GPT-6 Astra is OpenAI's newest flagship, and it's callable on the withConflux gateway today — the same cflx_ key, the same OpenAI-compatible endpoint. You just pass gpt-6-astra as the model.
Key takeaways
- Astra is live on the gateway ~6 days after launch — no integration change to add it.
- You pay provider list price plus our flat 25%, from one prepaid balance.
- We meter Astra at the long-context rate so a large request never runs anyone underwater.
- Reach for Astra on hard reasoning; keep high-volume paths on a cheaper model.
What Astra is good at
Astra is a large-context reasoning model (~1M tokens) with vision and tool use. It's the one to reach for on hard, multi-step work — long documents, agentic tool chains, careful reasoning — where a smaller model would drop the thread halfway through.
What it costs through withConflux
One prepaid balance, one flat markup, no per-seat math. Astra is context-tiered upstream — it costs roughly 2× above the short-context threshold — so we deliberately meter it at the long-context rate. That way a big request can never quietly run us, or you, underwater.
| Model | Vendor (per 1M, long-context) | Your price (+25%) |
|---|---|---|
| GPT-6 Astra — input | $20.00 | $25.00 |
| GPT-6 Astra — output | $75.00 | $93.75 |
Every model's live price is on the models page — it reads from the same catalog the gateway bills against, so what you see is what you're charged.
Where we route it
Our standing advice: send the heavy lifting to Astra (or Claude Opus 5), and keep high-volume, latency-sensitive paths on something cheaper — GPT-5.6 Luna or Claude Sonnet 5. You don't have to commit to one model; you pick per request.
Managed vs BYOK
On managed keys you switch models per request with a single model name — billing, routing and spend caps are handled for you. On BYOK the call runs on your own OpenAI account; the gateway still meters, dedupes and caps, but the upstream bill is yours.
When not to reach for it
The most expensive model is rarely the right default. Astra earns its price on genuinely hard work — not on classification, short replies, or high-QPS endpoints.
For those, a mid-tier model is faster and a fraction of the cost. The gateway makes that a one-word change, so there's no reason to overpay by habit.
One key, one bill, every model — including the newest one, less than a week after it shipped. That's the whole idea of the gateway.