Models
The models available on each Perch plan, and how Roost chooses between them.

Perch runs on a range of models rather than a single one. Roost picks between them for you on every task, and you can pin a specific model by hand: Starter inside the included pool, Pro across every model except the Ultra-only ones, Ultra across the full registry. This page lists what each plan includes. The Pro plan guide and the Ultra plan guide explain the separate Roost and manual-model usage limits.
This list evolves
The rates below are what our inference providers charge us for each model, and they are the rates we meter your usage at, with no markup. They are per million tokens: input, output, and cached input (prompt-cache hits). For a few models the provider's rate differs slightly from the lab's own list price. Cached input applies only to tokens the provider reuses from cache; everything else uses the input rate. OpenAI charges more for GPT-6 and GPT-6.1 prompts above 272K input tokens. Claude Haiku 5.5 has higher rates for prompts above 100K input tokens, noted below. Rates change from time to time; we reflect changes here. Gemini 4 Hydrogen's rate was added November 12, 2026; Gemini 4 Carbon's rate was added November 10, 2026; GPT-6.1 Terra's rate was added October 26, 2026; Kimi K3.1 Pro's rate was added October 23, 2026; DeepSeek V4.1 Flash (1017)'s rate was added October 17, 2026; Beam 1.0's rate was added October 16, 2026; Gemini 4 Argon's rate was added October 13, 2026; Claude rates were updated October 7, 2026; the Qwen launch-model rates were updated October 6, 2026; the other rates were last reviewed September 29, 2026. Your account's Usage view is the source of truth for the PT recorded by each turn.
Starter
Starter has a rotating list of frontier models, subject to availability. Every model available on Starter is free within the included allowance, and its usage still counts. Roost picks from the current routing pool for you; on Desktop and CLI you can select the models below yourself. Starter runs the Standard and Standard 1M tiers and includes up to 20,000 PT of hosted usage each month.
Models you can select
| Model | Availability | Desktop and CLI command |
|---|---|---|
| DeepSeek V4.1 Flash 1017 (Starter) | Current lineup | /model deepseek-flash-starter |
| GLM 5.3 Flash (Starter) | Current lineup | /model glm-5.3-flash-starter |
| Qwen 3.8 Max | Until November 20, 2026, noon Eastern | /model qwen-3.8-max |
Qwen 3.8 Max from Alibaba is on Starter from October 25, 2026, until November 20, 2026, at noon Eastern. It shows in the picker as Qwen 3.8 Max. It reads images and has a 262K-token context window. Its effort control runs off, low, medium and high. When the offer ends, Qwen 3.8 Max leaves the Starter lineup. It is also on Pro.
The selectable lineup can differ from the models Roost routes to automatically. The model picker shows current account availability. Gemini 4 Hydrogen, Gemini 4 Carbon, Gemini 4 Argon, Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, GPT-6.1 Sol and Terra, GPT-6 Sol and Luna, GPT-5.6 Sol, Terra and Luna, and the Pro and Pro 1M tiers remain on Pro.
Pro
Every model in the registry except Claude Fable 5.2, Claude Fable 5.1 and GPT-6 Astra, which are Ultra only, left to Roost or pinned by hand, across all four Roost tiers (Standard, Standard 1M, Pro, Pro 1M). Pro includes the full Starter pool plus the premium models below. Roost has no fixed monthly cap and runs on fair use over five hours and seven days. Manual selection has its own 15,000 PT five-hour limit, 37,500 PT seven-day limit, and 75,000 PT billing-period allowance. See the Pro plan guide for how the two lanes work.
Pro includes every current Starter model and the premium registry below, including Claude Sonnet 5.5 for coding, writing, and project work and Claude Haiku 5.5 for quick tasks and fast iteration.
