Both frontier models now cost $10 in and $50 out per million tokens. Everything interesting is in the structure around that number.
In September 2026, Anthropic released Claude Fable 5.1 and OpenAI released GPT-6 Astra. Both arrived at the same headline price: $10 per million input tokens, $50 per million output.
That convergence is not the story. The story is that each vendor then built a completely different pricing structure around the same number, and those structures encode opposite assumptions about how you are going to use the model.
Pick the wrong one for your workload and the sticker price tells you nothing about your bill.
Anthropic cut the cost of re-reading. Fable 5.1 drops cache reads to $0.25 per million, a 75% reduction. Anthropic puts the resulting saving at around 25% for typical workloads and up to roughly 45% for highly agentic work.
OpenAI raised the cost of going long. Astra prices cached input at $1 per million, and once a request exceeds 272,000 input tokens the entire request reprices: input doubles to $20, output rises to $75, cached input doubles to $2.
Side by side, on the things that differ:
| Fable 5.1 | GPT-6 Astra | |
|---|---|---|
| Input | $10 | $10 |
| Output | $50 | $50 |
| Cache reads | $0.25 | $1 |
| Long-context penalty | None published | Everything reprices above 272K input |
Cache reads are four times cheaper on one, and the other has a cliff the first does not.
The question worth asking is not "which model is better." It is "what shape is my token consumption?"
Long sessions over a stable prefix. An agent that loads a large, unchanging context once and then works against it for hours re-reads that prefix constantly. Cache reads dominate the bill, and a 4x difference on the dominant line item is decisive. This is the shape Anthropic priced for, and it is the shape most coding agents actually have.
Many short, varied requests. Classification, routing, extraction at volume. Little reuse, so caching barely matters and neither structure helps. Here the two are genuinely equivalent on price and you should choose on quality and latency.
Occasional very large single requests. Analysing an enormous document in one pass. This is the shape that runs into Astra's threshold, and the penalty applies to the whole request rather than the overage. If this is your workload, model the cost before committing.
The uncomfortable case is the workload that starts as the second shape and drifts into the first. Agents accumulate context. A tool that began as short request-response becomes a long session as you add memory and retrieval to it, and the pricing structure that suited it at launch may not suit it six months later.
Pricing only matters if the models are close enough that you have a real choice, and on coding they are.
Fable 5.1 scores 55.8% on Terminal-Bench 4.0, up from 42.0% for Fable 5. Astra reports 57.9% on the same benchmark. Those are close enough that benchmark ranking should not drive your decision on its own.
The more interesting Anthropic number is elsewhere. On Terminal-Bench-Science 0.1, which measures agentic scientific research, Fable 5.1 scores 52.6% against Fable 5's 24.7%. That is more than double in one generation, on a benchmark specifically about multi-step autonomous work.
Astra's standout number is on reliability: 4.2% hallucination against 12.2% for its predecessor. If your failure mode is confident wrongness rather than insufficient capability, that reduction may matter more to you than any coding benchmark.
Anthropic also ships Mythos 5.1 alongside Fable 5.1, which scores higher still on Terminal-Bench 4.0 at 60.9%.
Put the two structures next to each other and the strategic position is legible.
Anthropic is betting that agent workloads concentrate in long sessions over stable context, and has priced the dominant cost of that pattern down aggressively. The Terminal-Bench-Science jump points the same direction: multi-day autonomous work is the thing being optimised for.
OpenAI is betting that a very large context window is a premium capability rather than a default operating mode, and has priced it accordingly. The million-token window is there, but the pricing says use the first quarter of it routinely and the rest deliberately.
Neither is wrong. They are different predictions about where the workloads land, and you get to resolve the question for your own case rather than in the abstract.
Want help modelling what your agent workload actually costs under each structure? Get in touch.
This article is a live example of the AI-enabled content workflow we build for clients.
| Stage | Who | What |
|---|---|---|
| Direction | Tom Hundley | Set the angle: a structural comparison, not two launch write-ups |
| Research | Claude Opus 5 | Figures taken from each vendor's own announcement and model documentation |
| Drafting | Claude Opus 5 | Written on the Claude subscription in an interactive session |
| Fact-Check | Human + AI | Every price and benchmark confirmed against a primary source; comparisons made only where both vendors publish the same measure |
| Editorial | Tom Hundley | Final review for accuracy, tone, and value |
A note on what is not compared here. Cache write pricing is published for one of these models and not clearly for the other, so it is left out rather than estimated. The same applies to context-window figures where the two vendors describe the limit differently. We would rather leave a cell empty than fill it with a number we cannot stand behind, and the same rule governs the numbers we put in front of clients.
Want to build this capability for your team? Let's talk about AI enablement →
Discover more content:
