I wrote about GPT-6 Astra's release notes last week. The obvious follow-up is the comparison post — Astra against Claude, Gemini and DeepSeek — so I went and built the table from each provider's own pricing page rather than from the aggregator sites.
Two things came out of that. The table is below, because it's the thing people came for. But building it convinced me the headline rate is close to the least useful number in the comparison, and I want to spend most of this post on why.
Same caveat as last time: I haven't shipped production traffic on Astra. This is a read of published pricing, not a review.
The table
Per million tokens, input / output, as published on 10 September 2026.
| Model | Input | Output | | --- | --- | --- | | GPT-6 Astra | $10 | $50 | | GPT-5.6 Sol | $4 | $20 | | GPT-5.6 Terra | $2 | $12 | | GPT-5.6 Luna | $0.20 | $1.20 | | Claude Fable 5.1 | $10 | $50 | | Claude Opus 5 | $5 | $25 | | Claude Sonnet 5 | $2 | $10 | | Claude Haiku 4.5 | $1 | $5 | | Gemini 3.8 Flash | $0.75 | $3.75 | | Gemini 3.1 Pro Preview | $2 | $12 | | DeepSeek V4-Pro | $0.435 | $0.87 | | DeepSeek V4-Flash | $0.14 | $0.28 |
Astra and Claude Fable 5.1 land on identical headline numbers, which is a good place to start being suspicious.
You are not comparing the same token
Anthropic's pricing page carries a note that almost nobody quotes: Claude 4.7 and later use a newer tokeniser that produces roughly 30% more tokens for the same text.
Take that seriously for a second. Two models both advertise $10 per million input tokens. Feed both the same document. One of them counts about 30% more tokens for it. The cheaper model on the table is the more expensive model on the invoice, and nothing in the headline rate tells you which.
Tokenisation differs across every provider — it always has — but a vendor documenting a 30% swing inside its own product line is a useful reminder of how soft the unit is. Price per million tokens is only comparable if a token means the same thing on both sides, and it doesn't.
The long-context shape differs more than the rate does
This is the part that actually decides bills, and all four providers have arrived at genuinely different answers.
- GPT-6 Astra charges 2× input and 1.5× output on any request over 272,000 input tokens. A step function, and the only warning is the invoice.
- Claude 4.6 and later include the full 1M window at standard pricing. Anthropic spells it out: a 900k-token request bills at the same per-token rate as a 9k-token request.
- Gemini 3.1 Pro Preview tiers at 200,000 tokens — $2/$12 below it, $4/$18 above.
- DeepSeek has announced peak-hour pricing keyed to two daily UTC windows, doubling rates inside them. As of publication it has been announced but not switched on.
So: one provider punishes long context with a cliff, one charges flat, one tiers at a different threshold entirely, and one is proposing to price by time of day. Those are four different billing models wearing the same $/token costume.
If your workload is retrieval-heavy or agentic — anything that accumulates context across a run — the shape dominates the rate completely. A model at twice the headline price with flat long-context pricing can easily be cheaper than a "cheaper" model with a cliff sitting in the middle of your token distribution.
The cache is where the money actually is
Every provider discounts cached input, and the multipliers are not close to each other.
Anthropic's standard cache read is 0.1× base input, but Fable 5.1 and Mythos 5.1 read at 0.025× — $0.25 per million against a $10 base. That's a 97.5% discount on the repeated part of a prompt. OpenAI caches Astra input at $1 against $10, the more conventional 10%.
But Astra's cache write costs $12.50 per million — more than sending the tokens uncached. That is fine when the prefix gets reused many times and actively harmful when it doesn't. It means the cache is a bet on your access pattern, not a free optimisation, and getting the bet wrong costs you 25% extra on every miss.
For any system with a large stable prefix — a system prompt, a document set, a codebase — these multipliers will swamp the headline rate. The comparison that matters is not "what does this model cost" but "what does my prompt shape cost on this model", and that ordering flips depending on how much of your prompt repeats.
The numbers have a shelf life, and it's short
Everything above is true on the day of writing and some of it has a published expiry:
- Gemini 3.8 Flash is $0.75/$3.75 through 31 December 2026, then doubles to $1.50/$7.50 on 1 January 2027.
- Claude Sonnet 5 launched at $2/$10 as introductory pricing due to rise to $3/$15 on 1 September 2026. Anthropic has since confirmed the rise won't happen and $2/$10 is now standard.
- DeepSeek's peak-hour surcharge is announced with no confirmed activation date.
One of those went up, one went down, one is in limbo — inside a single quarter.
Worth adding: while researching this I pulled the numbers from aggregator "LLM pricing comparison" pages first, and several were wrong. One had GPT-5.6 Sol at $5/$30 against the documented $4/$20; another had DeepSeek's peak pricing described as live. If you're making a real decision, go to the provider's own pricing page. The comparison sites are optimised for ranking on "LLM pricing", not for being right this week.
What I'd actually do
Don't pick a model from a table, including mine.
Take the two or three candidates that plausibly clear your quality bar, then model your workload against each provider's actual billing shape: your real token distribution, your cache hit rate, the fraction of requests that cross a threshold. A spreadsheet with your own numbers in it will beat any comparison post, this one included, and it takes an afternoon.
Then build the swap in. Put the model behind an interface with one implementation per candidate so that switching is a config change rather than a refactor. Given that pricing moved three times in three months across three providers, the ability to move is worth more than picking correctly today.
The benchmark leader will change. The pricing will change. The billing shape — cliffs, tiers, cache multipliers, tokenisers — changes more slowly, and it's the part that decides what you actually pay.