GPT-6 Astra in Codex, Cursor, Cline and DSH: Working Configs (2026)

GPT-6 Astra speaks the same OpenAI-compatible protocol every coding agent already uses, so “adding” it is a model-string change rather than an integration. What actually needs a decision is the effort level and the cache-read rate. This page has a working config for Codex CLI, Cursor, Cline and DeepSeek Harness, and the two settings worth getting right before you make it a default.

Model ID:      openai/gpt-6-astra
Aliases:       gpt-6-astra, gpt-6-astra-2026-09-03
Endpoints:     /v1/chat/completions, /v1/responses
Base URL:      https://api.ofox.io/v1
Price:         $10.00 in / $50.00 out per 1M
Cache:         $1.00 read / $12.50 write
Context:       1,050,000 in / 128K out
Effort levels: low, medium, high, xhigh, max

Read from the live /v1/models endpoint on 5 September 2026, the day Astra was listed.

Set These Two Things First

Before the config blocks, the two settings that decide whether this is a good idea for your workload.

Effort: start on high, not max. Astra exposes five levels. Artificial Analysis measured max at 61 on its Intelligence Index and high at 60 — one point — while max wrote 42M output tokens against high’s 16M and cost $3,013 against $1,429 to run the same index. Reasoning tokens bill at the output rate whether you read them or not, so the effort setting is a price multiplier with a quality hypothesis attached. Verify the point exists on your tasks before paying for it.

Cache reads cost 4x what Claude Fable 5.1 charges. Both models are $10 / $50 per million on the headline. Astra reads cached input at $1.00 per million; Fable 5.1 reads it at $0.25. Coding agents resend the same system prompt, tool definitions and file context every turn, so cached input is usually the largest line on a long session. If your agent loop replays a long prefix, that row moves your bill more than the headline rate does.

The honest summary from the independent benchmarks: Astra is built and priced for agents. On the Coding Agent Index it scores 67 against GPT-5.6 Sol’s 65 while using about a third of the tokens in the Codex harness, so a full task costs roughly what Sol costs. On short prompts you are paying 2.5x per token for a general score that did not move.

Codex CLI

Two environment variables and a model flag.

export OPENAI_API_KEY=<your ofox key>
export OPENAI_BASE_URL=https://api.ofox.io/v1

codex --model openai/gpt-6-astra "Migrate this module off the deprecated client"

That is the whole setup for a single endpoint. The full Codex configuration guide covers installation and verification.

When the env-var route stops working

The two-variable trick hard-codes one endpoint. The moment you want OpenAI direct and a gateway live in the same terminal, or different projects on different models, you need ~/.codex/config.toml:

[model_providers.ofox]
name = "ofox.ai gateway"
base_url = "https://api.ofox.io/v1"
env_key = "OFOX_API_KEY"
wire_api = "chat"
request_max_retries = 4

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

codex --profile astra "Refactor the auth layer"

Two keys people get wrong here:

• wire_api = "chat", not "responses". Both values are accepted, and OpenAI’s own first-party config defaults to "responses", but third-party OpenAI-compatible gateways expose the Chat Completions schema at /v1/chat/completions. Set "chat" for a gateway route.
• env_key names the variable, it does not hold the key. Codex reads that environment variable at runtime. Do not paste a token into the TOML.

The multi-provider deep dive covers the rest of the block, including http_headers for gateways that need organization scoping.

Cursor

Settings → Models → Override OpenAI Base URL.

1. Enter https://api.ofox.io/v1 in the Override OpenAI Base URL field. Include the /v1 — Cursor appends /chat/completions to whatever you provide.
2. Put your gateway key in the OpenAI API Key field. The label says OpenAI; the key is sent to the endpoint you configured above.
3. Click + Add Model and type openai/gpt-6-astra exactly. The string has to match what the endpoint expects in the request body.

Cursor’s override is global rather than per-model, which is the known limitation of this route: turning it on points every OpenAI-labelled model at your endpoint. The Cursor and Cline setup guide covers what that breaks and the workarounds.

