AI model prices are falling: the case against AI vendor lock-in

Editorial illustration of AI vendor lock-in, showing falling model costs, a crumbling barrier and multiple pathways representing the freedom to switch between AI providers.

AI vendor lock-in now has a visible price. Leading AI models are improving and getting cheaper quickly, and a system tied to one provider cannot take advantage of either without a rebuild.

In a single week, Google launched Gemini 4 Argon at aggressive introductory pricing and Meta released its Muse Wonder models. Both arrived about ten weeks after OpenAI cut GPT-5.6 Luna prices by 80%. Leading models are improving and getting cheaper quickly.

Gemini 4 Argon launched on 13 October and Meta's Muse Wonder on 12 October. Ramp's data shows the effective price per million tokens down 41% from its March peak, and its 14 October index found almost no growth in token spend even as usage rose.

Our argument is for designing bespoke AI systems that can switch models, so you benefit from each price drop or performance gain without a rebuild.

This article is for CTOs, CFOs and product owners responsible for AI budgets. It covers the releases of the week of 12 October and what the price changes mean in practice, the cost of lock-in, how to build in model choice, and why model choice should be an ongoing cost-and-quality decision rather than a one-off purchase.

This week's releases and what the price changes mean in practice

Two frontier launches landed within a day of each other, on top of a summer of price cuts. Each one changes the answer to the same budget question: which model gives you the quality you need at the lowest cost?

Gemini 4 Argon: a frontier model at half its standard price, for now

On 13 October, Google introduced Gemini 4 Argon, which it describes as its latest frontier model, available through its API to developers and enterprises. Google says the model "launches today at an introductory price" of $2 per million input tokens and $10 per million output tokens. Its own footnote sets the end date: after the introductory period expires on 30 November 2026, the price rises to $4 and $20. AI Weekly reported the same rates and the same 30 November cut-off.

Google also claims Argon outperforms other frontier models on multi-step tasks. That is Google's own assessment, and the only test that matters for your budget is how Argon performs on your own work.

Muse Wonder: another frontier option from Meta

On 12 October, Meta released Muse Wonder and Muse Wonder Full, which it calls "our most capable models to date", built for agentic work. Muse Wonder is available to developers through the Meta Model API and Muse Code, and Muse Wonder Full is limited to Muse Code subscribers. Meta's announcement gives no prices and no benchmark figures. AI to ROI's 16 October roundup reports that Artificial Analysis lists Muse Wonder at $4 per million input tokens and $20 per million output tokens.

GPT-5.6 Luna: an 80% cut three weeks after launch

OpenAI cut GPT-5.6 Luna prices by roughly 80% on 30 July, three weeks after launching GPT-5.6. Luna fell from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens, according to Implicator. In the same move, GPT-5.6 Terra fell by only 20%, to $2 and $12. Price cuts arrive model by model, not provider by provider.

Ramp: businesses pay less per token and switch more often

Ramp's spend data shows the effective price American businesses pay per million tokens fell about 41% from its March peak, from $1.15 to $0.68, Fortune reported in September. Fortune also reported that since 1 August, OpenAI's effective price had fallen 38%, to $0.48, and Anthropic's 22%, to $0.90.

Ramp's October AI Index, published on 14 October, found two straight months of virtually no growth in token spend, even though token volumes rose. Ramp's lead economist, Ara Kharazian, puts this down mainly to "price declines at the frontier" pulling the blended price per million tokens down, and adds that "the labs are willing to provide concessions in the form of discounts on models from the sticker price."

Businesses are acting on it. Ramp's applied science team reported on 6 October that "the monthly share of firms switching providers, averaged over three months, reached a record 8% in September", and suggested this "could be an early sign that AI models are becoming commodities." These are US figures for businesses on Ramp's platform, so treat them as a signal of direction rather than a census.

What the price changes mean in practice

  • The cheapest capable model keeps changing.Three providers cut prices or launched new frontier models in under three months. A model choice that was right in July may already be the expensive option.
  • Introductory prices end.Gemini 4 Argon's price doubles after 30 November. A budget built on a launch price is built on a promotion, which is why we argued in the LLM pricing trap that prices move in both directions, not only down.
  • Savings only reach systems that can move.A price cut on a model you cannot use is not a saving. Every drop in this list benefits only the businesses whose systems can adopt the cheaper or better model quickly.

The cost of lock-in: rebuilding prompts, integrations and evaluations

The contract is usually not what holds a business to one AI provider. Model access through an API is usually paid per token, and Ramp's economist notes that labs will discount from the sticker price. The lock-in that costs money sits in the system itself, in three places.

Prompts

Prompts are written and tuned against one model's behaviour: how it follows instructions, how it formats answers, how it calls tools and how it returns structured data. Give the same prompt to a different model and the output can change in ways that break the steps that depend on it. Moving models means rewriting and retesting every prompt the system relies on.

Integrations

When an application calls a provider's software development kit directly, that provider's request formats, tool definitions, error handling and rate limits spread through the codebase. Search systems add a deeper dependency. Documents indexed with one provider's embedding model have to be re-embedded before another provider's embedding model can search them, because vectors from different models cannot be compared. The more places the provider appears in your code, the larger the rebuild.

Evaluations

Without a set of tests built on your own work, there is no way to show that a cheaper model is good enough, so the switch stalls on risk. Where tests do exist but were written around one model's habits, they need rebuilding too. Either way, the saving waits until the evidence exists.

The trap

Building directly on one provider's software and discovering how deep the dependency runs only when a better price appears.

The fix

Keep provider-specific code, prompts and evaluations separate from your business logic from the first release.

