At standard rates, GPT-6 Astra and Claude Fable 5.1 both cost $10.00 per million input tokens and $50.00 per million output tokens. Claude Opus 5 sits at half that — $5.00 and $25.00 — and the mid tier runs $2.00 to $4.00 input. Below that, GPT-5.6 Luna at $0.20 input and Claude Haiku 4.5 at $1.00 cover high-volume work.
Headline rates are not the bill, though. Output tokens cost four to six times input across every provider; cached input on GPT-6 Astra is $1.00 against $10.00, a 90% discount on repeated content; batch processing runs at half price; and prompts over 272,000 tokens bill at a multiplier. The number that actually matters is cost per completed task on your own workload.
Published rates per million tokens. OpenAI figures are from OpenAI’s own pricing documentation; Anthropic figures are first-party API rates. Both are standard-tier, short-context pricing.
| Model | Provider | Input | Cached input | Output | Context |
|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | $10.00 | $1.00 | $50.00 | 1.05M |
| Claude Fable 5.1 | Anthropic | $10.00 | — | $50.00 | 1M |
| Claude Opus 5 | Anthropic | $5.00 | — | $25.00 | 1M |
| GPT-5.6 Sol | OpenAI | $4.00 | $0.40 | $20.00 | — |
| Claude Sonnet 5 | Anthropic | $2.00 | — | $10.00 | 1M |
| GPT-5.6 Terra | OpenAI | $2.00 | $0.20 | $12.00 | — |
| Claude Haiku 4.5 | Anthropic | $1.00 | — | $5.00 | 200K |
| GPT-5.6 Luna | OpenAI | $0.20 | $0.02 | $1.20 | — |
Dashes indicate a figure we have not verified from a primary source at the time of writing, not an absence of the feature. Anthropic publishes cache-read pricing separately; check their pricing page for current rates.
Across every provider, output costs roughly four to six times input. GPT-6 Astra is 5x; Claude Opus 5 is 5x; GPT-5.6 Terra is 6x. If your bill is high and you have not looked at response length, that is almost always the cheapest thing to fix — asking for concise output is a one-line change that can move the number more than switching models.
On GPT-6 Astra, cached input is $1.00 against $10.00 uncached — a 90% reduction on the repeated part of the prompt. Any workload with a stable prefix (a long system prompt, a document, a fixed tool list) should be using it.
The constraint is that caching is a prefix match. A single changed byte early in the prompt invalidates everything after it, so a timestamp or a per-request ID near the top of a system prompt silently destroys the discount. Put stable content first, volatile content last, and verify by checking that your cache-read token count is non-zero across repeated calls.
OpenAI prices Batch and Flex at 50% of standard rates, and Fast mode at 2x. Long prompts also cost more per token: on GPT-6 Astra, requests above 272,000 input tokens bill at 2x input and 1.5x output for the entire request, not just the excess. A workload that occasionally spills past that threshold can cost far more than its average token count suggests.
Prices span roughly 50x between the cheapest and most expensive models listed above. That range only makes sense if you match the tier to the job:
The pattern that works for most teams is a cheap model handling the volume, detecting the cases it is not confident about, and escalating only those to an expensive model. Most requests are easy; paying frontier rates for all of them is the most common avoidable cost in production AI systems.
For a capability-oriented view rather than a price one, see our comparison of Claude, GPT, Gemini and the leading Chinese models, our GPT-6 Astra explainer, or our note on Claude Fable 5.1.
Frontier pricing has converged: GPT-6 Astra and Claude Fable 5.1 are identical on headline rates, with Claude Opus 5 undercutting both at half the price.
Choose the tier by what a failure costs you, use caching wherever your prompt has a stable prefix, and judge every option on cost per completed task rather than the rate card.
Not every AI decision needs a spreadsheet. XAIWebsites generates your complete business website — copy, layout, mobile version and hosting — from a short questionnaire, on a simple monthly plan.
OpenAI's published rate is $10.00 per million input tokens and $50.00 per million output tokens at standard short-context pricing. Cached input is $1.00 and cache writes are $12.50. Prompts above 272,000 input tokens bill at 2x input and 1.5x output across the whole request.
Among frontier-tier models, Claude Opus 5 at $5.00 input and $25.00 output per million tokens is half the per-token price of GPT-6 Astra and Claude Fable 5.1, which both sit at $10.00 and $50.00. Below the frontier tier, prices drop sharply — Claude Haiku 4.5 is $1.00 and $5.00, and GPT-5.6 Luna is $0.20 and $1.20.
Output tokens are generated one at a time and cannot be batched or parallelised the way input processing can, so they cost providers substantially more to serve. Across every major provider output runs roughly 4x to 6x the input rate, which is why controlling response length is usually the fastest way to cut a bill.
Substantially, for any workload with a stable prefix. Cached input on GPT-6 Astra is $1.00 versus $10.00 uncached — a 90% reduction on the repeated portion. The catch is that caching is a prefix match: any byte change early in the prompt invalidates everything after it, so put stable content first and volatile content last.
Rarely. The metric that matters is cost per completed task, not cost per token. A cheaper model that needs three attempts and a human review can easily cost more than a pricier model that succeeds first time — and the reverse is just as common on simple work. Measure both on your own tasks.
Often enough that any published comparison should be treated as a snapshot. Promotional rates expire, new models reset the tiers, and providers periodically reprice older models. Check the provider's own pricing page before committing to a budget.
Last updated 19 September 2026. Model pricing changes frequently and promotional rates expire — verify against each provider’s own pricing page before budgeting. See our methodology.