Our thoughts on Sol, Terra and Luna

Luna

According to OpenAI, Luna is the fastest and most affordable model of the bunch, which is somewhat marketing speak for “the least capable model.” However, its job is not to be the most capable—and it is in no way incapable. Its job is to be cheap and quick. We think OpenAI is taking aim at a market segment where it currently has little to no foothold: the budget LLMLarge language model: the kind of AI model behind tools such as ChatGPT and Claude. market, a segment currently dominated by Chinese open-sourceSoftware whose source code is publicly available to inspect, modify, and redistribute. For AI models, the exact level of openness can vary. alternatives.

The model's “sticker price” is $1 per 1M input tokensSmall chunks of text or code that an AI model processes. Input tokens are sent to the model; output tokens are generated by it. and $6 per 1M output tokens. But, as you might know if you read our last note on pricing, that alone does not tell us much about the actual cost. We also need to look at the model's efficiency.

What these charts show us, is that Luna is knocking it out of the park in the budget segment. Its sticker price is higher than those of the other models—GLM-5.2 at $1.40 input and $4.40 output, Qwen3.7 Plus at $0.40 input and $1.60 output, and DeepSeek V4 Pro at $0.435 input and $0.87 output—but its efficiency means that it can still operate at a lower cost per task. At some reasoning levelsSettings that control how much computational effort a model spends working through a problem. Higher levels can improve results but usually take longer and cost more., it also achieves a higher score than the Chinese open-source models.

That said, benchmarks are not necessarily the best way to test a model's capabilities. Benchmarks can be gamed, and labs tend to train their models to perform well on the ones they know people pay attention to. The results above come from Artificial Analysis. Because each score is an index composed of several benchmarks, it should be harder to game than any single benchmark. Still, the results are worth taking with a grain of salt.

Our assessment from using the model at Closp is that it is still a budget model, and it gets into the same kinds of hijinks that budget models usually do. There is absolutely a trade-off in not only capability but also reliability when using budget models, and that trade-off is clearly evident in this model as well.

Primary use case: Agentic workflows where you need a cheaper option.

Terra

This model is probably the least interesting of the three; it is the “Sonnet” of the bunch. It offers neither top-of-the-line performance nor the eye-watering low costs of the other two. We think the OpenAI press release says it best:

Terra's sticker price is $2.50 for input and $15 for output.

We'll leave it at that.

Primary use case: Anywhere you used to use GPT-5.5. Terra will do roughly the same work at a much lower price point.

Sol

The question on everyone's mind about this model is whether it is better or worse than Fable.

We have seen people describe Fable 5 as something akin to a wise owl and GPT-5.6 Sol as a wild Rottweiler. We cannot help but agree. Sol grabs a problem by the neck and does not let go until it has finished tearing through it. It is an absolute workhorse.

However, the verdict on whether Sol is a more capable model than Fable is… mixed. Looking at the benchmarks, Sol seems to be more capable of agentic work, while Fable appears to have greater general capability.

Sol's sticker price is $5 for input and $30 for output—the same as GPT-5.5, with some differences when it comes to cachingReusing previously processed input so repeated context does not need to be computed and charged at the full input-token rate again.. See below for more details. If we are to trust the benchmarks as a measure of efficiency, Sol is considerably more efficient and ultimately costs less than Fable.

Working with Sol in an agentic coding context feels good. It is steerableAble to follow direction and adjust its approach when a user gives feedback or changes course. but still creative, meaning you can get away with vaguer descriptions of what you want and still arrive at favourable results. Here are what we consider to be some of Sol's strengths and weaknesses.

Strengths

• Extremely determined, in a good way.
• Token-efficient.
• Understands overly vague requests and makes good assumptions about what you want.
• If you use it for coding, you will find that it is better than GPT-5.5—especially for frontend work and mobile development.
• Good orchestratorA model that coordinates a larger task by planning the work, delegating parts of it, and combining the results.. It takes the time to define what a subagentA separate AI agent assigned a smaller, focused part of a larger task by the main agent. should do in a way that prior models would only do if explicitly instructed.

Weaknesses

• Extremely determined, in a bad way.
  • Does some weird things to get around privileges. Run it in a VMVirtual machine: an isolated computer environment that runs in software, helping contain risky or unexpected actions. or on a remote server.
  • Does not know when it does not have the answer.
  • At risk of going down “rabbit holes” and unnecessarily consuming a large number of tokens.
• Has a tendency to overengineer.
• A confusing number of reasoning levels.
• Consumes a significant amount of your usage allowance when run through a subscription plan.
• A poor model choice for inexperienced developers. When paired with an inexperienced developer, Sol's confidence may reinforce poor decisions, leading to unnecessary complexity and token use.

It is a really good model. We are just not blown away.

Primary use case: Advanced coding tasks that would benefit from a strong orchestrator model.

That is it for the models themselves. Next, we will look at a few things surrounding the release that are not directly related to model capability.

Sol on Cerebras

One of the more exciting pieces of news, besides the models themselves, is this quote from the release:

Cerebras builds alternative computing systems that can run AI models instead of conventional GPU-based infrastructure. Its hardware is purpose-built for AI workloads, which allows models to generate responses at exceptionally high speeds. And 750 tokens per second is insanely fast. For anyone who does not have an intuitive sense of just how fast that is, the demo below puts it into perspective.

Speed is absolutely an underrated vector, and we plan to discuss it further in a future note.

Changes in caching

For the first time ever, OpenAI will charge money for cache writesStoring newly processed tokens so they can be reused in later requests without processing the same content from scratch., thereby joining Anthropic in how it prices cache writes. Tokens written to the cache will carry a 1.25× premium over the standard input-token price. It has been pretty surprising to see that they haven't started charging for cache writes earlier, since charging for cache writes more accurately reflects the model's cost to serve, as cached tokens occupy memory.

OpenAI has also introduced a way to set explicit cache breakpointsMarkers that identify the end of a reusable prompt prefix that can be stored and reused in later requests.. A breakpoint marks the end of a reusable prompt prefix, allowing the content that follows it to change without invalidating the cached prefix. This is useful when you know that a piece of context will remain unchanged throughout a back-and-forth. The system promptThe instructions given to an AI model that define its role, rules, and general behaviour for a conversation., for example, is unlikely to change during a conversation.

Ultra and max

You can now run the models at an even higher reasoning level called max. It gives GPT-5.6 even more time than xhigh to reason and explore alternatives, run checks, and revise its approach.

They are also introducing a reasoning level called ultra. It is debatable whether it should be considered a “reasoning level” or more of a “usage setting,” since rather than reasoning even longer than max, it coordinates multiple agents across parallel workstreams. The results from using ultra look promising. However, it should be used with great care, since costs can run very high. ultra is available to Pro and Enterprise users. In Codex, it is available to Plus and higher plans.

Sources

• Artificial Analysis — Intelligence vs. Cost per Task
• OpenAI — GPT-5.6: Frontier intelligence that scales with your ambition
• OpenAI API documentation — Prompt cache breakpoints
• Artificial Analysis — AI Coding Agent Benchmarks and Leaderboard