The most noticeable change in Codex 5.6 is not just model capability. It is how quickly the usage meter can start reflecting the real amount of work being done. The family has three clear roles: Sol for the hardest tasks, Terra for a balance of capability, speed, and cost, and Luna for the fastest, lowest-cost work. OpenAI uses Sol, Terra, and Luna as durable capability tiers within the GPT-5.6 generation. OpenAI’s GPT-5.6 announcement
One naming detail is worth clearing up. I have heard the phrase “Solar multi-agent” used informally, but the official model name is GPT-5.6 Sol. The multi-agent feature is Ultra mode. It goes beyond a single agent by using subagents to accelerate complex work.
Choosing between Sol, Terra, and Luna
| Model | Positioning | Best fit |
|---|---|---|
| GPT-5.6 Sol | Highest capability | Complex repositories, long debugging chains, architecture decisions, and work that needs repeated verification |
| GPT-5.6 Terra | Balanced capability, speed, and cost | Everyday development, feature work, normal analysis, and iteration |
| GPT-5.6 Luna | Fastest and lowest cost | Quick questions, small edits, and tightly scoped tasks |
My simple rule: start with Terra or Luna for normal work. Move to Sol—and especially Ultra—when the task genuinely needs deep reasoning, cross-file understanding, and several execution loops. Using the strongest model as the default is usually not the most efficient choice.
Headline rate versus actual consumption
The important pricing detail is that, in Codex’s token-based rate card, GPT-5.6 Sol has the same input, cached-input, and output credit rates as GPT-5.5: 125, 12.5, and 750 credits per million tokens. Terra is roughly half of Sol’s rate, while Luna is lower again. Codex rate card
That means Sol is not more expensive than 5.5 on the published per-token rate. It does not mean that a task will cost the same.
Sol Ultra is specifically designed for multi-agent work. When it delegates paths to subagents, checks alternatives, and then brings those results back together, the reasoning, tool calls, and output behind one visible task can all grow. OpenAI does not publish a fixed Ultra multiplier, but it is not sensible to treat a multi-agent workflow as the same unit of work as a single-agent task.
Fast mode compounds the difference. OpenAI notes that Fast mode consumes credits at a higher rate for supported models, while total use still depends on input, cached input, output, and the work itself.
My usage: 20 minutes to a five-hour limit
This morning, I used Sol Ultra with Fast mode for about 20 minutes. That was effectively enough to exhaust my five-hour limit.
This is not a benchmark that can be applied to every account or every task. Repository size, context, subtask count, tool calls, output length, and account-level limits all matter. But it makes one point very clearly: the same headline rate does not mean the same endurance.
With 5.5, it was easy to ask only whether the new model cost more. With 5.6, the better question is whether the task will invoke deeper reasoning, Fast mode, or multi-agent coordination. If it does, the allowance may disappear much faster than expected.
Early external feedback: stronger capability, not a verdict yet
I also looked at the first public discussion around 5.6. The more useful writing does not declare an absolute winner. It separates official positioning, personal tests, and what still needs verification. That distinction matters because independent review is still limited at this stage. Axios’s early coverage
Early community feedback follows two tracks. On the positive side, one developer said Sol Ultra found more issues in a project review and understood the full repository better; another described its workflow as feeling like several 5.5 agents working together in a more structured way. Early project-review report Codex discussion
The other track is usage and speed. Users have also reported Ultra exhausting an allowance quickly, while others did not find it noticeably faster than 5.5—only potentially smarter. These are not reproducible benchmarks, but together with my own result they support a practical test: judge Ultra by whether it removes human judgment and rework, not only by how long a single task runs. Usage and speed discussion
How I will use it
- Use Luna for narrow, obvious tasks.
- Use Terra for most daily development and normal iteration.
- Use Sol for work that needs judgment: difficult debugging, cross-module refactors, and pre-deployment checks.
- Turn on Sol Ultra only when parallel exploration or independent verification is genuinely worth it.
- Reserve Fast mode for time-sensitive work instead of treating it as the default.
Codex 5.6 is most useful when its capability is allocated by task difficulty. Sol Ultra is powerful, but I now think of it as a short, expensive strike team. Use it on the hard problems where it can change the outcome—not on everything, only to watch the usage limit reach zero faster.
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.