GPT-6 Astra Setup: Codex, DSH and Client Support Checks

To use GPT-6 Astra in a coding agent, match the provider, model ID and API protocol before changing the default model. OpenAI documents Responses as required for Astra tool calling, and Codex requires it for custom providers. DeepSeek Harness exposes an explicit Responses route; Cline and Cursor require separate client-support checks. A model being listed does not prove a complete agent loop works.

Configuration and public catalog checked on 16 September 2026. These are documented configuration examples, not a paid inference or end-to-end agent test.

Choose the endpoint and model ID together

Use a key issued by the same provider. The Ofox catalog currently lists Astra for both /v1/responses and /v1/chat/completions, with a 1,050,000-token context window and a 128,000-token maximum output. Catalog visibility is separate from account access, quota and transport compatibility.

The OpenAI model guide lists low, medium, high, xhigh and max reasoning effort; none is not supported. high is a reasonable trial setting, not a measured best choice for every repository. Compare completed-task cost and accepted changes before increasing effort. The pricing guide includes output, cache writes and cache reads in the calculation.

Codex: use a Responses provider profile

The current Codex configuration reference permits only responses for model_providers.<id>.wire_api. An older wire_api = "chat" example should be replaced.

First check codex --version: OpenAI’s access guidance requires Codex CLI 0.153.0 or newer for Astra. Update the desktop app separately if using that interface.

Merge this into ~/.codex/config.toml, retaining your other profiles:

[model_providers.ofox]
name = "Ofox"
base_url = "https://api.ofox.run/v1"
env_key = "OFOX_API_KEY"
wire_api = "responses"

[profiles.astra]
model = "openai/gpt-6-astra"
model_provider = "ofox"
model_reasoning_effort = "high"

Set OFOX_API_KEY in the environment that launches Codex, using your normal secret-management method. Then select the profile:

codex --profile astra

A successful text response is only the first check. In a disposable checkout, request a small read-only inspection and verify the tool call, returned tool result and final answer all complete. Then try one bounded edit with an existing test. Check streaming completion and usage records as well as the visible answer. A gateway that accepts Responses JSON may still differ in streaming or tool-continuation behavior.

For a direct OpenAI connection, use your existing OpenAI provider and its gpt-6-astra model ID; do not combine an OpenAI key with the Ofox profile. The Codex config guide explains profile selection in more detail.

Cursor: check the supported model picker first

Cursor’s current API-key documentation describes OpenAI custom keys for standard, non-reasoning chat models. It also says custom keys apply to chat while specialized features such as Tab use Cursor’s own models. This does not establish support for an Astra reasoning model through an arbitrary custom base URL.

Check Settings → Models and the model picker in your installed version and account. If Astra is available through Cursor’s own supported integration, follow that integration’s billing and model settings. If the custom-key flow does not offer a supported Astra route, changing a model-name string is not sufficient evidence that it works. Use a documented custom-provider client for that route, or wait for explicit Cursor support; do not assume a chat connection enables every agent feature.

Cline: distinguish a text connection from agent support

Cline’s OpenAI Compatible documentation identifies these connection fields. They describe a compatible connection, not verified Astra agent support:

OpenAI’s current model guide requires Responses for Astra tool calling. A Chat Completions-only provider path does not meet that requirement, even if text works. Verify that your installed Cline integration explicitly supports the required Responses tool workflow, or retain a supported model/use a documented Responses client. Do not assume a gateway translates this workflow without provider documentation and a successful tool-continuation test. Keep context/output limits within the catalog bounds and do not invent capability flags to enable UI options. No Cline release was tested against Astra for this article.

DeepSeek Harness: declare the custom route’s model

The official DSH provider documentation describes custom routes with an API protocol, base URL and nonempty model list. A configuration example in $DSH_HOME/settings.yaml is:

llm-pi-ai:
  providers:
    ofox:
      apiKeyEnv: OFOX_API_KEY
      api: openai-responses
      baseURL: https://api.ofox.run/v1
      models:
        - id: openai/gpt-6-astra

apiKeyEnv references an environment variable, so the key itself does not need to be stored in this file. Here openai-responses selects Responses, matching the official Astra tool-calling requirement. The DSH provider documentation recognizes this protocol; that does not prove every gateway or account supports the complete workflow. A models list replaces that route’s model catalog rather than appending to it. Include the other models you intend to select on this route, or use the documented catalog override mechanism where appropriate.

Restart or reload the configuration as required by your installed DSH version, select the route and model, then perform the same small tool-round check. A local UNKNOWN_MODEL error can indicate a missing route declaration; an upstream access error needs separate investigation.

Verify the migration before making Astra the default

1. Save the previous profile and client version. Check the provider’s current model list, then confirm the effective profile uses the intended URL and model ID.
2. Start a fresh, small task. Check one text response, one tool call and its continuation, then one complete streamed response. Keep the previous model available while checking failures.
3. Compare the same bounded task on each model with the same repository snapshot, permissions, context and acceptance test. Record accepted output, retries, elapsed time and billed usage. Include failed attempts in task cost.
4. Investigate errors by layer: credentials and account access, URL and model ID, request parameters, then tool history and streaming. A successful basic request does not prove the client is the only remaining source of failure.

When copying an older raw API preset, remove unsupported temperature, top_p and top_logprobs; also remove Chat Completions logprobs or the Responses include entry message.output_text.logprobs. The official model guide marks these as unsupported for Astra. Apply this check to the actual outgoing request; client configuration fields may differ.

When migrating a preset that used reasoning.effort = "none", choose a supported Astra value. OpenAI’s migration guide also replaces prompt_cache_retention with prompt_cache_options.ttl for migrations from GPT-5.5 or earlier; check whether your client exposes that field before copying raw API settings into its configuration.

For a budget decision, use the Astra–Sol comparison alongside the benchmark review. Published benchmark savings are a reason to run this trial, not a promise about your own agent’s bill.