New on Pro: Gemini 4 Hydrogen
In the model picker, Pro's named models appear in this order: Gemini 4 Hydrogen, Gemini 4 Carbon, Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol, GPT-6.1 Terra, Kimi K3.1 Pro, Hy4, DeepSeek V4.1 Flash (1017), GPT-6 Luna, Claude Sonnet 5.5, Claude Haiku 5.5, Kimi K3, Grok 4.7, Grok 4.6, GLM 5.3, GLM 5.3 Flash, Qwen 3.8 Flash, Qwen 3.8 Max, Qwen 3.8 27B, MiniMax M3, Nemotron Ultra, Nemotron 3.5 Lightning, DeepSeek V4 Pro 0813, DeepSeek V4 Flash 0731, Qwen 3.7 Plus, Inkling, Beam 1.0, Kimi K2.7 Code, GLM 5.2, GPT-6 Sol, GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna.
Gemini 4 Carbon is Google's most capable Gemini 4 model, released November 10, 2026. Pick it in the model picker on Perch AI Web and Desktop; in the terminal, update the CLI and run /model gemini-4-carbon. It accepts images, and its 1M-token context is shared between input and output. Its effort control runs low, medium and high; it always thinks, so off runs as low. It is not one of the automatic tiers, and it is not available on Starter.
GPT-6.1 Terra is OpenAI's newest model, released October 26, 2026. On Perch it is a separate pick from GPT-5.6 Terra, which stays on Pro. Pick it in the model picker on Perch AI Web and Desktop; in the terminal, update the CLI and run /model gpt-6.1-terra. It accepts images and has a 1.05M-token context window. Its effort control runs off, low, medium, high, xhigh and max. It is not one of the automatic tiers, and it is not available on Starter.
Kimi K3.1 Pro is Moonshot AI's newest model, released October 23, 2026. On Perch it is a separate pick from Kimi K3, which stays on Pro. Pick it in the model picker on Perch AI Web and Desktop; in the terminal, update the CLI and run /model kimi-3.1-pro. It accepts images and has a 1M-token context window. Its effort control runs off, low, high and max. It is not one of the automatic tiers, and it is not available on Starter.
DeepSeek V4.1 Flash (1017) is DeepSeek's October 17, 2026 update to V4.1 Flash. The earlier DeepSeek V4.1 Flash pick on Pro is folded into it, so /model deepseek-v4.1-flash now runs this build. Pick it in the model picker on Perch AI Web and Desktop; in the terminal, update the CLI and run /model deepseek-v4.1-flash-1017. It accepts images, and its 1M-token context is shared between input and output. Its effort control runs off, low, high and max. It is not one of the automatic tiers, and it is not available on Starter.
Hy4 is Tencent's model, on Pro and Ultra. It sits right after Kimi K3.1 Pro in the model picker; in the terminal, update the CLI and run /model hy4. It accepts images. Its effort control runs off, low, medium and high. It is not one of the automatic tiers, and it is not available on Starter.
The table below lists published base rates.
| Model | Input / 1M | Output / 1M | Cache / 1M |
|---|---|---|---|
| Gemini 4 Hydrogen | $1.50 | $7.50 | $0.075 |
| Gemini 4 Carbon | $4.00 | $20.00 | $0.20 |
| Gemini 4 Argon | $4.00 | $20.00 | $0.20 |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.20 |
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.01 |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.10 |
| GPT-6.1 Terra | $1.00 | $5.00 | $0.05 |
| GPT-6 Sol | $2.00 | $10.00 | $0.20 |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
| GPT-5.6 Sol | $4.00 | $20.00 | $0.40 |
| GPT-5.6 Terra | $2.00 | $12.00 | $0.20 |
| GPT-5.6 Luna | $0.20 | $1.20 | $0.02 |
| GLM 5.3 | $1.75 | $5.50 | $0.325 |
| Kimi K3 | $3.30 | $16.50 | $0.33 |
| Kimi K3.1 Pro | $2.00 | $10.00 | $0.20 |
| Grok 4.7 | $2.00 | $6.00 | $0.50 |
| Grok 4.6 | $2.20 | $6.60 | $0.55 |
| Qwen 3.8 Max | $2.00 | $6.00 | $0.25 |
| Qwen 3.8 Flash | $0.15 | $0.47 | $0.016 |
| GLM 5.3 Flash | $0.15 | $0.50 | $0.05 |
| GLM 5.2 | $1.40 | $4.40 | $0.26 |
| DeepSeek V4 Flash 0731 (262K context) | $0.14 | $0.28 | $0.007 |
| DeepSeek V4.1 Flash (1017) | $0.30 | $1.20 | $0.006 |
| Kimi K2.7 Code | $0.94 | $4.00 | $0.19 |
| MiniMax M3 | $0.29 | $1.20 | $0.06 |
| Nemotron Ultra | $0.75 | $2.75 | $0.15 |
| Nemotron 3.5 Lightning | $0.10 | $0.25 | $0.05 |
| Qwen 3.7 Plus | $0.40 | $1.60 | $0.08 |
| Qwen 3.8 27B | $0.40 | $3.00 | $0.15 |
| DeepSeek V4 Pro 0813 | $1.74 | $3.48 | $0.14 |
| Inkling | $1.00 | $4.05 | $0.17 |
| Beam 1.0 | $1.00 | $3.00 | $0.10 |
Claude Haiku 5.5 rates above apply to prompts up to 100K input tokens. For prompts over 100K, the rates are $0.50 input, $2.50 output, and $0.05 cached input per million tokens.