Cline

OpenAI Compatible provider.

In the Cline settings panel, choose OpenAI Compatible as the API Provider and fill three fields:

The model ID needs the provider prefix. gpt-6-astra on its own is an alias the gateway resolves, but Cline’s “model not found” reports almost always trace to a string entered without the openai/ prefix against a route that wanted it. The Cline configuration guide has the Anthropic-provider path as well, which matters for Claude models but not for this one.

DeepSeek Harness (dsh)

Settings → Models → Add a custom provider. The form asks for five things:

Or write it directly into $DSH_HOME/settings.yaml:

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

Three dsh-specific traps, all verified on a clean install:

• The models list replaces the route’s catalog, it does not extend it. A model you have not declared fails with UNKNOWN_MODEL before any request leaves the machine. There is no “send it anyway” path on a custom provider.
• The provider ID is permanent. Sessions, defaults and credential references all key off it, so renaming means creating a new provider and deleting the old one.
• Defining a provider does not make it the default. The headless profile boots the built-in DeepSeek route and asks for DEEPSEEK_API_KEY even when another provider is fully configured. Set the default model in the web app first.

dsh is still a developer preview and says so in capitals in its own README. The dsh setup guide and the version and stability notes cover what that means before you point it at a repository you care about.

Verifying the Route

Same check regardless of which agent you configured — if this returns a completion, the transport is fine and anything still broken is agent-side config:

curl -X POST https://api.ofox.io/v1/chat/completions \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-6-astra",
    "messages": [{"role": "user", "content": "Reply with exactly: ok"}],
    "reasoning": {"effort": "high"}
  }'

Read usage on your own prompts before forecasting cost. At $50 per million output tokens with reasoning billed at that rate, the token count matters more than the rate card.

When model_not_found Comes Back

Almost always a stale or malformed model string. The valid identifiers are openai/gpt-6-astra, plus gpt-6-astra and gpt-6-astra-2026-09-03 as aliases.

Two specific traps:

• Older Codex guides recommend openai/gpt-5.4-mini-codex. That model is no longer in the catalog and returns the same model_not_found. The current Codex-family options are openai/gpt-5.1-codex-mini at $0.25 / $2.00 and openai/gpt-5.3-codex at $1.75 / $14.00.
• There is no tier suffix on GPT-6. GPT-5.6 shipped as Sol, Terra and Luna; GPT-6 ships as Astra and Astra Pro. Any router or config template that builds a model string by appending a tier will not find a match. The generational comparison covers what else that naming change breaks.

GET https://api.ofox.io/v1/models is the authority on what is callable, rather than any page including this one.

Which Model to Actually Point These At

Astra is not automatically the right default just because it is newest. At the same $10 / $50, three other routes are worth benchmarking against it on your own tasks:

Fable 5.1 leads the Coding Agent Index at 70 and reads cache at a quarter of Astra’s rate, which is the combination that matters most on long agent sessions. Astra’s counter-argument is token efficiency: about a third of Sol’s tokens in the Codex harness, which is why its per-task cost lands near Sol’s despite the higher sticker. The head-to-head has the full split.

Since every agent above reaches all four through the same endpoint, switching between them is one string. Run your own three tasks on two models and read usage — that comparison costs a few cents and beats any benchmark table, including the ones on this site.

Sources

• https://ofox.io/models/openai/gpt-6-astra
• https://developers.openai.com/codex/config-reference
• https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra
• https://ofox.io/docs/integrations/cline

Model IDs, rates, context lengths and endpoints were read from the live Ofox /v1/models endpoint on 5 September 2026. Coding Agent Index scores and the effort-level cost comparison are from Artificial Analysis, read 4 September 2026. Codex config.toml keys follow OpenAI’s config reference. The dsh provider form fields and the three dsh traps were verified on a clean install for our earlier dsh setup guide and re-checked against it here; dsh remains a developer preview with breaking changes promised by its maintainers.