Skip to content
TrustList
Blog

What an AI task costs: a 200-fold spread across 38 priced models on our board

Editorial

By TrustList Editorial

A 50,000-token summary costs a quarter of a cent on the cheapest text model on our board and 60 cents on the dearest. We priced five tasks across 38 models, and list the launch prices and premiums that change the bill.

About What an AI task costs: a 200-fold spread across 38 priced models on our board

What an AI task costs: a 200-fold spread across 38 priced models on our board

Summarising a 50,000-token document, roughly a 35,000-word report, costs a quarter of a cent on the cheapest text model on our AI model board and 60 cents on the dearest. That is a spread of more than 200 times for the same job, and it holds for every everyday task we priced: answering a chat message, drafting a document, running an agent, or classifying ten thousand support tickets.

The spread inside a single vendor's range is almost as wide. Anthropic's smallest current model, Claude Haiku 5.5, does that document summary for 0.6 cents, its mid-range Sonnet 5.5 for 12 cents and its largest Opus 5.5 for 24 cents. OpenAI's GPT-6 Luna, GPT-6.1 Sol and GPT-6 Astra run from 0.6 cents to 60 cents. For a business putting AI into a product or a workflow, the choice of model tier matters far more than the choice of vendor.

This piece prices five common tasks across every text model on our board that has a published per-token price, shows which half of the price list matters for which job, and lists the price changes already scheduled for models people are building on today.

How we counted. Our AI model board lists 121 models. 46 have a published price per million input and output tokens in our records, each read from the vendor's own pricing page, model card or announcement and linked from the model's page. We left out eight that cannot be compared on a text task: two text-to-speech models, two real-time voice models, a voice model with a launch discount, two models priced in a different way, and one whose maker says API access "will be provided". That leaves 38 text models. Prices are the standard list rate on 14 October 2026, before batch, caching or volume discounts. Costs are our arithmetic.

Five everyday tasks, priced

We defined each task by the tokens it sends and receives, which is how every vendor bills.

Task Input tokens Output tokens Cheapest Median Dearest
Answer a chat message 2,000 500 $0.0002 $0.007 $0.045
Draft a document 3,000 3,000 $0.0006 $0.024 $0.18
Summarise a long document 50,000 2,000 $0.003 $0.08 $0.60
Run an agent through a multi-step job 500,000 50,000 $0.03 $1.13 $7.50
Classify 10,000 tickets 10,000,000 500,000 $0.58 $17.50 $125.00

The middle of the board sits at roughly 1.40 to 2 dollars per million input tokens and 6 dollars per million output tokens. A chat reply at the median costs less than a cent, which is why chat features rarely break a budget. The agent run is where costs bite: half a million tokens of input is a modest agent job once it has read files, called tools and fed its own results back in, and at the median it costs over a dollar each time it runs.

The dearest models are not dearest by a little. Four of the 38 charge more than 20 dollars per million output tokens, and the top tier at 50 dollars charges more than 300 times what the cheapest charges for output.

Input or output: which price matters

Every vendor prices output tokens higher than input, because generating text costs more compute than reading it. On our board the output price is a median of four times the input price, ranging from two times to ten times. Eighteen of the 38 models charge five times or more for output.

Which of the two prices drives a bill depends on the job:

Task Share of cost from output, median model
Summarise a long document 14%
Classify 10,000 tickets 17%
Run an agent 29%
Answer a chat message 50%
Draft a document 80%

For reading jobs, such as summarising contracts, searching a knowledge base or classifying records, the input price is what to compare, and a model with cheap input and expensive output can be the best buy. For writing jobs, such as drafting emails, reports or code, the output price dominates. Comparing models on a single headline number hides this, and many comparison tables quote only the input price.

Reasoning models complicate the picture further. Several vendors bill the tokens a model spends thinking before it answers as output, so a model that reasons at length can produce a short visible answer and still charge for thousands of output tokens. Google's pricing page for Gemini 4 Argon, for example, states that its output price includes thinking tokens.

The ladder inside one vendor

The large vendors now sell their models as a ladder, and the rungs are far apart:

Vendor Small tier Middle tier Top tier
Anthropic Haiku 5.5: $0.10 in, $0.50 out Sonnet 5.5: $2 in, $10 out Opus 5.5: $4 in, $20 out
OpenAI GPT-6 Luna: $0.10 in, $0.50 out GPT-6.1 Sol: $2 in, $10 out GPT-6 Astra: $10 in, $50 out

The two vendors price their small and middle tiers identically. The middle tier costs 20 times the small one; OpenAI's top tier costs 100 times its small one. An agent run that costs 7.5 cents on a small model costs 1.50 dollars on the middle tier and 7.50 dollars on GPT-6 Astra.

The practical consequence is routing. A system that sends every request to the top tier pays for intelligence most requests do not need. The common pattern is to send simple requests, such as classification, extraction and short replies, to a small model, and escalate to a larger one only when the small model is unsure or the task is hard. Done well, most traffic runs on the cheapest rung.

