Token prices have dropped roughly 1,000-fold since 2021. GPT-3-quality output that cost $60 per million tokens in late 2021 sold for $0.06 by late 2024. If you assumed that meant your AI bill was heading toward zero, Gartner published a correction on August 17: AI inference costs per agentic workflow will increase more than fivefold through 2028, because as AI products evolve from assistive features to multistep execution, product leaders face a new margin challenge - falling model prices are subsidizing more complex workflows while escalating total AI costs.
Gartner calls this the inference paradox. Token prices fall roughly 95% by 2030, and inference costs per agentic workflow rise more than fivefold through 2028. Both are true, because each new generation of agent capability spends more tokens per task - and often on a more expensive model tier - faster than the per-token price falls.
This is not an abstract forecast. Salesforce announced on September 11 that 7 billion Agentic Work Units have been delivered across Agentforce and Slack, including 3.2 billion in the second quarter alone. The volume is already real. The question is whether teams understand what is driving those costs.
Why an agent costs more than a chatbot per task
The price on your provider's pricing page is per token. The cost of a completed task is tokens multiplied by steps, context size, model tier, and retry rate. Those multipliers are what Gartner is measuring.
Where a simple chatbot must read and interpret a query and quickly respond with a probabilistically reasonable answer, an AI agent must constantly reason, negotiate, and question itself. That loop is expensive in a way that per-token pricing hides. A simple chatbot interaction may involve one request and one response. An agentic workflow can involve a sequence of model calls: planning a task, selecting tools, retrieving information, interpreting results, checking work, correcting mistakes, and deciding what to do next.
Anthropic measured this directly in production. Agents typically use about 4× as many tokens as chat interactions, and their multi-agent research system uses about 15× as many tokens as chat. The 15× figure is not a worst case - it is what happens when a lead agent orchestrates subagents that each carry their own context windows across parallel workstreams. The cost multiplier is the price of admission to the architecture. If the task does not decompose into parallel directions, you pay it without earning it.
Gartner identified three factors behind the trend: falling costs for foundational AI models, the use of more powerful models for advanced applications, and the much higher token requirements of complex AI workflows.
The third factor is the one most teams underestimate, because it is invisible on a pricing page.
What the Salesforce Agentforce launch shows about where those tokens go
Salesforce introduced seven named Agentforce AI agents - Casey, Paige, Carter, Hunter, Marshall, Piper, and Fin - each built for a specific business function in sales, service, commerce, IT/HR, supply chain, and customer experience on September 11, 2026. The launch is a useful anatomy of what production agents actually do.
Six agents are generally available now, while Hunter remains in pilot and is the first to use a new long-horizon runtime that pursues goals over weeks instead of a single chat session. That last point is where the inference paradox gets concrete. A workflow that runs over days re-reads its context on every step - tool schemas, prior decisions, business rules, accumulated output. Each re-read is billed.
Hunter helps business-to-business salespeople find leads, craft outreach emails, and prepare for presentations. It uses a module called the long-horizon runtime to maintain work plans across chat sessions. A workflow like that - researching a prospect, tracking a deal across weeks, drafting outreach - is dozens of model calls, not one.
The early numbers from Salesforce customers are telling. 70% of Autism Queensland's administrative requests are resolved by its employee service agent, while 90% of core shopper journeys are handled by Hibbett AI, which went live in six weeks. Asana's website agent is now driving 4× the conversation volume. Resolution rate and conversation volume are the wins being cited. Cost per resolved interaction is not mentioned.
Salesforce did not disclose pricing for any of the agents or the supporting capabilities. That gap - between headline resolution rates and actual per-workflow inference cost - is exactly where teams get surprised.
Beagle in action#ops-budget, Monday morningThe fix: inference tiering, not token price negotiation
"Product leaders cannot rely on more efficient token economics to rationalize AI costs," Gartner analyst Will Sommer said in the August 17 report. There is no reliable, economical one-size-fits-all model on the horizon. Producing competitive AI products will require developing and maintaining complex multimodel ecosystems.
The prescribed fix is inference tiering: routing each step in a workflow to the cheapest model that can handle it reliably, rather than sending everything through the same frontier model. The logic is that most steps in an agent loop do not require the full capability of a top-tier reasoning model. Retrieval parsing, format conversion, decision routing - these are cheaper jobs.
| Step type | What it needs | Model tier to consider |
|---|---|---|
| Planning / goal decomposition | Strong reasoning | Frontier (Claude, GPT-4o) |
| Tool selection | Instruction following | Mid-tier (Haiku, GPT-4o-mini) |
| Document retrieval parsing | Pattern extraction | Small, fast model |
| Result synthesis | Reasoning + coherence | Frontier |
| Format / output generation | Instruction following | Mid-tier |
"Defaulting to generic autonomous intelligence will result in unbounded costs orders of magnitude higher than those of optimized product ecosystems," Sommer said. That sentence is worth putting in front of whoever is approving agent deployments.
What companies need to verify is not whether they have lowered the per-token price for each model. It is whether they can track, workflow by workflow, the cost per accepted outcome, the success rate, the number of retries, and the burden of human review.
The non-obvious consequence here: teams that centralize all agent work in one tool - and many are, given Slack's position as the surface where Agentforce AWUs are counted - have poor visibility into which workflow is burning which budget. The bill arrives as a total. The steps are invisible unless you instrument them.
A teammate like Beagle, running inside Slack with a draft-and-approve model, keeps a human in the loop on each send - which also means each send is a natural checkpoint where cost-per-outcome becomes auditable rather than invisible.
Agentic AI inference costs: common questions
What is Gartner's inference paradox?
Gartner's own name for a situation where the price per token keeps falling while the cost of a completed agentic workflow keeps rising. The two trends do not cancel because workflows consume more tokens per task - and often more expensive tokens - as agent capability increases. Gartner forecasts per-workflow costs up more than fivefold by 2028, while per-token prices fall 95% by 2030.
Why do AI agents use more tokens than chatbots?
Agents re-read their full context on every step: instructions, tool schemas, prior reasoning, tool outputs. Anthropic's engineering team measured this directly in production: agents typically use about 4× as many tokens as chat interactions, and their multi-agent research system uses about 15× as many tokens as chat. The multiplier grows with the number of agents and the length of the task.
What is inference tiering for agentic workflows?
Inference tiering routes each step in an agent workflow to the cheapest model capable of doing that step reliably, instead of sending all calls to one frontier model. Ensuring ROI from advanced AI, like reasoning agents, demands exponentially higher returns relative to basic models, or highly optimized inference-tiering, routing and orchestration to calibrate complex tasks relative to more cost-efficient intelligence.
Do reasoning models make agentic workflow costs worse?
Reasoning models can consume 100× more tokens internally than they output, creating a cost paradox where cheaper per-token pricing leads to higher total bills. Using a reasoning model for every step in a workflow - rather than only the steps that need it - is one of the fastest ways to make the inference paradox visible on your bill.
What did the Salesforce Agentforce launch show about agent costs?
Salesforce reported that Agentforce and Slack have collectively delivered 7 billion Agentic Work Units, with 3.2 billion in Q2 alone
- confirming that agent volume at enterprise scale is already significant. Salesforce did not disclose pricing for any of the agents , which means teams adopting Agentforce need to model their own per-workflow inference costs rather than waiting for the bill to arrive.