meta introduced Muse Spark 1.1 today, marking its entry into paid models. While neither leads the field, both are considered top-tier models.
Google remains the only major U.S. lab yet to refresh its flagship model during this period. Its top-tier offering, Gemini 3, has been available since November 2025, making it the oldest of the major models still in circulation, while OpenAI, Anthropic, xAI, and Meta have all released new models in the past week.