Models
A chat session on Subscriptions or workspace Credentials runs on one model and reasoning effort, such as Opus 5.5 (xHigh). Personal tasks without a Subscription use Auto - Sensei free, whose model stays hidden. Pick it in the Model row when you create a session, or from the model picker in the chat header. A session you do not set starts on its Assignee's Default Model.
Pick a model
- Open the model picker. In a session it sits on the right of the chat header. In the New Session dialog and in a Backlog session's details it is the Model row.
- Hover over a model to open its Reasoning effort menu.
- Pick an effort.
Some OpenAI models also have a Speed group under the efforts. See Choose a speed.
Workspace pickers offer models run by workspace Credentials only. Personal pickers show Auto when available and provider groups while you hold a Subscription. With none, only Auto - Free, with the prompts left, and Connect a subscription appear. Free reads Auto - Sensei free once selected. A login at its limit or needing reconnect still counts, including SuperGrok. Limits, reconnects and provider failures never move work to Free. Connecting the first Subscription and removing the last change personal choices from their next turn, as AI providers explains.
Reasoning effort sets how long the agent thinks before it acts. A higher effort suits hard problems. A lower one is faster for mechanical edits.
Muse models run in Muse Code. Muse Wonder Full requires Power Usage or Maximum Usage. Auto and the Model Router route Muse when the allowed source can run it. Delegated Runs support Meta. When Auto or a Delegated Run selects Muse Wonder Full and the login’s plan lacks it or cannot be read, it uses Meta’s next model below it in the Model table.
Gemini models run in Antigravity on your sign-in for personal work on Subscriptions, or the workspace's selected Google sign-in or Gemini API key. There is no personal method selection. A limit or reconnect never crosses the credential boundary. Auto, the Model Router and Delegated Runs support Gemini. Auto steps down from Argon to Flash on a workspace Gemini API key. See Gemini.
Let Auto pick the model
Auto over personal Subscriptions or workspace Credentials uses the Model Router described below. Auto - Sensei free is available only without a Subscription and keeps its separate Free routing. Sensei chooses one Source per Free chat session, Sensei thread Turn, Deploy or Cron run. Delegated Runs use their parent Turn's Source. Free works like a personal Subscription, within its allowance and staff maximum Turn Level. Sensei moves work automatically when a Source reaches its limit, is paused or is removed, without showing the model, Source, provider or move. On Auto over Subscriptions or workspace Credentials, Sensei picks the model, and with it the provider, for each turn. The picker then reads, for example, "Auto - Sonnet 5.5 (Medium)". Pick a model to turn Auto off.
For each message, Sensei sets a Turn Level: Light, Standard, Hard or Hardest. Its classifier, Jev, reads your message and the session so far. Jev judges code and any other work you send the agent, such as writing, research, marketing, finance or legal work. A fixed rule turns Jev's answers into the level:
- The level starts at how hard the work most likely is.
- It moves one level up when Jev leans toward harder work, or one level down when Jev leans toward easier work.
- A change to security, stored data, money, work that runs in parallel or legal text adds a level when Jev gives it at least a 35% chance.
- When you say the agent's last work was wrong, the turn gets at least one level more than the turn before.
Each provider has one model for a level. An empty cell means the provider has no model strong enough for that level:
| Level | Claude | Codex | GLM | Kimi | DeepSeek | Gemini | Muse | Grok |
|---|---|---|---|---|---|---|---|---|
| Light | Sonnet 5.5 (Medium) | GPT 6.1 Terra (Medium) | GLM-5.3-Flash (Max) | Kimi K3 (Max) | DeepSeek V4.1 Flash (Low) | Gemini 3.8 Flash (High) | Muse Wonder (Minimal) | Grok 4.7 (Low) |
| Standard | Opus 5.5 (Medium) | GPT 6.1 Sol (Medium) | GLM-5.3 (Max) | Kimi K3.1 Pro (Low) | Gemini 4 Argon (Low) | Muse Wonder (xHigh) | Grok 4.7 (xHigh) | |
| Hard | Opus 5.5 (High) | GPT 6.1 Sol (xHigh) | Kimi K3.1 Pro (High) | Gemini 4 Argon (Medium) | Muse Wonder (Max) | |||
| Hardest | Opus 5.5 (xHigh) | Kimi K3.1 Pro (Max) | Gemini 4 Argon (High) | Muse Wonder Full (Max) |
Light turns get the fastest model. Standard and Hard turns get the model that costs least at that level. Hardest turns get the strongest model.
