Claude Fable 5.1 arrived on 1 September and GPT-6 Astra on 4 September, both at exactly $10 per million input tokens and $50 per million output. Two labs, three days apart, landing on the same number for the top of their range. That is not coincidence, and it tells you something about how this market now prices.
Identical pricing, different shapes
Fable 5.1 carries a 1,000,000-token context window and a 128,000-token maximum output, accepting text, images and files. GPT-6 Astra carries 1,050,000 and 128,000, with the same input modalities. Structurally they are close to interchangeable.
Astra is the first model in OpenAI's GPT-6 generation, positioned as their most capable and most aligned. Fable 5.1 succeeds Fable 5 in Anthropic's top line, above Opus.
Both sit at roughly double their own vendor's mid-tier. GPT-5.6 Sol launched at $5 and $30, so Astra is twice that on input. Claude Opus 5 is $5 and $25, so Fable 5.1 is twice on both sides.
The per-message cost is worse than the rate card suggests
Comparing per-token rates understates the gap between these and the tier below, because top-end models tend to produce longer answers. They reason more, they qualify more, and they write more.
A 3,000-token prompt with a 1,500-token answer costs about $0.105 on either of these, against $0.053 on Claude Opus 5. That is twice the price. If the flagship writes twice as much, which in our experience it often does, the real ratio approaches three and a half.
This is why we price usage on measured tokens rather than a flat per-message figure. A flat price has to assume the long answer, which means everyone asking short questions subsidises the few asking for essays.
When the top tier is worth it
Our honest position, having run a lot of work through both tiers: for most tasks, the mid-tier model is indistinguishable in output quality and costs around half as much per message, and far less once its shorter answers are counted.
The flagship earns its price on a narrow set: long-horizon agentic work where errors compound across many steps, genuinely hard reasoning where the mid-tier visibly gives up or confabulates, and tasks where a subtle mistake is expensive enough that a small reduction in error rate pays for a large increase in cost.
What it does not earn its price on is ordinary work made to feel important. Drafting, summarising, routine analysis and most coding sit comfortably in the tier below, and the difference does not show up in the output.
The way to find your own line is to stop guessing at it. Run the same prompt through the flagship and the tier below, read both answers, and ask whether you could tell which was which without the labels. On Deepest that is one message rather than two subscriptions and an afternoon.
A note on "most aligned"
OpenAI describes Astra as its most intelligent and aligned model. Alignment claims are worth treating carefully: they are made by the vendor, measured on evaluations the vendor selected, and "aligned" means something specific and internal rather than a general guarantee of good behaviour.
That is not cynicism about the work, which is real and hard. It is a caution about reading a marketing sentence as a safety property you can rely on in your own deployment. Your obligations to test the model on your own use case do not transfer to the vendor because their announcement used a reassuring word.
Frequently Asked Questions
Is GPT-6 Astra better than Claude Fable 5.1?
They are priced identically and shaped almost identically, so the answer depends entirely on your work rather than on a general ranking. Run both on your own prompts. Anyone who tells you one is definitively better without asking what you do is selling something.
Should I default to a $10/$50 model?
Almost certainly not. The tier below costs around half as much per message and handles the large majority of real work indistinguishably. Reach for the flagship when you have a specific class of problem where you have confirmed the cheaper model falls short, rather than as a general policy.
Why do flagship models cost more per message than the rate suggests?
Because they write more. Reasoning tokens bill as output, and top-tier models reason at length before answering. The per-token rate is the floor of the difference, not the whole of it, which is why comparing rate cards understates the real gap between tiers.
