OpenAI didn’t ship one new model. It shipped three, Sol, Terra, and Luna, and then claimed a state-of-the-art win on the one benchmark it chose to highlight. Look at the benchmarks OpenAI didn’t show, the ones a competitor still leads, and the “who’s best” picture stops being clean. Here is what the three tiers actually are, how they really compare, and why the most useful answer is not the one the launch slides imply.
OpenAI did something unusual this week, and the part everyone is missing matters more than the part everyone is repeating.
The repeating part is that GPT-5.6 launched. The missing part is the shape of what launched. It is not one model, it is three, Sol, Terra, and Luna, released together as a tiered family under a new naming idea, the number marks the generation and the name marks a durable capability tier that can improve on its own schedule. That alone reframes picking a model from a default choice into a deliberate decision about how much intelligence each job needs.
But the more interesting story is buried in the benchmark slides. OpenAI’s flagship, Sol, posts a state-of-the-art number on the benchmark OpenAI chose to highlight, a meaningful lead over its closest competitor. Look at the benchmarks OpenAI did not highlight, where the same competitor has published its own last-known scores, and the lead reverses. On those, the competitor is still ahead. Which means the honest answer to “is GPT-5.6 the new best model” depends entirely on which benchmark you let into the room, and nobody is saying that out loud. So this piece does two things, it walks through what the three tiers actually are and what each is for, and then it lays out the comparison honestly, where Sol wins, where it does not, and why the answer is messier than the headline.
The whole point of the GPT-5.6 family is that the three models are built for different jobs, and the pricing tells you immediately which is for which.
Sol is the one you reach for when the problem is genuinely hard. Complex reasoning, long multi-step coding sessions, advanced agent-driven workflows, security work. It’s the most capable and the most expensive, 5 dollars per million input tokens and 30 dollars per million output, the same headline price as the previous flagship. You pay for the ceiling.
Terra is the one you make your default. Priced at 2.50 dollars input and 15 dollars output, exactly half of Sol, OpenAI positions it as competitive with GPT-5.5, the previous flagship, at roughly half the cost. That’s the most quietly important claim in the whole release, because if a half-price model genuinely matches last generation’s best, Terra becomes the sensible default for most serious work and you only reach up to Sol for the hardest cases.
Luna is the one you call a lot. At 1 dollar input and 6 dollars output, it’s built for high-volume work where you’re running many calls and need each one quick and cheap. It trades ceiling for speed and cost, aimed at the bulk tasks where being good enough at low cost beats being excellent at high cost.
The structure rewards a simple habit, match each job to the cheapest tier that clears its quality bar, heavy reasoning to Sol, steady production work to Terra, high-volume jobs to Luna. And because the tier names are meant to be durable, you set that routing once and it holds as each tier improves over time.
Beyond the tiers, GPT-5.6 adds two capabilities that matter for hard work.
The first is a new maximum reasoning setting for Sol, which lets the model spend more time thinking through a difficult problem before answering. More thinking time tends to help on the hardest tasks, the ones where a fast answer is a wrong answer, and this gives Sol room to work when the problem warrants it.
The second is more interesting, an ultra mode that goes beyond a single model response by using subagents to tackle complex work. In plain terms, instead of one model doing everything in a single pass, the system can spin up multiple coordinated workers to handle a big task in parts. That’s the same multi-agent idea that has been reshaping how people build with AI, now built directly into the flagship, and it’s where Sol posts its very highest scores.
Here’s what you came for, the comparison, with the honest caveat attached to every figure, these are mostly OpenAI’s own reported results, not independent third-party benchmarks.
