Models, reasoning, and usage

Choose a model, reasoning effort, and speed with the controls below the reply box. The available options depend on the connected agent and machine.

Select a model and reasoning effort

Open the model picker and select a model. On desktop, older supported generations appear under More models when available, and Claude threads also accept /model to open the picker or /model <model> to select a known model directly.

The model picker lists available models and provides access to more supported models. Select the image to enlarge it.

If the model offers reasoning controls, choose how much effort it should spend on the task. More reasoning can help with difficult work, such as investigating a bug with several plausible causes, but may take longer or use more of your allowance. Supported Codex models also offer Standard and Fast speed, and some offer Ultrafast as well; Fast and Ultrafast will increase usage.

Use the reasoning picker to choose the effort and speed available for the selected model. Select the image to enlarge it.

Claude offers Fast mode on models that support it, such as Opus 5.5. On desktop, turn on Enable fast mode in the model picker or type /fast in the reply box; on mobile, choose Fast under Speed. Fast mode will increase usage. It needs extra usage turned on in your Claude account and a Claude account connected as a claude.ai login; if your account says Setup token in Configure agents, choose Switch to claude.ai login from its ⋯ menu. If Claude can't use fast mode, the picker will explain why, and Claude will keep working at standard speed.

Claude's model, reasoning, and fast mode choices stay with the thread when you reconnect or reopen it on desktop or mobile, and changing them will not change another existing thread.

Run Claude tasks in parallel with Ultracode

In a Claude thread, open Reasoning on desktop or the model options on mobile and turn on Ultracode. Claude will use workflows to run large tasks in parallel at a much higher usage rate. You can choose the reasoning effort separately, including Low.

The choice stays with that thread when you reopen it or switch devices. It does not become the default for your next thread. New threads follow your Claude settings, where Ultracode is off unless you enabled it there.

Availability depends on the selected model and your Claude settings and account. If Claude cannot use Ultracode, the menu will explain that and the thread will continue normally. Your choice will stay selected so it can take effect if Ultracode becomes available again.

A running workflow shows its progress in the conversation. Choose Stop on its card to stop that workflow, or use the composer's Stop to stop the current response and all running workflows in the thread. Closing the thread will leave the work running.

Choose the model new threads start with

Your new threads in a project start with the model, reasoning, and speed you picked most recently for that agent, including a change you made in an existing thread. To keep new threads on one setup, open the model picker in a Codex or Claude thread or draft and choose Pin for new threads. Your new threads in that project will start with that model, reasoning, and speed until you unpin it, and changing another thread will affect only that thread. Codex and Claude each have their own pin, and teammates keep their own.

When the thread you're in uses the pinned setup, the model picker shows Pinned for new threads; choose it to unpin. When the thread uses a different setup, choosing Pin for new threads will replace the pin with the current thread's setup. To see what's pinned now, hover over that option on desktop; on mobile, the pinned setup is listed under it. To see or clear both pins, open Project settings and go to Models for new threads.

Use Codex with another model provider

For a new Codex thread, Provider above the reply box selects OpenAI or a saved model provider connection. The model picker will show that connection's supported choices. After the thread starts, the provider appears in its header and stays fixed; model and reasoning choices remain adjustable within the same connection. To use another provider, choose Start new thread with… from the header's provider menu. The new thread will start without the old conversation history.

OpenRouter and Vercel offer GLM-5.3 and Kimi K3 in boxes.dev. OpenRouter also offers their Exacto routes. The picker contains these reviewed choices rather than either gateway's entire catalog. Kimi K3 accepts images; GLM-5.3 and advanced custom endpoints use text input. A model without reasoning options has no reasoning control. OpenAI's speed tiers and subscription quotas do not apply to these connections.

Codex keeps its file, command, skill and configured MCP tools. Built-in web search and image generation are not enabled for these connections; an MCP or command can still provide the tools a task needs. Tool use, reasoning quality and latency depend on the chosen model. A gateway can route the same model to different serving hosts, including different quantizations. Exacto does not pin a host. Check the provider's routing settings when that matters for your work.

Understand the context window

The context window is how much of the conversation the model can work with at once. As a Codex conversation fills that window, you may see an auto-compaction warning. To compact a Codex or Claude conversation yourself on desktop, click the context meter below the reply box and choose Compact context now. Compaction reduces the retained context so the conversation can continue; it does not reset your usage allowance. To change when Claude or Codex compacts conversations in a project, use Agent settings.

Check your remaining allowance

Usage allowances come from the connected provider account. On desktop, the agent's rate-limit indicator sits with the model controls below the reply box; open it to check the reported allowance and reset times.

Codex shows session, weekly, and available model-specific quotas, highlighting limits for the selected OpenAI model. If your ChatGPT account has credits, the indicator also shows your balance. Codex keeps working on credits after your session or weekly limit runs out. That limit shows red at 0% and your balance is marked In use, while the indicator below the reply box stays amber. If your account has usage limit resets, the indicator shows how many are available and when they expire.

For OpenRouter, Vercel or a custom endpoint, check that provider's dashboard for available credits, limits and billing totals. boxes.dev shows recorded token and context activity when the provider supplies it. Missing measurements are unavailable, not zero usage.

Claude reports session and weekly quota shared by machines using the same Anthropic organization and account. After you sign in with a different account, an already-running thread may still use the previous credentials and show that account's quota. If extra usage is turned on for your Claude account, Claude keeps working after a session or weekly limit runs out, and Anthropic bills that work as extra usage. The limit still shows red at 0%, and the indicator adds Using extra usage.

Inspect recorded activity

On desktop, choose View usage from the rate-limit indicator to see the current thread's recorded activity: tokens (the units of model input and output), model calls, average and largest input context, and the balance of new input, cached input, and output.

For a Codex model provider thread, choose the View usage chart button beside the context indicator. Its activity belongs to that thread's provider connection; the dialog does not show OpenAI account limits.

Choose the 24-hour view for hourly bars or the seven-day view for hourly or daily bars, and refresh to load newer activity. The chart compares this thread with one combined Other activity series; it does not name or link to other threads.

These measurements come from saved conversations and may differ from the provider's quota calculation. The dialog identifies activity that cannot be matched to the selected account and warns when missing history makes the totals incomplete.

Machine usage is measured separately in boxes.dev box-hours; see Plans, seats, and box-hours.

When Claude changes the model

Claude can switch a Fable conversation to Opus when Fable's safeguards decline part of the conversation and Claude retries it with another model. The thread will show a switch notice, and the picker will show the model now in use. You can select Fable again, though another switch may occur. A refusal without a retry on another model will leave the selected model unchanged.

Claude may first retry a declined response once with the same model. The thread will show that retry as a collapsed Claude Code notice, which you can expand to read Claude's message.