A model's strength is the average of its scores on DeepSWE, SWE Atlas QnA and Terminal-Bench, three coding benchmarks from Artificial Analysis. Each score counts relative to the score of Opus 5.5 (Max) on the same benchmark.
Some models run only inside their provider's coding agent and have no public API, so Artificial Analysis cannot time or price them. Auto picks such a model only for Hardest turns.
Some models are only on the higher plans of a subscription. When your login's plan does not include the model for the level, the turn runs on the provider's next model below it.
Auto picks the provider when a new conversation starts: on the first turn, after /clear, and when the context is 80% full. It takes the provider with the most credits left. A subscription comes before an API key, which costs money on every turn. Later turns stay on that provider. They move to another provider only when:
- every login of the provider is at its limit or needs reconnecting,
- the provider's subscription is almost at its limit and another subscription has credits to spare, or
- the provider has no model for the level, or none that reads your attachments.
A move starts a new conversation with the conversation so far handed over. The transcript names the reason, for example "Anthropic credits low: switched from Opus 5.5 (High) to GPT 6.1 Sol (Medium), conversation handed over".
When Jev fails or does not answer in time, or your message has only attachments, a later turn keeps its model, and a first turn starts on the provider's default model.
Auto routes every released provider the work's source can run, including Gemini, Muse and Grok. Anthropic on AWS Bedrock does not count because its key may not run every model in the table. You can make Auto your Default Model.
Grok runs chat, Auto, Sensei, Crons and Deploys. Delegated Runs can use Grok with full access only; read-only runs use another provider. See Grok for credentials and Deploy limits.
Available models
| Provider | Model | Reasoning efforts | Context window (tokens) |
|---|---|---|---|
| Anthropic | Fable 5.1 | Low, Medium, High, xHigh, Max, Ultracode | 1,000,000 |
| Anthropic | Opus 5.5 | Low, Medium, High, xHigh, Max, Ultracode | 1,000,000 |
| Anthropic | Sonnet 5.5 | Low, Medium, High, xHigh, Max | 1,000,000 |
| Anthropic | Haiku 5.5 | Low, Medium, High, xHigh, Max | 1,000,000 |
| OpenAI | GPT 6 Astra | Low, Medium, High, xHigh, Max | 872,000 |
| OpenAI | GPT 6.1 Sol | Low, Medium, High, xHigh, Ultra | 872,000 |
| OpenAI | GPT 6.1 Terra | Low, Medium, High, xHigh, Ultra | 872,000 |
| OpenAI | GPT 6 Luna | Low, Medium, High, xHigh | 872,000 |
| Z.ai | GLM-5.3 | Low, High, Max | 1,000,000 |
| Z.ai | GLM-5.3-Flash | Low, High, Max | 1,000,000 |
| Moonshot AI | Kimi K3.1 Pro | Low, High, Max | 200,000 |
| Moonshot AI | Kimi K3.1 Pro Highspeed | Low, High, Max | 200,000 |
| Moonshot AI | Kimi K3 | Low, High, Max | 200,000 |
| Grok | Grok 4.7 | Low, Medium, High, xHigh | 500,000 |
| DeepSeek | DeepSeek V4.1 Flash | Low, High, Max | 1,000,000 |
Anthropic, GLM, Kimi and DeepSeek models run in Claude Code. OpenAI models run in Codex. Grok 4.7 runs in Grok Build and takes no image attachments. That choice decides a few things in the session:
- Skills start with
/in Claude, GLM, Kimi and DeepSeek sessions, and with$in Codex sessions. - A
/compactfocus is Claude Code's prompt for the summary. Codex compacts without a prompt, so your focus stays in the conversation's history as a note, which Codex may not follow.
GLM-5.3 reads text only. For images and PDFs, use GLM-5.3-Flash or a Kimi model. OpenAI models and DeepSeek V4.1 Flash read images but not PDFs, so the composer does not take a PDF attachment on them. A DeepSeek agent reads a PDF in the repository by extracting its text with a shell tool. See Messages, Up next and steering for attachment limits.
Kimi K3.1 Pro Highspeed needs a Kimi Code Pro membership or higher. On a lower membership, its turns fail with a message that says so. Pick Kimi K3.1 Pro instead.
Retired models sit under Legacy at the bottom of the picker.
