Models and effort
Which model answers, how hard it thinks, and what that costs you.
Models
Every model below is included in every plan, the Free plan too; the Free plan limits how many messages a day and how long it thinks, not which model answers. There is nothing to set up and no API key of your own to bring. Pick one in the model pill at pocketphantom.com — that picker is the whole list today, because the Windows app cannot be installed while it is rebuilt around accounts. Auto lets us pick; today that is Claude Opus 5.
Anthropic
| Model | Good for | Speed |
|---|---|---|
| Claude Opus 5 | The hardest work: long reasoning, tricky code, careful writing. This is what Auto picks. | Slowest |
| Claude Fable 5.1 | Anthropic's newest and most capable model. Released September 2026. | Slow |
| Claude Sonnet 5 | Most things. Sharp and fast. | Fast |
| Claude Haiku 4.5 | Quick questions, short edits, lots of small asks. | Quickest |
OpenAI
| Model | Good for | Speed |
|---|---|---|
| GPT-6 Astra | OpenAI's newest and strongest: computer use, browsing, software engineering, long documents (a million tokens of context). Released September 2026, rolling out in phases. | Slow |
| GPT-5.6 Luna | One of three 5.6 models. OpenAI has said nothing about what sets them apart, so we claim nothing either: reasoning, code, long documents. | Slow |
| GPT-5.6 Sol | The second of the three, same family and same generation. | Slow |
| GPT-5.6 Terra | The third. They are peers; take whichever answers you best. | Slow |
| GPT-5.5 | The previous flagship; still strong everywhere. | Slow |
| GPT-5.5 Pro | The 5.5 that thinks longest before it answers, and waits longest for it. | Slowest |
| GPT-5.4 mini | Everyday questions, cheaply. | Fast |
| GPT-5.4 nano | The smallest of them: short answers, small jobs. | Quickest |
GLM, uncensored
GLM is an open model family from Zhipu AI; ours are builds with the refusal training taken out. They answer the questions the big labs decline — security research, fiction that goes somewhere dark, blunt medical or legal questions, a subject somebody decided was awkward. They sit in the picker with the rest, marked with a blue Uncensored tag and included in every plan. What you may do with the answers is still the terms; the model refuses less, the law does not.
| Model | Good for | Attachments |
|---|---|---|
| GLM-5.3 | The newest and strongest of the three: hard reasoning, long documents, a million tokens of context. Carries the green Frontier tag too. | None |
| GLM-5.2 | The previous large model, with the same million tokens of context. | None |
| PocketPhantom Noir | The one that reads pictures: screenshots, photos, scans. 256K context. | Images |
When the Windows app is back it will have a picker of its own, with Claude Sonnet 5 as its default and the Claude and GPT models above in it; the GLM three stay on the web. Older names there (Claude 4.x, earlier GPT versions) are answered by their successors, and the Google Gemini entries left over from earlier versions are in no plan and are being retired.
Effort
Effort is how long the model may think before it answers. The sliders button in the chat bar opens the effort rail; drag it, or use the arrow keys.
| Level | What it means | Use it for |
|---|---|---|
| Low | Quick answers, little deliberation. | Simple questions, lookups, short edits. |
| Medium | Balanced. The default. | Most things. |
| High | Thinks longer before answering. Slower, more thorough. | Planning, debugging, anything with several steps. |
| Max | Thinks as hard as it can. Slowest. | The hard stuff: a proof, a gnarly bug, a long document. |
Higher effort uses more of your monthly AI. The setting is remembered in this browser.
Web search
Always on. When an answer needs something current, the model looks it up (a few searches per answer) and tells you where it found it. There is no switch, on purpose.
Something missing or wrong? Tell us.