GPT-6 Astra on OpenClaw and Hermes: Setup, Pricing, and Rollout

New on September 29: OpenAI released GPT-6.1 Sol, which it positions as near-Astra performance for complex work at $2 input and $10 output per 1M tokens, a fifth of Astra's price. It also runs on Hermes Agent through a ChatGPT subscription. Compare the two on your own tasks before paying for Astra.

What OpenAI launched

OpenAI positions GPT-6 Astra as its most capable model for complex reasoning, coding, computer use, research, and document creation. The official model page lists a 1,050,000-token context window, up to 128,000 output tokens, and an April 30, 2026 knowledge cutoff. It accepts text and images and produces text; audio and video are not supported model inputs.

OpenAI opened Astra to a limited set of organizations on September 3 and said broader access would follow across ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API, Microsoft Azure, and AWS Bedrock. Two days later the API side is clearly live: OpenRouter added Astra on September 4 and routes it through both OpenAI and Azure endpoints, and the higher ChatGPT paid tiers report access. ChatGPT Plus is the tier most often still waiting, and OpenAI's release notes still stop short of calling Astra generally available. Publication never guaranteed that a given organization can call the model, so check your own account rather than the announcement date.

GPT-6 Astra specifications and price

The headline context limit needs a cost warning. Once an input exceeds 272,000 tokens, OpenAI applies 2× input and cache pricing and 1.5× output pricing to the full request. Batch and Flex processing are 50% of standard rates; Fast processing is 2×. Start with a modest context budget and only pay for the million-token window when the task genuinely needs it.

Use GPT-6 Astra on OpenClaw Launch

The verified managed path at publication time is Hermes Agent. Add your OpenAI key as a user credential, then use Hermes's custom model field. The key is not shared with other customers.

1. Open API Keys and add your OpenAI API key.
2. Open the model picker for your running Hermes Agent instance.
3. In the custom model fields, choose platform openai and enter model ID gpt-6-astra.
4. Apply the change. For Hermes Agent, begin a new session so the new default is used.

Managed OpenClaw does not offer Astra yet. OpenClaw resolves a model against the catalog compiled into its own release, and Astra is absent from every published build so far, including 2026.9.1, so a managed instance needs the same provider-model entry described in the self-hosted section below before the ID resolves. Until that entry ships in the managed image, typing the ID into a managed OpenClaw instance produces a bot that fails on its first turn, so wait until Astra appears in the OpenAI section of the model picker.

If OpenAI returns a model-not-found, permission, or organization-access error, do not keep changing the spelling. Confirm Astra is visible to the same API organization that issued the key. During rollout, waiting for account enablement may be the only fix.

Self-hosted OpenClaw setup

This is the step most guides get wrong, so it is worth being precise. OpenClaw uses provider-prefixed model references, and its built-in OpenAI provider reads OPENAI_API_KEY and calls OpenAI's Responses API. The reference for Astra is openai/gpt-6-astra. On 2026.9.2 and newer that string works on its own, because that build carries Astra in its bundled OpenAI catalog. On earlier releases it does not: the newest OpenAI entry in 2026.7.1 and 2026.9.1 is the GPT-5.6 family, and the registration step below is required there.

Setting the model still appears to succeed. The failure arrives on the first agent turn:

Unknown model: openai/gpt-6-astra. Found agents.defaults.models["openai/gpt-6-astra"],
but no matching models.providers["openai"].models[] entry. Add
{ "id": "gpt-6-astra", "name": "gpt-6-astra" } to models.providers["openai"].models[]
to register this provider model.

OpenClaw is telling you the fix. Register the model against the built-in OpenAI provider, then pin it. The --merge flag matters: models.providers is a protected path, and assigning without it would drop any provider models you already have.

export OPENAI_API_KEY="sk-..."
openclaw onboard --auth-choice openai-api-key

# Register Astra against the built-in OpenAI provider
openclaw config set models.providers.openai.models \
  '[{"id":"gpt-6-astra","name":"GPT-6 Astra"}]' --strict-json --merge

# Then make it the default
openclaw models set openai/gpt-6-astra

The equivalent in openclaw.json is both blocks together:

{
  "models": {
    "providers": {
      "openai": {
        "models": [{ "id": "gpt-6-astra", "name": "GPT-6 Astra" }]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "openai/gpt-6-astra"
      }
    }
  }
}

With the entry in place the request leaves the box and reaches https://api.openai.com/v1/responses, which is the signal worth watching. That difference separates a local naming problem from everything else: Unknown model is raised by OpenClaw and never reaches OpenAI, so no key change affects it. Any response carrying an OpenAI request ID means the reference resolved locally, and you diagnose that response on its merits — authentication, organization access, rate limits, request parameters, or a transient server error. Drop the entry once an upstream release ships Astra in its own catalog.

