Models and providers

Open Customize → Models to choose model access and manage your connections. The model picker shows the models available with your current setup.

First-run setup requires an account; choosing your own provider keeps model requests on that provider's connection. See Pricing before starting paid work. A model's presence in a public catalog does not guarantee that your account can use it.

Use Ace

Select Ace in Customize → Models. Sign in when prompted, check the funding workspace, then confirm purchased or valid promotional credits. Add funds if needed; automatic reload is optional and requires separate consent. Select an available model and return to your conversation.

The Wallet control opens account and funding details. Changing model access to Keys & subscriptions does not cancel automatic reloads; manage them in Wallet. See Ace and your account.

Connect your own API key

Select Keys & subscriptions, choose your provider, and follow its connection form. You can also connect from a terminal:

openscience keys add
openscience keys list
openscience models

Use an API key issued by the provider you choose. A consumer chat subscription and an API account may have separate billing and access.

For terminal sessions, supported provider environment variables are another option:

export ANTHROPIC_API_KEY="your-provider-api-key"
openscience

Other examples include OPENAI_API_KEY, GEMINI_API_KEY, and DEEPSEEK_API_KEY. Set the variable required by your provider. Desktop apps launched from an application launcher may not inherit a terminal's environment; use the connection form for that setup.

Remove a saved connection in Customize or with openscience keys rm. If a key is still present in your environment or shared workspace, remove that source too.

GPT-6.1 Sol

Connect an OpenAI API key, then choose 6.1 Sol in the model picker. OpenScience uses the Responses API for tools and offers low, medium (default), high, xhigh, and max reasoning. Fast mode costs twice the standard token rates and is unavailable with EU data residency. The previous 6 Sol model remains selectable. Availability depends on your OpenAI API account; this does not add a ChatGPT subscription or Ace route.

Use a supported provider sign-in

Where a connection offers browser sign-in, follow that flow instead of entering an API key. For supported ChatGPT access:

openscience keys signin

Complete the browser sign-in and select an available model. Access and limits depend on your provider account. Signing into a provider does not add funds to Ace.

ChatGPT access tokens renew automatically while work continues. Disconnecting or replacing a provider connection still stops work using the previous credentials.

To disconnect that connection:

openscience disconnect codex

Select a model

Use the conversation's model picker. Each row leads with what a request on that model is charged to: Wallet, Your key, Subscription, or Local. When a model is available through more than one connection, the row names the one that choosing it selects: the credential you are already using stays in use, then your configured default, then the Model access mode. A connection the app cannot classify, such as a cloud profile, has no label.

For a terminal run, copy an exact identifier from openscience models --flat:

openscience models --flat
openscience run --model <provider>/<model> "Review the analysis in this project"

Replace <provider>/<model> with an identifier from your list. An unknown or disconnected model produces an error; connect it before retrying.

To set a default in openscience.json:

{
  "model": "provider/model-id"
}

Replace the example value with your configured model. See Configuration for file locations.

Reasoning, context, and rates

Use the effort control beside the model name in the composer to choose the model's supported effort, Fast mode, and context window. Available choices depend on the exact model and access route; there is no universal list. View current prices in Customize → Models → Rates and limits. Ace shows Wallet input/output rates per million tokens in both places, including Fast and long-context tiers. Variable prices say Up to for input and output; their cache prices are estimates. Fixed schedules show exact rates. Provider-key connections show catalog estimates billed by your provider.

Verified effort choices remain usable while Ace refreshes pricing. Fast appears only after its route and prices are verified. If the options report that rates are unavailable, choose Refresh options; this only reloads metadata and does not send a model request, change your selection, or purchase anything.

Open the context indicator for the latest request's token breakdown and the session's recorded model cost. The header and details include the lead's cost and completed direct worker costs, and display small charges below one cent. Recorded Ace turns stay labeled Ace if you later switch to provider keys.

The token breakdown separates ordinary input, cache reads, cache writes, and output when reported by the model service. Cache writes are counted once in the total. A reported request charge takes precedence over an estimate from the token rates. Ace applies the same rounding to each reported request charge as Wallet settlement; provider-key costs retain the provider's reported precision.

