Published 2026-09-05 · 林启明
Direct answer
GPT-6 Astra lists at $10/$50 per million tokens — until your prompt passes 272K tokens, where the tier doubles to $20/$75. The math, worked examples, and when cache saves you. This guide is written for Australian product and platform teams comparing model quality, spend, routing policy, and production rollout risk.
The two tiers
The catalog lists GPT-6 Astra at $10 input / $50 output per million tokens (USD) with cache hits at $1, for prompts up to 272,000 tokens. Past that boundary the listed tier is $20/$75 with $2 cache hits. The cliff is prompt-size-based: the bigger your context, the higher the per-token rate on the listed tier.
Worked examples
A 100K-token document submitted once: about $1.00 input at the low tier, plus output at $50/M. A 500K-token repository slice: about $10.00 input at the high tier (500K/1M x $20) before a single output token. A 1M-token prompt approaches $20 input per call at the listed tier — before output. At these rates the model's own cache matters: verified repeated prefixes bill at $1/$2 instead of $10/$20.
When the cliff changes your architecture
Between roughly 250K and 300K prompt tokens, trimming context below the boundary can halve the input rate. That is sometimes the wrong trade — the long-context retrieval quality is why you chose Astra — but it should be a decision, not an accident. Watch prompt size as a cost metric, not just a capability metric.
Compare before you commit
GPT-5.6 Sol covers the same 1,050,000-token window at a flat $5/$30. Estimate your real monthly mix with the cost calculator, and check the live pricing page — catalog rates move.
What is unverified
All rates here are the NextModel catalog listing, which mirrors OpenAI's published API prices at launch. Cache-hit billing depends on verified exact-prefix matches; there is no first-party CacheSafety row for this model yet.
FAQ
Does the 272K tier apply to the whole prompt or the overflow?
The catalog lists per-tier rates by prompt size; treat any request above 272K prompt tokens as billing in the $20/$75 tier and budget accordingly. For exact invoice semantics on edge cases, check your request receipts — every call is itemized.
How much is a cached GPT-6 Astra call?
Listed cache hits are $1 per million input tokens in the low tier and $2 in the high tier, versus $10 and $20 uncached. Repeated system prompts and document prefixes are where this pays.