AI SaaS pricing compared: token-based vs per-seat vs outcome-based models

Anthropic charges $10 for every million input tokens sent to its Claude Fable 5.1 model. Microsoft charges $30 per user per month, paid yearly, for Microsoft 365 Copilot, and Intercom charges $0.99 for each outcome its Fin agent delivers. Comparing the three takes a buyer's own numbers: tokens in and out, seats in active use and conversations resolved each month. Every price in this comparison is a list price, read on the vendor's own page on October 6, 2026, and not a quote.

What does token-based AI pricing cost per million tokens?

OpenAI's pricing page names no single flagship; gpt-6-astra and gpt-6-luna are the highest and lowest priced of its gpt-6 rows, and gpt-5-nano is lower still at $0.05 input and $0.40 output. The older gpt-4.1, which earlier comparisons quoted, is still listed at $2 and $8. Those gpt-6 rates cover requests of up to 272K input tokens; above that, gpt-6-astra is $20 input and $75 output. At an assumed 100 million input tokens and 20 million output tokens a month, the bill is $2,000 on gpt-6-astra and $20 on gpt-6-luna.

On Anthropic's pricing page no model is priced above Claude Fable 5.1, and Claude Haiku 4.5 is the lowest-priced model not marked retired. Claude Sonnet 4, at $3 and $15, and Claude Opus 4, at $15 and $75, are still printed there, each marked retired except on named cloud platforms. Google's pricing page labels no model its flagship and calls Gemini 3.8 Flash its most intelligent Flash model. Of the three Google rows in the table, only Gemini 3.8 Flash carries the January 1, 2027 increase, and Google's output prices include thinking tokens. The older Gemini 2.5 Flash-Lite is still listed at $0.10 and $0.40.

OpenAI's batch guide, Anthropic's pricing page and Google's pricing page each state a 50 percent discount for batch processing. OpenAI's guide adds that a batch completes within 24 hours. OpenAI's caching guide says reused input tokens are discounted up to 95 percent, and that on GPT-5.6 and later models a cache write costs 1.25 times the uncached input rate, so a prompt that is written to the cache and never reused costs more than one sent without caching. Anthropic prices a cache hit at 0.1 times the base input rate on most models and a five-minute cache write at 1.25 times. Google's caching guide says cost savings are passed on when a request hits a cache and gives no percentage. OpenAI's cost guide lists selecting a smaller model among its ways to cut cost, the idea behind routing easy requests to cheaper models, and puts no percentage on it.

How much does a per-seat AI license cost?

OpenAI's Business FAQ points larger deployments to ChatGPT Enterprise and gives no price for it.

A Microsoft Copilot Blog post dated September 25, 2026 says everyday use stays on the per-user license while Cowork, Code, Autopilot and frontier models run on usage-based billing, which stays off for enterprise customers until an admin creates a spending policy; the post prints no credit rate. At list price, an assumed 1,000 seats cost $360,000 over a yearly term, and if 400 of those users are active the cost is $75 per active user per month, before the Microsoft 365 license underneath and any metered usage.

On its January 28, 2026 earnings call, Microsoft reported 15 million paid Microsoft 365 Copilot seats and, in separate remarks, more than 450 million paid Microsoft 365 commercial seats. Dividing the first figure by the second gives about 3.3 percent, which is this page's arithmetic and not a share Microsoft stated. On its July 29, 2026 call and in the earnings release of the same day, Microsoft reported more than 30 million paid Copilot seats; the call passages read do not repeat the commercial seat total, so no later share can be worked out from them.

What counts as a billable outcome in outcome-based pricing?

Four of Fin's outcome types cost $0.99 on Intercom's Fin pricing page: a resolution, a Procedure handoff, a disqualification and self-serve routing. Intercom renamed the metric from resolutions to outcomes on March 12, 2026. An unsuccessful attempt is not billed: Intercom's help article on Fin outcomes says a customer who asks for a human, or a Procedure that fails to complete, is not an outcome. An assumed resolution is billed: the article counts one when a customer leaves after Fin's last answer without asking for more help, or stays silent for 24 hours. A handoff to a human costs the same $0.99 when the buyer configured a Procedure to end in that handoff.

The Fin pricing page prints no $49 figure, while the help article still shows a $49 monthly base plan that includes 50 resolutions; 50 outcomes at $0.99 come to $49.50, so the two do not reconcile. Fin on Intercom's own helpdesk also carries a seat price, which the Intercom pricing page lists at $29, $85 or $132 a month on annual billing; Fin on an existing helpdesk has no seat cost. At list price, an assumed 2,000 billable outcomes in a month cost $1,980.

Zendesk's pricing page, last modified on October 6, 2026, puts AI agents in each of its Suite and Support plans, and the three plans it prices run from $19 to $115 per agent per month on yearly billing. Fin's outcome and Zendesk's automated resolution are different units: a handoff the buyer configured is billable at Intercom, while a case passed to a person falls outside Zendesk's definition, so the two prices do not measure the same thing. Salesforce's Agentforce pricing page also sells a per-user add-on at $125 per user per month and calls its prices informational and subject to change.

What should a buyer model before choosing a pricing model?

A buyer needs three volumes from their own records before any price here can be turned into a budget: tokens in and out each month for an API, seats that will be in active use for a seat license, and conversations that end in a billable result for an outcome plan. Outside those volumes sits every cost the headline price leaves out: a required base license, a monthly minimum, metered agents or usage billed on top of a seat. The vendor's written definition of the billable unit matters as much, because that definition sets the count the price is multiplied by. Before signing, a buyer should have the vendor's definitions of an assumed resolution, a configured handoff and a conversation written into the order form.