OpenAI has launched three versions of its ChatGPT 5.6 frontier AI model: Sol, the flagship model; Terra, which is balanced for everyday work; and Luna, a fast and affordable variant.
The release, following closely behind Anthropic’s Fable 5, has sharpened an increasingly important question for customer experience leaders around whether enterprises should keep chasing the most powerful frontier large language models (LLMs), or focus on models that are smaller, cheaper, more controllable, and better aligned to specific CX workflows.
The answer appears to be both, but not for the same jobs.
Frontier models are advancing quickly, and their reasoning, multimodal capabilities, coding performance and agentic potential continue to stretch what enterprises believe AI systems can do. For CX operations, that raises the prospect of more sophisticated virtual agents, richer knowledge retrieval, better personalization and AI systems capable of handling complex, multi-step customer journeys.
Yet the practical reality inside many contact centers is that much of customer service does not require the most powerful model available. It requires fast, reliable, compliant execution of repeatable tasks from classifying intent to summarizing conversations, retrieving policy information, checking answers against approved knowledge, routing cases, detecting sentiment and escalating when needed.
For those use cases, a smaller language model that has been specifically trained or tuned for a particular workflow or customer service domain may be not only sufficient, but more effective.
As Ashish Nagar, Founder and CEO of enterprise AI software firm Level AI told CX Today:
“If I want to go from Mountain View to San Francisco Airport, I don’t want to take a Boeing 747 to do that. I just want to take an Uber. Using a GPT-5 for a simple CX task is like taking a Boeing 747 to go to the airport, which is 30 minutes away. You need an Uber, which is a small language model that’s specifically trained to navigate this path.”
Many CX applications are narrow, high-volume and cost-sensitive, and in those environments the question is not “what is the most intelligent model?” but “what is the right level of intelligence for this job?”
That distinction is becoming more significant as enterprises encounter the operational costs of large-scale AI deployment.
The Token-Burn Problem
As GenAI moves from experimentation into production, token consumption has become a board-level concern. A proof of concept may look compelling when usage is limited. But once an AI agent begins handling thousands or millions of interactions, inference costs can rise quickly.
In a recent CX Today interview, Rebecca Wettemann, CEO and Principal Analyst at Valoir, highlighted two fears now shaping enterprise adoption.
“A lot of them are hearing the stories about two things,” she said. “Either AI running off in the middle of the night and doing something deleterious, or waking up in the morning and finding that they have a huge token bill that they didn’t expect.”
But many of the most dramatic examples today come from coding use cases, rather than customer service.
“Most of the time that those two things are happening today are with things like using AI for coding, where it’s deleting files or databases or that sort of thing. It’s not in the customer service context necessarily,” Wettemann noted. “But we hear that and people say, ‘wait a minute, I want to make sure before I put this in production that it’s actually going to do a positive interaction with my customer if it’s customer facing, and that it’s not burning through tokens.’”
That anxiety is significant because customer service is a high-risk environment for AI, with models often acting directly in front of customers. Regulated industries in particular have a low tolerance for improvisation.
At the same time, customer experience is also one of the areas where AI has the clearest ROI potential. Reducing handle time, automating common queries, improving agent assist and increasing self-service containment can all deliver measurable value. The challenge is implementing AI in a way that is economically viable and operationally safe.
Salesforce Emphasizes “Precision Over Power”
That is why model right-sizing is moving from an engineering detail to a strategic CX issue.
As Jayesh Govindarajan, EVP of Software Engineering at Salesforce AI, wrote in a blog post, most enterprise work needs “the right intelligence for each job—not the most intelligence for all of them.”
Govindarajan described how Agentforce previously relied on a single rented model, causing token bills to grow linearly with traffic. Rather than simply passing those costs on to customers, the company rebuilt its architecture, breaking tasks apart and tuning specific open-source models for defined jobs.
The company’s conclusion was that a general-purpose frontier model can perform many tasks, but often “more slowly, more expensively, and less precisely” than a model designed for a specific function.
