GPT-6 Astra, OpenAI's newest flagship model, is now selectable in CRHQ through the Codex provider, and it is live on every instance. OpenAI is calling it their most intelligent and most aligned model yet.
What it is
Astra is OpenAI's largest training run to date and their new top-tier model, built for the hardest reasoning, agentic, and long-context work. It runs with a 1 million token context window, up to 128K output tokens, and reasoning effort you can set from low to max. We verified it live on our own subscription before enabling it across the fleet.
How it benchmarks
The figures below are OpenAI's own reported results, not our measurements. The short version: Astra posts step-changes over GPT-5.6 Sol across reasoning, computer use, long context, and reliability.
| Benchmark | What it measures | GPT-5.6 Sol | GPT-6 Astra |
|---|---|---|---|
| OSWorld 2.0 | Computer use | — | 72.6% (about 47% faster per task) |
| FrontierMath Tier 4 | Hardest math | — | 97.6% |
| ARC-AGI-3 | Novel reasoning | — | 99.9% |
| ExploitBench | Cybersecurity | 78.5% | 100% |
| Hallucination rate | Making things up (lower is better) | 12.2% | 4.2% |
On long-context retrieval, OpenAI reports Astra holding 100% accuracy up to 512K tokens and 96.3% between 512K and 1M. In plain terms: it stays sharp deep into very long inputs, and it makes things up far less often than the model before it.
Pricing
Astra is priced at $10 per million input tokens and $50 per million output tokens, with cached input at $1.00. That is the flagship tier, about 2.5 times the cost of GPT-5.6 Sol, so use it deliberately for high-value work.
Alongside the launch, GPT-5.6 Sol got cheaper: its rate dropped to $4 in and $20 out per million tokens. So the everyday Codex workhorse is now more affordable at the same time the new flagship arrives.
| Model | Input, per 1M | Output, per 1M | Cached input |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 | $1.00 |
| GPT-5.6 Sol | $4.00 | $20.00 | $0.40 |
These are equivalent token values based on OpenAI's published rates. Actual billing runs through your own OpenAI or Codex plan.
How to use it
- Start a new session, or edit an agent's default model.
- Choose provider Codex.
- Pick GPT-6 Astra from the dropdown.
- Optionally set an effort level. Astra defaults to our top reasoning tier.
- Send your message as normal. Delegation, artifacts, skills, and tools all work identically.
Your satellite needs an OpenAI or Codex connection, the same one that already powers the GPT-5.6 models. Per-turn cost appears in each session's Costs tab.
Availability
| Fact | Value |
|---|---|
| Availability | All CRHQ satellites |
| Live since | 2026-09-05 |
| Provider | Codex (OpenAI) |
| Client action required | None, beyond an OpenAI or Codex connection |
| Billing | Through your own OpenAI or Codex plan |