token pricing is effectively meaningless now. 3.8 flash looks 13x cheaper when measured per token, but Astra is cheaper per task since it's far more efficient. measure your costs per task, not per token.
cheaty
@cheatyyyy
Sep 3
GPT-6 Astra is on the pareto-frontier of cost efficiency due to being EXTREMELY token efficient. It is in a whole league of it's own. It is cheaper than Gemini 3.8 Flash per task (a model 13x cheaper than Astra).

Sep 3, 2026 · 11:53 PM UTC

66
111
2,161
195,546
also "tokens" isn't a standard unit of measurement across providers, or even models from the same provider. Anthropic snuck in a 30% cost increase to Opus 4.7 just by making their tokenizer less efficient
Maxime Labonne
@maximelabonne
Apr 29
Turns out Anthropic basically snuck in a 30% cost increase by changing the tokenizer
1
51
6,324
Sort replies: Relevant Recent Liked
Replying to @stevenheidel
4
425
Replying to @stevenheidel
this is an argument someone who's never used gemini OR chatgpt would make. 95% of tasks gemini is going to be cheaper and use fewer tokens. how could you even possibly come to this conclusion????? rage bait ahh post
1
21
1,435
Replying to @stevenheidel
"tasks" can not be quantified tho
2
9
627
Replying to @stevenheidel
Per task is relative.
9
799
Replying to @stevenheidel
how is a task defined
4
315
Replying to @stevenheidel
but if you give Astra a long running complex task presumeably it has the ability to absolute drain the budget
3
528
Replying to @stevenheidel
measure tokens per task, not price per token.
2
314
Replying to @stevenheidel
One “task” is an excellent unit of measurement, just like a “foot” 🦶
196
Replying to @stevenheidel @BorisMPower
How do you meter the cost of supplying the models if work is done in the latent space? 🤔 GPU hours?
798
Replying to @stevenheidel @ShanuMathew93
Not always, it must depend on the type of task
32
Replying to @stevenheidel
Wrong, the meaningslessness comes from Gemini's "task results" being absolute cancer, not that it uses more tokens
129
Replying to @stevenheidel
Are all tasks created equal? What qualifies as a task?
119
Replying to @stevenheidel
Andrii Chornyi
@postmortemdaily
Sep 4
Here are 6 levels of developer Claude, look at which one you're on?
382
Replying to @stevenheidel
Astra isn't even available. So what do you mean its cheaper?
147
Replying to @stevenheidel
sorry but in real Life for us the huge input increase will result in much more costs since people LOVE to just input all unnecessary things in prompt
139
Replying to @stevenheidel
so every interaction with an LLM has to be a task? Why?
398
Replying to @stevenheidel
You need to stop brand hating and see models objectively for what they are good for. 3.8 flash has same AA intelligence as astra-medium while being cheaper and faster. The graphs literally show that too
569
Replying to @stevenheidel
Oh no, useless benchmarks are useless. Keep us posted
238
Replying to @stevenheidel
what's a task?
408
Replying to @stevenheidel
Good takes only
49
5
537
41,291
Replying to @stevenheidel @BorisMPower
Token pricing has always been an incomplete picture. Gotta measure all 3 Es Efficacy, Efficiency, and Expense
1
187
Replying to @stevenheidel
Yes. The number that matters is what you pay to finish a job But again, that’s something you can’t really calculate since everyone’s expectation of what counts as completed can be very different.
32
Replying to @stevenheidel
yeah exactly
177
Replying to @stevenheidel
I wouldn't even measure it per task but per project, you want to delegate an entire project to Astra not a task
92
Replying to @stevenheidel
Per-task cost is the useful metric only when it includes retries, review, and the failures you still have to catch.
6
435
Replying to @stevenheidel
I’ve been using Gemini and Sol high pretty equally for the past couple days. Gemini is still about 5x cheaper and faster per task. I don’t think Astra will change things that much. Get anyone who says it’s cheaper per task to show you the receipts. It’s still not close.
1
4
390
Replying to @stevenheidel
For real work, price per task is the useful metric; token counts are only a proxy.
3
173
Replying to @stevenheidel
My god. Astra can also complete tasks. Flash can “do some” things. Astra is a senior engineer with a team. Flash is … a guy who can bang out SQL.
1
3
225
Replying to @stevenheidel
Yes, this is the measurement I pay attention to. Seeing these charts, I'm now expecting my cost per task will go down despite Astra's price per token being higher than sol. Total spend/usage will continue up though, because now we can solve bigger problems!
1
229
Replying to @stevenheidel
Soon we won’t care which model did the work. Just whether the job got done.
1
35
Replying to @stevenheidel
If token efficiency scales like this and smarter models make larger models even cheaper (as we saw by the luna price cuts), 30T models might be completly common sense and normal within a few months.
1
572
Replying to @stevenheidel
Trust me bro cheaper
1
139
Replying to @stevenheidel
Token price still matters as task is undefined. You can’t control TCO without control unit cost.
14
Replying to @stevenheidel
Thats exactly it. People dont understand token efficiency drops cost even if the api pricing is higher per token. Fable is costing about 1.5-10x cost (task dependant) in comparison to astra.
37
Replying to @stevenheidel
Just as token cost is wrong metric for a model. Proper metric is “how much value do I get out of this model”.
15
Load more