Ultra
Ultra includes the full Pro lineup above, left to Roost or pinned by hand, across the same four Roost tiers, plus the models only Ultra can pick: Claude Fable 5.2 and GPT-6 Astra, with Claude Fable 5.1 still available beside them. All three are pinned only: you choose them yourself, and Roost never picks them. Claude Fable 5.2 has a 1M-token context window and GPT-6 Astra a 1.05M-token context window, and both can write up to 128K output tokens. They draw from Ultra's manual allowance like every other pinned model. For Claude Fable 5.2 and Claude Fable 5.1, the model's host keeps prompts and replies for up to 30 days and may have people review them for safety. See the Ultra plan guide for Ultra's limits.
The table below lists published base rates.
| Model | Input / 1M | Output / 1M | Cache / 1M |
|---|---|---|---|
| Claude Fable 5.2 | $8.00 | $40.00 | $0.20 |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 |
| GPT-6 Astra | $10.00 | $50.00 | $1.00 |
Bring your own key
Available on every plan, including the free one. Add a key from any OpenAI-compatible provider, or point Perch at your own endpoint and name any model you want. Perch usage limits do not apply to inference you pay for yourself, so your included allowance stays untouched while you use it. A cloud key works on Web, Desktop, and CLI. A model running on your own computer works in Perch Desktop and the CLI. See Bring your own key and Local models.
At a glance
| Starter | Pro | |
|---|---|---|
| Model selection | Included pool, automatic or pinned | Every Pro model, automatic or pinned |
| Gemini access | No | Gemini 4 Hydrogen, Carbon, and Argon |
| Claude access | No | Opus 5.5, Sonnet 5.5, and Haiku 5.5 |
| GPT-6.1 access | No | Sol and Terra |
| GPT-6 access | No | Sol and Luna |
| GPT-5.6 access | No | Sol, Terra, and Luna |
| Premium-only registry | No | Yes |
| Claude Fable 5.2, Claude Fable 5.1, and GPT-6 Astra | No | No; Ultra only |
| Manual model pin | Included pool only | Any Pro model |
| Roost tiers | Standard, Standard 1M | All four (adds Pro, Pro 1M) |
| Included usage | Up to 20,000 PT / billing period | Separate Roost fair-use windows + 75,000 PT manual / billing period |
| Bring your own key | Yes | Yes |
| Images and scans | On models that see images | On models that see images |
In Perch Desktop, each chat keeps the model and effort level you pick in it. See Model and effort per chat.
Model choice is not only for the model you talk to. On Perch Desktop you can also set which model the helpers Perch spawns run on, which is how you have one model check another's work. See Sub-agents.
For prices and plans, see Pricing and the Pro plan guide. For how the routing works, see Perch Roost.
Images and scans
Models that see images get your screenshots, photos, and scanned pages directly. A model that cannot see images works from text, and tells you when an answer depends on an image it could not read.
MCP servers
Perch connects to outside apps and tools through the Model Context Protocol. Add connectors in Settings, choose whether each one runs on your computer or in the cloud, or write the config file yourself.