Cost Per Token vs Cost Per Task: Why LLM Pricing Comparisons Keep Misleading Founders
A model priced 2.5x higher per token can still produce a smaller bill, because what you actually pay for is tokens consumed per task, not the rate on the pricing page.
Key takeaways
- GPT-6 Astra lists at $10 per million input tokens and $50 per million output tokens, with a fast mode at $20/$100 for up to 2x the speed.
- That is roughly 2.5x the per-token rate of the model it replaces, yet several independent evaluators measured a lower cost per finished task.
- Artificial Analysis reported Astra using about one third of its predecessor's tokens inside the Codex harness, and about one fifth of Claude Opus 5's at high effort.
- The same model came out roughly 75% more expensive per task on their general Intelligence Index at max effort. Cheaper and pricier, same model, different workload.
- "Which model is cheaper" is not a property of the model. It is a property of your workload and your harness.
What actually changed with the Astra pricing
OpenAI published standard API pricing for GPT-6 Astra at $10 per million input tokens and $50 per million output tokens, plus a fast mode at double the rate for up to double the throughput. Latent Space summarised the launch as "2.5x pricier per token, but WAY cheaper per task", which is an unusual sentence to read about a price increase and a good reason to look closer.
The reconciliation is simple arithmetic. Your bill is the rate multiplied by the tokens you consume. If the rate goes up 2.5x and token consumption drops to a third, the product falls. Vendors have started optimising the second term, and buyers are still shopping on the first.
Why do the published numbers disagree with each other?
Because they measure different jobs.
On the coding side, Artificial Analysis put Astra at roughly one third of GPT-5.6 Sol's token consumption in the Codex harness and about one fifth of Claude Opus 5's at high effort, landing at less than half the cost of Claude Fable 5 for an equivalent coding-agent score. Cognition reported coming within 0.4 points of Fable 5 on FrontierCode 1.1 at 64% lower cost. Perplexity measured $11.98 per task on their WANDR benchmark, 13.5% better than Fable 5.1 at 6.1% lower cost.
On general intelligence, the same evaluator found Astra using only about 10% fewer output tokens at max effort. A 10% efficiency gain against a 2.5x price increase leaves you roughly 75% more expensive per task. Both results are correct. They describe different workloads.
The variable nobody prices: your harness
Here is the part that rarely makes it into the launch coverage. Token consumption per task is not fixed by the model. It is set by the scaffolding around it: how much context you resend, how you compact long sessions, how many retries you tolerate, how many tool calls a run makes.
On ARC-AGI-3, the reported score moved from roughly 63% to near 100% purely by changing the harness, and the cost per run moved with it. Vendors publish figures from harnesses they built and tuned. Yours is different. Their cost per task is therefore evidence, not a quote.
The practical consequence for a founder: treat published cost-per-task numbers as an upper bound on how confident you should be, and measure your own after a week of real traffic.
How do you actually compare two models on cost?
Run the comparison in this order.
- 1Fix the job, not the model. Pick one representative task from production, for example "resolve one support ticket end to end".
- 2Measure tokens, not requests. Log input and output tokens per completed task, including retries and failed runs. Retries are where quiet models get expensive.
- 3Multiply through at both rate cards. Only now do the published $/1M figures matter.
- 4Add the latency premium if you need it. A fast mode at 2x price is a real cost line if your product promises a response time.
- 5Divide by users. Cost per task times tasks per user per month is your COGS per seat. That is the number that decides whether your price holds.
A worked example, illustrative only: say a task takes 4,000 input and 1,500 output tokens on a $10/$50 model. That is $0.04 + $0.075, about $0.115 per task. If a cheaper model at a quarter of the rate needs three times the tokens to finish, it costs about $0.086, close enough that reliability, not price, should decide it.
What this means for your pricing page
If your COGS per task can swing by 3x depending on which model you route to, then a flat per-seat price is a bet on your routing staying stable. Most founders make that bet without noticing they made it. The fix is not necessarily usage-based pricing, it is knowing your margin per tier before the next frontier launch resets the rate card again.
Takeaway: stop benchmarking rate cards and start benchmarking your own cost per completed task, because the rate card is the one variable you do not control.
If you want to run that comparison without building a spreadsheet, you can model both providers at your own token split in the Calcaas provider comparison.
Frequently asked questions
What is cost per task for an LLM?
Cost per task is the total token spend needed to complete one unit of real work, including retries and tool calls, rather than the price of a single request. It is calculated by multiplying the tokens consumed per completed task by the provider's per-token rate.
How much does GPT-6 Astra cost?
OpenAI lists GPT-6 Astra at $10 per million input tokens and $50 per million output tokens for standard processing. A fast mode is offered at $20 per million input and $100 per million output for up to twice the speed.
Can a more expensive model produce a smaller bill?
Yes. If the higher-priced model finishes the same job using substantially fewer tokens, the product of rate and volume can fall. Independent evaluators reported exactly this pattern on coding workloads while finding the opposite on general-purpose ones.
Why do vendor cost comparisons differ from mine?
Vendors measure inside harnesses they built and tuned, with their own context management, retry policy and tool setup. Token consumption per task is highly sensitive to those choices, so your production numbers will differ.
Should I switch models every time prices change?
Not automatically. Switching costs include re-testing prompts, re-tuning the harness and revalidating output quality. Re-run your cost per task first, and only move if the gap survives your own workload. Place the JSON-LD above inside a <script type="application/ld+json"> tag in the page head. Source and further reading: Latent Space, [AINews] GPT-6 Astra, https://www.latent.space/p/ainews-gpt-6-astra-openais-biggest
More from the blog
LLM EconomicsCompute Capacity Belongs in Your AI Cost Model, Not Just Your News Feed
When a provider adds capacity and raises your rate limits, your effective cost per user can fall even though the list price per token has not moved at all.
LLM EconomicsA Cheaper Model Does Not Lower Your AI Bill: What Fable 5.1 Actually Changes
Anthropic's Fable 5.1 is built to reduce token cost, but your monthly AI bill is set by cost per completed task, not by the number on the model card.
LLM EconomicsOpenAI's Jalapeño Chip and What Custom Inference Silicon Means for LLM Pricing
OpenAI's custom Jalapeño inference chip beat the current state-of-the-art Nvidia system on tokens served per user and throughput per kilowatt, a result that points toward cheaper inference and more pressure on provider pricing.
Pricing math, in your inbox.
One short note a week on AI pricing, token economics, and margin. No spam, unsubscribe anytime.