Salesforce’s architecture now uses targeted models for steps such as safety screening, intent detection, grounding and answer validation, while still reserving a frontier model for core multi-step reasoning.
“Today, these precision models run a growing share of the stack, and they’re already cutting costs, running more effectively, and giving us more control over our products and roadmap,” according to Govindarajan. “A frontier foundation model still handles the core muti-step reasoning, but now that reasoning is governed by the harness we’ve built around it—and that means that we, and our customers, can swap the foundation models freely, instead of being dependent on any single vendor.”
The larger lesson for CX leaders is that model selection is not a binary choice between frontier and small models; it is becoming an orchestration problem.
A modern AI architecture may route different parts of the customer journey to a small model for intent classification, a retrieval model for knowledge grounding, a specialist model for compliance checks and a frontier model only when complex reasoning or novel problem-solving is required.
The Risk of Outsourcing the AI Roadmap
In a recent discussion with CX Today at Agentforce World Tour in London, Salesforce Futures VP Mick Costigan framed the issue as one of business risk as much as cost.
“One of the questions from a scenario point of view has always been: if you fine-tune today’s model, what about the next model? There’s that kind of uncertainty. And then what people are increasingly becoming aware of is the token cost. So there’s an efficiency question that’s really important.”
Costigan noted that frontier models may be capable of “incredible things,” but enterprises still have to consider whether access is stable, affordable and controllable enough for production environments.
“If you find it impossible to use them, or if access to them is not stable enough, then are you outsourcing or creating a risk in your business?” Costigan said.
If an enterprise builds its customer-facing automation entirely around a single external frontier model, it may become exposed to pricing changes, latency issues, availability constraints, model behavior shifts and vendor roadmap decisions.
For many organizations, particularly those with established service operations, the safer strategy may be to build a model-agnostic architecture, using frontier models where they add differentiated value, but avoiding making them the default engine for every task.
Why Small Models Fit CX, But Frontier Models Still Matter
Small language models (SLMs) are not simply cheaper versions of larger models. In CX, their advantage can come from specialization.
A smaller model trained on contact center data and industry-specific terminology may perform well on the bounded tasks that dominate customer service. It may also be easier to evaluate, govern and deploy within enterprise constraints.
For example, CX teams may not need a frontier model to identify whether a customer is asking about a refund, a delivery delay, a billing issue, or a password reset. Nor do they necessarily need one to summarize a call, detect dissatisfaction, suggest the next best action, or check whether an agent’s response includes a required disclosure.
In many cases, using a large frontier model for those jobs introduces unnecessary cost and latency. It can also make the system harder to control.
The fact that Fin (formerly Intercom) has developed its own customer service-specific AI model is part of what made it an interesting acquisition target for Salesforce, Costigan noted.
Nagar’s Uber vs. Boeing 747 analogy captures the operational reality: the best tool is the one optimized for the journey.
The CX AI Stack Is Becoming Layered
None of this means frontier models are irrelevant to CX. Their role may become more important as organizations move from basic automation to more ambitious AI-enabled service transformation.
Costigan argued that companies need to hold two thoughts at once. There is immediate ROI available from applying AI to known workflows without necessarily relying on the most advanced models. At the same time, innovation teams should explore how frontier models could reshape the organization more fundamentally.
“There are ways of getting value, and you need to be pushing ahead on that. And then also you need an innovation group that’s thinking in a more fundamental way about how to restructure, and they probably have a higher reliance on frontier models.”
The release of more powerful models will expand what AI can do, with each new release bringing stronger benchmarks and broader capabilities.
For CX tech buyers, that means asking more specific questions, such as which tasks genuinely require frontier-level reasoning, how the organization will monitor token consumption, how answers are grounded in approved enterprise data and what happens when model pricing or availability changes.
In customer experience, the most effective approach may be the one that uses frontier intelligence when necessary, small language models when sufficient, and enterprise architecture to hold the whole system together.
