On July 9, 2026, OpenAI moved its most capable system into broad public availability. The notable part is not that GPT-5.6 exists. It is that OpenAI released it as three models on the same day, plus an agent designed to complete jobs rather than answer prompts. The lineup is named Sol, Terra and Luna, and the shape of the release tells you as much about where the market is heading as any individual score.
This piece walks through what each tier is, how they compare on independent benchmarks, what the new ChatGPT Work agent is trying to do, and the strategic logic behind splitting one launch into three products.
Three models, one family
For most of the GPT era, an OpenAI launch meant a single flagship with smaller variants trailing behind. GPT-5.6 inverts that. The three tiers are positioned as deliberate points on a curve rather than a headliner plus its budget cousins.
| Model | Positioning | Price (input / output per 1M tokens) | Intelligence Index (max) |
|---|---|---|---|
| Sol | Frontier reasoning, long-horizon agentic work | $5.00 / $30.00 | 59 |
| Terra | Balanced everyday model, GPT-5.5-class quality | $2.50 / $15.00 | 55 |
| Luna | Fastest and cheapest, high-volume work | $1.00 / $6.00 | 51 |
All three share the same underlying platform features: a one-million-token context window, up to 128,000 output tokens, and a February 16, 2026 knowledge cutoff. They are available through ChatGPT, the Codex coding environment, and the OpenAI API.
GPT-5.6 spreads across the leaderboard, not just the top
Artificial Analysis Intelligence Index v4.1 (9 evaluations). Data from Artificial Analysis and BenchLM. Higher is better.
The three GPT-5.6 tiers (green) sit at 59, 55 and 51, letting a buyer pick a point on the intelligence curve rather than a single flagship. Sol lands one point behind Claude Fable 5.
On the Artificial Analysis Intelligence Index v4.1, a composite of nine hard evaluations, Sol scores 59, one point behind Anthropic's Claude Fable 5 at 60. Terra lands at 55, level with the previous GPT-5.5 flagship, and Luna at 51, roughly matching GLM-5.2 and Gemini 3.5 Flash. On individual tests, OpenAI reports Sol at 94.6% on GPQA Diamond and 64.6% on SWE-Bench Pro.
The single most useful way to read that chart is not "Sol is nearly the best." It is that OpenAI now occupies three separate rungs of the ladder with one release. A buyer no longer chooses between OpenAI and a cheaper competitor. They choose which OpenAI tier fits the task.
The value question: intelligence per dollar
Raw capability is only half of any real purchasing decision. The other half is what that capability costs to run at volume. Plotting the same models by price and intelligence reframes the launch.
Mapping intelligence against price
Blended cost per million tokens (weighted 3:1 input:output) versus Intelligence Index. Toward the top-left is more capability for less money. Pricing from provider API pages, July 2026.
Terra and Luna (green) push OpenAI into the value corner Grok 4.5 opened up, while Sol competes at the top on capability rather than price.
Terra is the quietly interesting model here. At $2.50 input and $15 output per million tokens, it delivers GPT-5.5-level quality at roughly half the price of the model it replaces. Luna pushes further into the low-cost corner that Grok 4.5 opened earlier in the year, landing near a blended cost of just over two dollars per million tokens. Sol, by contrast, is not trying to win on price. It competes at the top of the capability axis, where a few points of Intelligence Index can matter more than the bill.
This is a familiar pattern in maturing technology markets. Once a capability frontier is established, competition migrates from "who is smartest" toward "who delivers a given level of quality most cheaply." The three-tier structure lets OpenAI fight on both fronts at once instead of picking one.
ChatGPT Work: from answers to jobs
Alongside the models, OpenAI introduced ChatGPT Work, an agent framed around completing multi-step tasks rather than returning a single response. The distinction is worth taking seriously because it changes what "good" means.
A chat model is judged on the quality of a reply. An agent is judged on whether a job finished correctly, how many steps it took, how many tokens it burned, and how often a human had to intervene. Those are different success criteria, and they reward different model traits: reliability across long tool-use chains, token efficiency, and graceful recovery from errors. Sol's positioning around "long-horizon agentic work" is a direct response to that shift.