Self-hosted Hermes Agent setup

Hermes Agent uses a provider name and a bare model name in separate fields. Put the key in ~/.hermes/.env:

OPENAI_API_KEY=sk-...

Then set the model in ~/.hermes/config.yaml:

model:
  provider: "openai-api"
  default: "gpt-6-astra"

You can also run hermes model to configure the direct OpenAI provider. The old LLM_MODEL environment variable is no longer the source of truth. Existing chats keep their current model, so start a new Hermes session after the change.

Model IDs: do not mix the two formats

Agent features and migration caveats

Astra supports streaming, function calling, and structured outputs. OpenAI's model guidance says agent tools are available through the Responses API, including web and file search, code interpreter, hosted shell, computer use, MCP, skills, and tool search. A Chat Completions request may generate text, but tool use requires Responses.

• There is no none reasoning mode. Migrate none or minimal workloads to low.
• Remove sampling parameters when reasoning is enabled. OpenAI tells migrating clients to remove temperature, top_p, and top_logprobs; Chat Completions clients should also remove logprobs.
• Review skills and agent instructions. Astra follows repository and skill instructions closely and can operate powerful tools. Keep permissions narrow, use approval gates for consequential actions, and test workflows before granting unattended access.
• Expect mid-run interaction. Astra adds asynchronous tool calling, mid-turn steering, configuration updates, and monitoring. Older wrappers may need an update before they expose all of those capabilities.

Audit your AGENTS.md and skills before switching

Astra reads instruction files more literally than earlier models, which turns a tolerant prompt into a strict one overnight. Rules your old model quietly ignored now get executed, vague rules make it stop and check with you, and two rules that contradict each other can stall it outright. The practical advice from OpenAI's own migration guidance is subtractive: go through AGENTS.md, SOUL.md, and your skill files and delete what accumulated for weaker models instead of adding more scaffolding.

OpenAI suggests asking the model to name the skill and the instruction that caused it to pause or change direction. On a bot with many loaded skills that is the quickest way to find the silent or conflicting line, and it works just as well against an OpenClaw workspace or a Hermes skill directory.

• It asks more questions. Astra stops mid-task to confirm choices. If your prompt already implies authorization, say so explicitly — tell it to treat “can you…” and “help me…” as instructions to do the work, and to finish reversible work before asking for approval.
• It formats heavily. Astra reaches for lists, tables, and Markdown by default. Chat channels such as Telegram and WhatsApp read better with prose, so ask for paragraphs and reserve lists for genuinely parallel items.
• Word choice changes its behavior. Early users report that asking for “excellent” work keeps it moving while asking for “perfect” work sends it into diminishing returns. Set a bar it can actually reach.
• It delegates to subagents. Astra is trained to split work across parallel subagents. If your harness supports that, tune how much it delegates rather than leaving it to default.
• Reasoning effort can change mid-conversation. A configuration_update input item raises or lowers effort without rewriting the prompt prefix, so the cached prefix survives the change.

Troubleshooting the launch window

• OpenClaw says Unknown model: openai/gpt-6-astra: this is the catalog gap, not your key. Add the models.providers.openai.models entry shown above. The error never reaches OpenAI, so no amount of key rotation changes it.
• Model not found or access denied from OpenAI: verify the API key's organization. The spelling is gpt-6-astra. Once OpenClaw stops raising Unknown model, the reference resolved locally, so you are reading OpenAI's response rather than debugging the model name.
• Model absent from a framework picker: update the framework, inspect its provider model list, or use the custom model field when available.
• Tools do not run: make sure the integration uses the Responses API. Tool support is not available through Astra's Chat Completions path.
• OpenRouter route fails: the slug is openai/gpt-6-astra. If you pinned openai/gpt-6-astra-pro, note that OpenAI's API reference publishes only the single gpt-6-astra snapshot, so fall back to it.
• Unexpectedly high bill: inspect input size. Crossing 272,000 tokens reprices the whole request at the long-context multipliers.

Primary sources

• OpenAI GPT-6 Astra model reference
• OpenAI GPT-6 Astra usage and migration guide
• OpenAI GPT-6 Astra safety overview
• OpenClaw model provider documentation
• Hermes Agent provider documentation

GPT-6 Astra FAQ

Related model guides

• GPT-6.1 Sol: near-Astra for less
• OpenAI models on OpenClaw
• Bring your own API key to OpenClaw
• Bring your own API key to Hermes Agent
• GPT-5.6 on OpenClaw
• Compare agent models