GPT-6.1 Sol with ChatGPT

After connecting ChatGPT in Customize → Models, choose 6.1 Sol and its ChatGPT access route. The CLI model ID is openai-codex/gpt-6.1-sol. It offers Low, Medium, High, Extra high, and Max reasoning, plus Fast mode. Usage counts against your ChatGPT/Codex allowance; the API-key route remains available separately.

Access depends on OpenAI's rollout and your plan or workspace settings. Enterprise and Edu administrators must enable the model. See OpenAI's current model availability.

The previous 6 Sol remains selectable through ChatGPT as openai-codex/gpt-6-sol, with the same reasoning levels and Fast mode. It is no longer offered through Ace.

GPT-6 Astra and Claude Fable 5.1

Both models offer Low, Medium, High, Extra high, and Max effort. Their transport contracts remain distinct:

• GPT-6 Astra: native OpenAI access uses Responses for tool calls. The public API supports a 1.05M context window; supported ChatGPT access has separate 272K/872K context choices. Ace uses Azure Responses behind its compatible managed gateway. Ace Fast uses OpenAI's own priority processing alongside Azure Global Standard. Direct OpenRouter connections retain their own supported options.
• Claude Fable 5.1: adaptive thinking is always on, with High as the default; there is no Off or manual thinking-budget option. Native Anthropic access preserves valid thinking blocks and uses the documented handling for blocks invalidated by a changed conversation prefix. Forced tool selection is unsupported. Fable 5.1 has no Fast mode.

Native Fable requests the provider's readable thinking and progress output (thinking.display: "summarized"), which the API otherwise omits by default. This is provider-supplied text, not an additional summary generated by OpenScience, and is not the model's private internal reasoning.

Model availability still depends on your account and the selected route. The exact IDs are openai/gpt-6-astra, openai-codex/gpt-6-astra, and anthropic/claude-fable-5-1 for those native connections. OpenRouter slugs are openai/gpt-6-astra and anthropic/claude-fable-5.1 (prefixed by openrouter/ in CLI model selections).

Fable 5.1 replaces Fable 5 in Ace and runs through Anthropic's API behind the managed gateway. OpenScience offers it when the Ace catalog confirms that route is available, preserving its bound thinking replay across tool turns. Native Anthropic support is available independently. The client still leaves Fable 5.1 out of automatic BYOK OpenRouter catalog imports; explicit custom routes remain user-controlled and their replay compatibility is unverified.

Provider contracts: Astra, OpenRouter Astra, and Fable 5.1.

openscience run --variant <tier> selects a supported reasoning tier for that run. Research effort, selected with --effort normal or --effort ultra, is a separate control for how broadly to investigate the task.

The conversation displays readable reasoning and tool activity when the model supplies them. Some models return only a final answer or a reasoning summary. The visible activity is useful for following progress, but the saved outputs and cited evidence are what you should verify.

Use Show reasoning and activity / Hide reasoning and activity above a turn to inspect its steps. Expansion is saved per turn, not applied to every earlier response. Live turns open once and stay open when they finish; answers and pending approvals remain accessible when steps are collapsed. The control does not stop the model or change the conversation sent to it. There is no global reasoning toggle or Detailed/Compact mode.

Readable passages appear as plain prose, with routine provider phase headings removed from the display. OpenScience does not generate another summary or modify the stored provider response. A model that exposes only a reasoning summary cannot be made to expose private internal reasoning through this setting.

An open response can pause between outputs. OpenScience shows the time since the last output and keeps waiting by default; after two minutes without model output the turn's status says so instead of "Thinking". Stop cancels your wait without undoing completed work. Explicit provider timeout settings still apply, as do provider-side limits. Resubmitting a stopped or failed request can incur a new charge.

For usage across conversations, open Customize → Usage. See Usage reports and CSV exports.

Local and custom endpoints

Use Local models for Ollama, LM Studio, and self-hosted endpoints. For a service not listed in the connection picker, see Custom providers.

Choose a model that supports tool calling for tasks that read files, search, or run code. A chat-only endpoint can answer text requests but cannot perform the full research workflow.

When a connection fails

Check the selected access mode, model ID, and account access. For Ace, also check the funding workspace and available Wallet balance. For your own provider, check key validity, API access, and that provider's usage limits.

If a request times out, the output already received stays in the conversation. Review it before resending the request. Troubleshooting covers stalled requests, missing models, and local endpoint checks.