The practical caveat is that agentic reliability is genuinely hard to benchmark. A model can look excellent on a static coding test and still stall on a messy real-world workflow with ambiguous instructions and flaky tools. Independent, task-based evaluations of ChatGPT Work will matter more than the headline scores, and those take time to accumulate.
The speed story: GPT-5.6 Sol on wafer-scale silicon
One number from the launch stands slightly apart from the rest. OpenAI confirmed that GPT-5.6 Sol is being served on Cerebras wafer-scale hardware at up to 750 tokens per second for select customers, which the companies describe as roughly fifteen times the speed of a frontier model on conventional batched GPU serving.
Speed is not a vanity metric for agentic work. When a model is executing a chain of dozens of tool calls, latency compounds at every step, and a task that takes minutes on slow infrastructure can take seconds on fast infrastructure. As agents move from demos into production, tokens per second starts to behave like a real product feature rather than a spec-sheet footnote. We looked at that shift in more depth in our analysis of the inference speed race.
What the tiered launch signals
Step back from the individual numbers and a strategy comes into focus. Releasing Sol, Terra and Luna together does three things:
- It segments the market by willingness to pay. Frontier buyers get Sol. Cost-sensitive high-volume workloads get Luna. The large middle gets Terra. No single competitor undercuts the whole stack.
- It reframes the benchmark conversation. When the story is "here is our smartest model," a rival one point ahead wins the headline. When the story is "here is a family that covers the curve," a one-point gap at the very top matters far less.
- It optimises for the agentic era. Different steps in an automated workflow need different models. A planning step may call Sol, a bulk classification step may call Luna, and a routine draft may call Terra. A single provider offering all three simplifies that routing.
That last point is the one most likely to shape how teams actually build. The emerging best practice is not to standardise on one model but to route each task to the model that fits it on quality, cost and speed. GPT-5.6 makes that easier within OpenAI's own lineup, and the same logic extends across providers: the right default for reasoning, coding, high-volume extraction and research are rarely the same model. Platforms like Metir AI exist precisely to make that cross-provider routing one decision instead of six subscriptions.
The honest limitations
A neutral read requires the caveats. First, Sol trails Claude Fable 5 on the composite index and sits close to several rivals rather than clearly above them, so "most powerful OpenAI system yet" is true without meaning "clearly the best model available." Second, the February 2026 knowledge cutoff means anything after that date requires retrieval or tool use, which matters for fast-moving domains. Third, the most differentiated claims, particularly around ChatGPT Work and the Cerebras speed tier, are the hardest for outsiders to verify today and will need independent testing before they can be treated as settled.
The takeaway
GPT-5.6 is less a single leap than a repositioning. By shipping three tiers and an agent at once, OpenAI is competing on the full curve of intelligence versus cost rather than staking everything on the top of the leaderboard. Sol keeps OpenAI within a point of the frontier, Terra resets the price of "good enough for most work," and Luna chases the high-volume, cost-sensitive base that has become one of the most contested parts of the market.
For anyone choosing models in mid-2026, the practical lesson is the same one the release itself embodies: the winning move is rarely a single model. It is matching each job to the tier, and the provider, that does it best.
Work across every frontier model in one place
The smartest way to use GPT-5.6 is not to commit to it exclusively but to route each task to whichever model handles it best. With Metir AI you get unified access to GPT-5.6, Claude, Grok, Gemini and other leading models in a single workspace: send long-horizon agentic work to Sol, high-volume drafting to Luna or Terra, and your hardest reasoning to whichever frontier model wins that week. Try Metir AI free and let the right model handle every job.
Sources:
- Previewing GPT-5.6 Sol, a next-generation model | OpenAI
- A preview of GPT-5.6 Sol, Terra and Luna | OpenAI Help Center
- The new GPT-5.6 family: Luna, Terra, Sol | Simon Willison
- GPT-5.6 has landed | Artificial Analysis
- GPT-5.6 Sol Benchmarks, Pricing and Speed | BenchLM
- Cerebras Runs OpenAI GPT-5.6 Sol at 750 Tokens per Second | Value Add Pulse