A worked budget: a support desk

To see how these choices add up, take a support team that handles 2,000 tickets a day and uses a model for two steps on each: classifying the ticket (about 1,000 tokens in and 50 out) and drafting a reply for an agent to check (about 2,000 tokens in and 500 out). That is 3,000 input tokens and 550 output tokens per ticket, or about 180 million input tokens and 33 million output tokens in a 30-day month.

Approach Monthly cost at list price
Everything on a small tier ($0.10 in, $0.50 out) $34.50
Everything on a middle tier ($2 in, $10 out) $690
Everything on the dearest tier ($10 in, $50 out) $3,450
80% small tier, 20% escalated to the middle tier $165.60

The gap between the first and third lines is a hundred times, for the same tickets. The routed approach in the last line costs less than a quarter of sending everything to the middle tier, and it is the pattern most production systems end up with once they measure where the small model falls short. Moving the classification step to a batch tier, where the vendor offers one, would cut that part of the bill in half again, because classifying a ticket rarely needs an answer within the second.

None of these figures includes the cost of the people who check the drafts, which for most support teams will still be the larger number.

Prices that change after you sign up

A per-token price is not fixed for the life of a project. Our records already carry five kinds of scheduled or conditional change:

  1. Launch prices that end. Google's Gemini 4 Argon is priced at 2 dollars in and 10 dollars out until 30 November 2026 and 4 dollars and 20 dollars from 1 December, so its costs double for every customer at once (our news item). Google's Gemini 3.8 Flash TTS and Flash-Lite TTS voice models carry launch prices until 31 December 2026, after which input doubles to 1 dollar and output rises to 18 and 12 dollars. Inception lists a 50 per cent launch discount on Mercury Voice.
  2. Long-context premiums. Claude Haiku 5.5 charges 0.10 dollars in and 0.50 dollars out for prompts up to 100,000 tokens, but 0.50 and 2.50 dollars above that, five times as much. GPT-6.1 Sol's long-context rate is 4 dollars in and 15 dollars out. A document pipeline that crosses the threshold can see its costs multiply without any change of model.
  3. Batch discounts. Anthropic and OpenAI both charge half price for batch requests that may wait for an answer, and OpenAI offers the same for its flex tier. Overnight jobs, such as classification and bulk summarisation, rarely need an instant reply.
  4. Caching. Repeated input, such as the same instructions or the same document sent with every question, is billed far below the list rate when cached. Gemini 4 Argon's cached input is 95 per cent cheaper than its input price, and Claude Haiku 5.5 charges 0.01 dollars per million for cache reads.
  5. Preview prices. Fireworks' Ember-1 and Mistral Large 4 were listed at research-preview or public-preview prices when we recorded them, which vendors usually revise at general release.

Meta takes the opposite position for Muse Wonder, stating that its standard mode has no long-context premium and no peak tier.

Open weights and hosted prices

Three of the 38 priced text models also publish their weights under an open licence: Agnes 3.0 Flash under Apache 2.0, and Xiaomi's MiMo-V2.6-Flash and MiMo-V2.6-Pro under MIT. A fourth, StepFun's Step 5 Preview, has announced open weights for 15 October 2026 but had not published them when we recorded it. The three open models' hosted API prices are among the lowest on the board; MiMo-V2.6-Flash charges 0.14 dollars in and 0.28 dollars out.

Open weights add a second option: running the model on one's own hardware or a rented GPU, which turns a per-token price into a fixed infrastructure cost. That pays off only at steady, high volume, and it brings the work of serving, scaling and securing the model. For most small businesses the hosted API of an open-weight model is the cheaper route, and the open licence is insurance: if the provider raises its price or closes the API, the model can be run elsewhere.

How to budget an AI feature

From the prices above, a few rules follow that a buyer can apply without us:

  1. Count tokens, not requests. Log the input and output tokens of a sample of real tasks before choosing a model, because a request can be 500 tokens or 500,000.
  2. Price the task on both halves of the price list. Reading-heavy work should be compared on input price, writing-heavy work on output price.
  3. Start on the smallest rung that does the job, and test a larger one only on the cases where the small one fails.
  4. Put every launch price and preview price in the calendar with its end date, and budget at the post-launch rate from the start.
  5. Move anything that can wait to batch, and cache what repeats. Both are discounts of half or more that need only a change of settings.
  6. Watch the context threshold. Splitting documents to stay under a long-context boundary can cut costs several times over.

AI models in our catalogue

Every model priced here has its own page on our AI model board, with its price, context window, licence and release date, each linked to the vendor's own source. Among them are Claude Haiku 5.5, GPT-6 Luna, Gemini 4 Argon, Muse Wonder, MiMo-V2.6-Flash and GPT-6 Astra. We add new models and price changes as vendors publish them, so the figures in this piece will drift; the model pages carry the current price.

Related on TrustList:

Sources

Categories & features

TrustList Weekly

The week in software and IT, in one email

The news that matters to buyers, new rankings and our own research. Every Thursday, free, and easy to leave.

We will email you to confirm. Unsubscribe with one click in any issue. Privacy policy

Keep exploring

More on TrustList

Everything here links back to the same verified catalogue. Pick your next stop.