Prices for access to generative AI models and features could tank in 2025 amid cheaper foundational models, open source advances and vendors becoming realistic about what customers will actually pay.

Welcome to the only thing in your life that'll be deflationary--generative AI. We're only a month into 2025, but there have been a series of moves indicating that the generative AI gravy train is unlikely because customers will soon be focused on cost per query.

Here are a few reasons why enterprise AI is likely to become less expensive.

Wither add-ons?

Rewind to a year ago and Wall Street was busy modeling revenue growth due to $30 per month per user add-ons for access to generative AI features. Later, those add-on prices became more like $20, but the working theory from vendors was that enterprises will pay for AI that drives productivity.

During this time, the vendors that went for genAI usage and raised prices on their overall SKUs were mocked. Wall Street analysts were almost indignant. When are you monetizing genAI? Zoom, Adobe and Workday all got flack for noting what was kind of obvious--genAI is a feature not the end game.

Fast forward to January 2025 and copilot add-ons are toast. Oracle just launched AI agents for its sales applications without charging extra. Google Workspace dropped Gemini add-on charges, but raised business and enterprise plan prices. Microsoft launched Copilot Chat and now includes copilots in Microsoft 365 in many cases. There are charges for AI agents, but many genAI features will be bundled into enterprise plans.

Sign up for Constellation Insights newsletter

Microsoft did something similar with its consumer Microsoft 365 plans and now copilot is everywhere (even if you didn't want it). For what it's worth, I'm more of a Notepad person. Like my appliances, I often like my tools to be on the dumb side. Don't bog me down with features I don't need or want.

Now that Google and Microsoft have ditched the add-on game, rest assured that other software vendors will follow. The argument that genAI is all just part of the software has won out.

Sure, you'll be hit with other AI charges and consumption models, but the add-on game is over.

LLM pricing is going to collapse

The big news this week is that a Chinese AI company called DeepSeek set out to blow up OpenAI's business model. DeepSeek, combined with other open source large language models, is going to be a real threat to model pricing, which revolves around tokens and API calls. ByteDance, owner of TikTok, also released Doubao 1.5 Pro, a model with strong performance.

Yes, there are concerns about censorship on DeepSeek and Doubao 1.5 Pro, but the idea that you can get API access to DeepSeek's R1 model for 14 cents for a million tokens compared to OpenAI's $7.50 is disruptive. OpenAI is clearly positioned as a premium LLM model, but that pricing is disruptive for a company that can't make money on a $200 a month subscription plan.

This DeepSeek news landed on Monday and was overshadowed by an inauguration in the US featuring technology bigwigs, AI infrastructure plans and chatter out of Davos and earnings season.

Nevertheless, it's worth taking DeepSeek for a spin. Many of these models have caught up to what OpenAI can do and are likely good enough for most use cases.

Some reading:

Simply put, LLM margin compression is here. The scary part is the LLM giants didn't have profit margins to begin with. LLMs are going to commodity in a hurry.

AI agents will require price transparency

Although vendors are wrapping in genAI as a bundle, the agentic AI as labor replacement/augmentation rap is just starting.

Enter the consumption model. Enterprises (allegedly) will pay for agentic AI conversations, problems solved and value created by the simple fact companies won't need to add a human.

The problem: The SaaS vendors pushing this agentic AI consumption model aren't used to the level of transparency needed yet.

Cloud providers have consumption dashboards, more transparent pricing and ways to manage costs. SaaS vendors simply don't.

Once consumption becomes part of the mix, SaaS vendors will have no choice but to be more transparent.

Today, SaaS packaging and contracts are complicated and the sales cadence is almost engineered so enterprises make tactical errors along the way. Companies will need to know exact costs by AI use case especially with agentic AI.

We'll have a hybrid seat, subscription and consumption model for the foreseeable future, but ultimately SaaS vendors are going to have to show the math behind the pricing. Value would be nice too.

What's next?

As early genAI pricing models are disrupted, you'll most SaaS vendors want to become platforms, LLMs bundled with broader software suites and a more holistic sales pitch.

SaaS vendors will want to be more like platforms that enterprises use to leverage AI. ServiceNow is already seen this way, but look for many vendors to follow the same path. What's unclear is whether CxOs who have been cross-sold to oblivion will suddenly think their vendors are platforms.

Meanwhile, LLM visionaries are going to attempt to look more like SaaS companies.

One thing that caught my eye out of Davos was Mistral's take that enterprises will move away from models and to systems. Mistral CEO Arthur Mensch told CNBC that models are merely part of systems that include data and tools that act as agents.

Cohere launched North, an AI platform that combines LLMs, search and agents in one collaboration platform. Anthropic is adding collaboration features as it expands use cases for its Claude models. OpenAI is also expanding but has largely focused on expanding into search to compete with Google.

If you follow the LLM players, it's clear that they think foundational models alone aren't going to pave the way to profitability.

The other thread to watch is how the hyperscalers fare in the agentic AI age. Consumption pricing is already there and AI makes it a lot easier for cloud giants to go vertical as well as horizontal.

In the end, 2025 is going to be a year where enterprise vendors attempt multiple models. And margin compression may be inevitable.

Business Research Themes

Data to Decisions, Next-Generation Customer Experience

Roles

Chief Information Officer

Tags

#AI #artificialintelligence #machinelearning #ML #generativeAI