On Terminal-Bench 2.1, a benchmark for command-line coding work that requires planning, iteration, and tool coordination, OpenAI reports that Sol sets a new state of the art. Standard Sol lands around 88.8 percent, and in ultra mode, using those subagents, it reaches about 91.9 percent. For comparison on that same benchmark, the reported numbers put Sol ahead of Anthropic’s Claude Fable 5, which scores around 83.4 percent, a lead of roughly 5 points for standard Sol and more for ultra. The cheapest tier, Luna, reportedly ties Anthropic’s Mythos 5 on this particular benchmark, which is a striking result for the lowest-priced model in the family.
But here’s where the honesty matters, because the picture flips depending on which benchmark you look at. On other coding and reasoning evaluations, Claude Fable 5’s last published scores actually lead, it sits at about 80.3 percent on SWE-Bench Pro, around 89.8 percent on LiveCodeBench, and 59 percent on Humanity’s Last Exam, areas where OpenAI hasn’t yet published Sol’s general-availability numbers. So the accurate read isn’t that one model is simply better. It’s that Sol leads on the terminal-agent benchmark OpenAI chose to highlight, while Claude’s models lead on several others, and which one wins depends heavily on the specific task and the specific test harness. Anyone telling you there’s a clean winner is overselling it.
A recurring theme in OpenAI’s results is worth pulling out, efficiency. Across several evaluations, the claim isn’t just that GPT-5.6 matches rivals, but that it does so while spending far fewer tokens. On one cybersecurity benchmark, OpenAI reports Sol matched the performance of Anthropic’s Mythos Preview while using roughly a third of the output tokens. If that holds up in real use, token efficiency can matter as much as raw capability, because it directly lowers what a task costs to run.
On price against the broader field, one honest point stands out. Even Luna, the cheapest model in the GPT-5.6 family, is a mid-priced model in the current market, and it remains more expensive than some frontier-level competitors, such as the openly available GLM-5.2. So GPT-5.6 isn’t competing to be the cheapest option. It’s competing on capability and efficiency, and you pay a premium relative to the lowest-cost frontier models.
Two things deserve emphasis before you draw conclusions, and they matter more than any single score.
The first is the benchmark caveat, repeated because it’s important. These numbers are overwhelmingly OpenAI’s own reported evaluations, run and presented by the company, not independent third-party tests. That doesn’t make them false, vendor benchmarks under consistent conditions are a normal way to show generational progress, but they’re a starting point, not a verdict. The sensible move, as always, is to validate against your own real tasks before betting anything important on a leaderboard number, because strong benchmark results don’t automatically translate into better results on your specific work.
The second is access, which in mid-2026 matters as much as capability. GPT-5.6 launched as a limited preview, available at first only through the API and a coding tool to a small set of partners, with broad availability promised in the coming weeks. A model you can’t yet call doesn’t help you today, however high it scores, and this is a real consideration when you are comparing options you can actually use right now against one that is still rolling out. Capability you can access beats capability you can’t.
Step back and the release tells a clear story. OpenAI is moving away from one model that does everything toward a tiered family where you deliberately match the model to the job, a flagship for the hardest work, a balanced tier that aims to deliver last generation’s best at half the cost, and a cheap tier for volume. It’s pairing that with more reasoning time and a built-in multi-agent mode for the hardest tasks, and it is leaning hard on efficiency, matching rivals while claiming to spend fewer tokens.
On the comparison everyone wants, the honest answer is that there’s no single winner. GPT-5.6 Sol leads on the agentic terminal-coding benchmark OpenAI highlighted, competing models lead on several others, and the right choice depends on your specific workload, which benchmarks actually match your use, and crucially, which models you can access today. The most useful question is not which model tops a chart, but which tier of which family clears the quality bar for your particular work at the lowest cost, and that’s a question only your own testing can answer. The tiered structure is the real news here. The leaderboard position is the part to hold loosely.
If you get access to the GPT-5.6 tiers, drop a comment with how Sol, Terra, and Luna performed on your actual work, especially against whatever you are using now, because independent, real-world results are worth far more than any vendor benchmark.
GPT-5.6 vs the Frontier. The Comparison Depends on Which Benchmark You Look At was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.