The contract is easy to leave. The build is not.

How to build in model choice: an abstraction layer, evaluation sets and routing between models

Each of the three components removes one of the lock-in costs above. Together they make a model-agnostic system: one where a model can be changed without changing the system around it.

How each component of model choice removes a cost of AI vendor lock-in
Component Lock-in cost it removes What it lets you do
Abstraction layer Rebuilding integrations and hard-coded prompts Add or replace a provider in one place
Evaluation sets Rebuilding evaluations, and switching on guesswork Prove a new model is good enough before it handles live work
Routing between models Paying frontier prices for routine tasks Send each task to the cheapest model that passes its tests

An abstraction layer

The application calls one internal interface rather than any provider directly. Behind it, a small adapter for each provider translates the request, and prompts are stored as versioned templates for each model rather than written into the code. Adding a provider then means writing one adapter and one set of prompt templates.

Keep the layer thin. Flattening every provider to the features they all share delivers the weakest version of each model, so provider-specific features you pay for should stay available behind the same interface. We set out that trade-off in our piece on LLM pricing. Off-the-shelf gateways can be a starting point: in August, Ramp launched Router, which TechCrunch reported "lets users and companies use and switch between various large language models through an API". Your own business rules still belong in your own code.

Evaluation sets

An evaluation set is a collection of real examples from your own workflow, each with an agreed correct outcome. Scoring is automatic where the answer is clear-cut, with a person reviewing the cases that need judgement. Every candidate model runs through the set before it handles live work, and the set runs again whenever a provider updates a model. Models will come and go. The evaluation set is the asset that lasts.

Routing between models

LLM routing sends each task to the model that passes its evaluation set at the lowest cost. Routine work such as classification and data extraction can go to small, cheap models, harder reasoning can go to a frontier model, and a second provider can take over when the first is unavailable. Ramp's Router offers routing strategies on the same principle, such as sending only the harder problems to more expensive models. Track the cost per completed task for each model, not just its price per token.

Routing also widens the choice of LLM providers you can use, including open-weight and Chinese models, as we covered in our piece on AI model alternatives for business, and models running on your own servers, which we looked at in on-premises AI.

Treating model choice as an ongoing cost-and-quality decision rather than a one-off purchase

The market now moves faster than most procurement cycles. Ramp counts at least nine model launch announcements from Anthropic and OpenAI alone since June 2026, and the share of firms switching providers each month is at a record high. Choosing a model once, at the start of a project, fixes a decision that the market will reopen within weeks.

What an ongoing decision looks like

  • A named ownerOne person owns model choice for each AI system, with a cost target and a quality target.
  • Clear triggers for reviewA new model release, a price change, an introductory rate ending, a model being retired, or a drop in quality in production.
  • The same test every timeEach trigger means re-running the evaluation set and comparing cost per completed task and quality for each candidate model.
  • A recorded decisionSwitch, split the work between models, or stay, and write down why, so the next review starts from evidence.
  • Commercial terms that can changeAI to ROI's 16 October roundup advises renegotiating often, avoiding commitments sized for last quarter's prices and testing cheaper models on routine tasks.

The trap

Treating the model as a purchase made once at procurement, then building everything around it.

The fix

Treat the model as a replaceable component, with an owner, a test and a standing review.

This is how we approach the systems we build. Our LLM and generative AI development work includes evaluation and accuracy testing, and after launch we move clients to better-value models as they appear.

Why your AI system should not be tied to one provider

Leading models are improving and getting cheaper quickly, and the gains arrive model by model, from different providers, often with little notice. A system tied to one provider captures only the gains that provider chooses to offer, and only on its timetable.

AI vendor lock-in is a design choice, even when nobody makes it deliberately. Build the abstraction layer, the evaluation sets and the routing in from the start, and every price drop or performance gain becomes something you can test and adopt, not something you watch competitors take.

Can your AI system switch models without a rebuild?

If moving to a cheaper or better model would mean rewriting prompts, integrations and tests, your system is paying for vendor lock-in.

At Imobisoft, we design and build AI solutions around the systems and data you already have, with the model treated as a component you can change, not a commitment you are stuck with.

Talk to us about model choice

Frequently asked questions

What is AI vendor lock-in?

AI vendor lock-in is when an AI system is built so closely around one provider's models that switching would mean rebuilding its prompts, integrations and evaluations. It stops a business from adopting cheaper or better models when they appear.

Are AI model prices falling?

On average, yes. Ramp's data shows the effective price US businesses pay per million tokens fell 41% from its March 2026 peak to $0.68. Individual prices can still rise: Gemini 4 Argon's introductory price doubles after 30 November.

How much did OpenAI cut GPT-5.6 Luna prices?

By about 80% on 30 July 2026, three weeks after launch. Luna fell to $0.20 per million input tokens and $1.20 per million output tokens.

How do you avoid AI vendor lock-in?

Put an abstraction layer between your application and the model, build an evaluation set from your own work, and route each task to the cheapest model that passes it. Then review the choice whenever prices or models change.

What is LLM routing?

LLM routing sends each task to the model best suited to it on cost and quality, such as a small model for routine extraction and a frontier model for harder reasoning, with a second provider ready if the first is unavailable.

Does a model-agnostic design lower quality?

Not if the abstraction layer stays thin. Abstract the call to the model and the evaluation, keep access to the provider features you pay for, and test every model against your own evaluation set before it goes live.

Sources

Have a project in mind? Let’s get to work.

Let’s chat about how we can help you. Fill in the details and we’ll get back to you as soon we can.