Summary: This review organizes GPT-5.6 Codex models into Sol for hard work, Terra for balanced development, and Luna for quick inexpensive tasks. Sol’s listed token rates match GPT-5.5, but Ultra’s subagents, Fast mode, and deeper reasoning can make task-level consumption much higher. The author reports a 20-minute Sol Ultra Fast-mode session nearly exhausting a five-hour allowance and recommends scaling model choice to task difficulty.
The most striking change in Codex 5.6 goes beyond stronger models. It lies in how fast the usage readout can start reflecting the real amount of work in motion. The lineup now breaks into three clear roles: Sol takes on the hardest jobs, Terra splits the difference between capability, speed, and price, and Luna handles the quickest, least expensive requests. Within the GPT-5.6 generation, OpenAI frames Sol, Terra, and Luna as lasting capability tiers. OpenAI’s GPT-5.6 announcement
A recurring naming slip deserves flagging. Shorthand like “Solar multi-agent” keeps circulating, but the model’s official name is GPT-5.6 Sol. Ultra mode is the multi-agent feature: rather than one agent working alone, it brings subagents in to accelerate complicated jobs.
Choosing between Sol, Terra, and Luna
| Model | Where it sits | What it fits |
|---|---|---|
| GPT-5.6 Sol | Top of the capability range | Large repositories, long debugging chains, architecture decisions, and any job needing repeated rounds of verification |
| GPT-5.6 Terra | A balanced blend of capability, speed, and cost | Everyday development, feature work, routine analysis, and iteration |
| GPT-5.6 Luna | Fastest and most affordable | Quick questions, small edits, and tightly scoped tasks |
The rule I actually follow is straightforward: route ordinary tasks to Terra or Luna before anything else. Sol—and Ultra most of all—comes out only when a job truly demands heavy reasoning, an understanding that spans files, and multiple execution loops. Rarely is the strongest model the efficient default.
Listed rates against real spending
The pricing fact worth holding onto: under Codex’s token-based rate card, GPT-5.6 Sol bills input, cached input, and output at identical credit rates to GPT-5.5 — 125, 12.5, and 750 credits per million tokens. Terra costs about half of Sol, and Luna undercuts even that. Codex rate card
Judged on the published per-token rate alone, Sol costs no more than 5.5. That says nothing about whether an actual task will cost the same.
Sol Ultra exists for multi-agent execution specifically. The moment it passes paths to subagents, weighs competing options, and merges the pieces back together, the reasoning, tool calls, and output hiding behind one visible task can each swell. OpenAI has published no fixed multiplier for Ultra, so booking a multi-agent workflow as the same unit of work as a single-agent task makes no sense.
Fast mode stretches that distance further. Per OpenAI, Fast mode consumes credits at a higher rate on supported models, while overall spend still reduces to input, cached input, output, and the work being done.
This morning: 20 minutes erased a five-hour limit
Earlier today I kept Sol Ultra running with Fast mode engaged for about 20 minutes. That alone was effectively enough to empty my five-hour allowance.
Read it as a single anecdote, not a benchmark that transfers to every account or task. Repository size, context, the number of subtasks, tool calls, output length, and account-level limits all feed into the outcome. It still hammers one point home, though: an identical headline rate promises nothing about matching endurance.
With 5.5, the natural question was whether the newer model simply cost more. With 5.6, ask instead whether the task will pull in deeper reasoning, Fast mode, or multi-agent coordination. When it does, your allowance can evaporate far earlier than expected.
Earliest outside takes: capability reads better, the final call hasn’t landed
I also went through the first public conversation around 5.6. The pieces worth your time decline to crown an outright winner; they hold official positioning, individual experiments, and everything still unverified in separate compartments. That discipline matters while independent evaluation stays this thin at the current stage. Axios’s early coverage
Community reactions so far split along two lines. The positive one: one developer reported that Sol Ultra flagged more problems during a project review and grasped the repository as a whole more completely; another described its workflow as feeling like a group of 5.5 agents working together with more structure. Early project-review report Codex discussion
The other line concerns burn rate and pace. Certain users watched Ultra drain an allowance quickly, while others found it no faster than 5.5 — only potentially smarter. None of these accounts qualify as reproducible benchmarks, yet stacked on my own result they support a workable yardstick: judge Ultra by whether it removes human judgment and rework from the loop, not merely by the time a single task takes. Usage and speed discussion
How I intend to use it
- Luna gets the tight, unambiguous jobs.
- Terra shoulders most day-to-day development and routine iteration.
- Sol takes anything that requires judgment: stubborn debugging, refactors spanning modules, and checks before deployment.
- Turn on Sol Ultra only where parallel exploration or independent verification genuinely earns its keep.
- Reserve Fast mode for time-sensitive work; it never becomes the default.
Codex 5.6 pays off most when model capability is scaled to task difficulty. Sol Ultra is genuinely powerful, yet I now picture it as a short, expensive strike force. Aim it at problems hard enough for it to decide the outcome—not at everything, just to watch the usage counter race to zero sooner.