Chat sessions do not offer Ultracode or Ultra. Both efforts make the agent run agents of its own, and a chat session starts every helper as a Delegated Run instead. See Where each provider runs.
Choose a speed
Some OpenAI models can answer faster at a higher price. Their menu has a Speed group under the efforts. Standard is the default and costs nothing extra.
- Open the model picker and hover over the model.
- Under Speed, pick Fast or Ultrafast. The menu stays open.
- Pick an effort. When the session is already on that model, the speed applies at once and keeps its effort.
The model then reads, for example, GPT 6 Astra (High, Fast). Changing the effort keeps the speed. Picking another model goes back to Standard. Every model picker has the Speed group, so you can save a speed on a session, a Cron, your Default Model or the deploy agent.
Ultra, a reasoning effort on GPT 6.1 Sol and GPT 6.1 Terra, is not a speed.
Whether a turn gets its speed depends on the OpenAI login it runs on, as AI providers lists. A turn that cannot get it stops, and the chat says why, for example "Fast isn't available to this OpenAI account". It never runs at Standard instead. Pick Standard to run it.
Set your Default Model
Your personal Default Model applies to new Sensei threads, personal tasks and personal Crons without their own choice. Existing work with a saved choice keeps it. A Deploy without a stored model takes the first connected provider in catalog order and its catalog default model.
- Open Settings > You > AI Providers.
- Under Default Model, choose Auto or a personal model and effort while you hold a Subscription, or Auto - Free when you hold none.
The Free selection reads Auto - Sensei free. With no Subscription and no saved default, it is selected already. Connecting the first Subscription moves a Free default to Auto over Subscriptions. Removing the last moves it to Free. Limits, reconnects and provider failures never make that move.
A workspace task uses your preferred model only when workspace Credentials run its provider, and never uses your Free choice. Otherwise workspace defaults and available workspace models apply. See AI providers.
If your Default Model's provider is not available in a workspace, settings say so, and new tasks there use an available model.
Set the Workspace Default Model
Owners and admins set the model a workspace task starts on when its Assignee has no usable Default Model of their own. A session booked to Sensei uses it too.
- Open Settings > your workspace > AI Providers.
- Under Default Model, choose Auto or a model and effort.
Auto shows only while the workspace Credentials can run it. With Auto selected, a new task routes each turn over the workspace Credentials, as a task on Auto does. Crons on Default never use Auto. They run on the last model picked there, or on an available model when none was picked. When the workspace Credentials can no longer run Auto, the card says so, and new tasks start on that model instead.
A new task starts on its Assignee's Default Model, whoever creates it. If you create a Backlog session without picking a model, the Model row shows "Default Model". The session's details then show its Assignee's Default Model, which it takes when it starts.
Switch model during a session
You can change the model of an Active session while the agent is idle. While a turn runs, the picker is locked.
- Open the model picker in the chat header.
- Pick the new model and effort.
- Send your next message. The new model takes effect with it.
For personal or workspace models, switching within the same provider keeps the conversation as it is. Switching to another provider starts a fresh conversation for the new agent. Sensei hands it the conversation so far as text. The text carries what you and the agent said, and the branch and sandbox still hold the work. The transcript marks the switch, for example "Switched from Anthropic - Opus 5.5 (xHigh) to OpenAI - GPT 6.1 Sol (xHigh), conversation handed over".
A very long conversation is shortened for the new agent. It keeps your opening prompt and the latest exchanges. After /clear, nothing said before the clear is handed over.
Where each provider runs
| What | Providers | Where the model is picked |
|---|---|---|
| Chat sessions | Every released provider available to the task's allowed source | The chat header, or the New Session dialog |
| Deploy agent | Every released provider on the work's allowed source; personal Deploys can use eligible Sensei free sources | Project settings or the Deploy Chat header |
| Sensei | Every released provider on your Subscriptions; Sensei free uses eligible Sources | Sensei's composer |
AI providers explains whose credentials run each one. Chat sessions and Sensei do not offer Ultracode or Ultra. A session or a Default Model that uses one runs at the highest effort below it that the model offers, such as Opus 5.5 (Max). The picker shows that effort for Subscription and workspace models. Crons and the deploy agent keep Ultracode and Ultra.
Reassigning a workspace task keeps its workspace Credentials. See Chat sessions and tasks. The API and the CLI name models by id. See the models reference.