OpenAI is now gradually rolling out GPT-5.6 worldwide. A limited preview, imposed due to concerns from the U.S. government about the model’s safety, preceded the actual release. For now, OpenAI does not seem able to match Anthropic’s Claude Fable 5, even though its efficiency (and relatively low price) are impressive.
New LLMs are popping up left and right these days: Fable 5, Sonnet 5, GLM-5.2, Grok 4.5, and Muse Spark 1.1 were all impressive in their own way. Now comes OpenAI’s response to the latest Anthropic models. Or is that really the case?
The last GPT-5.x?
Reportedly, GPT-5.6 is the last OpenAI model with this architecture and pre-training. GPT-6 is even expected to arrive sooner than previously thought, possibly as the company’s true response to Fable 5. At the same time, GPT-5.6 has brought us significantly closer, if we look at the Artificial Analysis score for general intelligence. With a score of 59 for GPT-5.6 Sol, 55 for Terra, and 51 for Luna, the new LLMs fall almost exactly between Fable 5 and Sonnet 5. The mid-sized Terra, a sort of Sonnet variant, as Anthropic would call it, seems to outperform Sonnet, at least.
More striking is the coding performance. In certain benchmarks, measured both externally by Artificial Analysis and by OpenAI itself, GPT-5.6 Sol is significantly stronger than Fable 5. This is an impressive achievement, especially considering that this model architecture may have reached its limits.
Exactly what this new architecture entails remains to be seen. With previous breakthroughs like GPT-4 and o1, it took a few months before the industry was largely familiar with the new “trick.” Think of o1’s “test time compute,” also commonly referred to as “reasoning.” Essentially, it was a new way to rely more on inference rather than training, with specific training of the model to tackle a task step by step. In less than half a year, DeepSeek R1 demonstrated that the o1 technique was not only reproducible but also relatively inexpensive.
We mention all this because GPT-5.6 is emphatically not a new GPT-4 or o1. Naming conventions are subjective, but we can say that the transition from Opus 4.8 to Fable 5 rightly involved a completely new number. For GPT-5 versus GPT-4o, the designation signaled that the underlying methodology had changed.
Sol, Terra, Luna
OpenAI has named its most powerful model Sol, which it claims delivers state-of-the-art results in coding, knowledge work, cybersecurity, and science. A new feature is the “Ultra” mode, in which four agents work on a task in parallel by default. Additionally, each model has a separate reasoning function, resulting in a total of 4 modes per model, 3 models, and 1 additional mode for Sol via Ultra.
The pricing clearly distinguishes the three models. Sol costs $5 per 1 million input tokens and $30 per 1 million output tokens. Terra is priced at $2.50 and $15, while Luna is priced at $1 and $6. The models are available via ChatGPT, Codex, and the OpenAI API. Plus, Pro, Business, and Enterprise users have access to Sol; free users must use Terra.
According to previously shared benchmarks, the variants are competitive with all LLMs except Fable, with Sol approaching that top level in Ultra mode. On ExploitBench, Sol scores 73.5 percent compared to 47.9 percent for GPT-5.5. However, according to independent measurements, Fable 5 remains the strongest in general intelligence.
Competition is heating up
To emphasize it once more, the timing makes this week a true LLM week. Shortly before GPT-5.6, Grok 4.5 from the rebranded SpaceXAI and Meta’s Muse Spark 1.1 were released. Both operate at a level comparable to Opus 4.8 and GPT-5.5 and have a strong focus on coding.
OpenAI had previously postponed its own IPO until 2027, while maintaining its target valuation of one trillion dollars. OpenAI linked the broad rollout of GPT-5.6 to a temporary security review by the U.S. government. OpenAI states that the models are more capable than previous models in both domains, biology and cybersecurity, but do not cross the critical threshold. According to the company, Sol’s cybersecurity safeguards block about ten times more potentially harmful activity than before.
For the most demanding security tasks, access remains limited to verified users within OpenAI Daybreak’s Trusted Access program. OpenAI says it conducted its most intensive security testing to date ahead of the launch, including approximately 700,000 GPU-hours of automated red teaming.