We might be looking at the end of cheap AI systems era, and those who are locked into a single vendor, will most likely find the bill is going to be more than what they signed up for.
Here is a thought experiment for anyone building something serious on AI right now. What happens to your business if your AI provider doubles its prices? Is that a line item and a sprint of engineering work, or is it a genuine problem nobody budgeted for?
Many soon-to-be mission-critical systems are being wired into single-vendor ecosystems as you read this. One company’s models, one company’s tooling, one company’s pricing. Early on that feels easy and safe, and honestly, it is.
The catch arrives later… and by the time it does, you’re not really a customer, you’re a tenant at best, hostage at worst.
For the first two years of the AI boom, the pricing story only moved in one direction: down, fast, for everything. The last twelve months broke that pattern, and they are the twelve months worth planning around. Two things happened at once: the meters on the serious tiers turned upward, and the systems being built became enormously hungrier for what those meters measure.
Start with the side of the equation nobody budgets for: consumption. The systems being built today are not the chat windows of two years ago. Agentic workflows loop, retry, and call tools dozens of times per task. Reasoning models think in billable tokens before they answer. Retrieval pipelines fill enormous context windows on every request. Each of these patterns became standard practice for a rational reason: tokens were cheap, so architecture stopped optimising on them. Industry analyses through 2025 and 2026 document the result: unit prices fell while token consumption exploded, and total enterprise AI spend rose sharply through the steepest price cuts in the history of computing.
That inversion changes what a price movement means. When your workload has grown tenfold or a hundredfold, your sensitivity to the meter rate has grown by the same multiple, and a rate change that would once have been a rounding error now moves the whole invoice. Cheap tokens taught everyone to build token-hungry systems, and the invoices now quietly assume the cheapness is permanent. That assumption is the exposure, and it is exactly the assumption the past twelve months have started to test.
“Cheap tokens taught everyone to build token-hungry systems, and the invoices now quietly assume the cheapness is permanent.”
Now look at what the meters themselves did over the same twelve months.
OpenAI’s flagship input price quadrupled in eight months, from US$1.25 per million tokens at GPT-5’s launch in August 2025 to US$5 for GPT-5.5 in April 2026, with output at US$30. Google’s workhorse Flash line went from US$0.10 and US$0.40 per million tokens to US$1.50 and US$9 across two generations. Anthropic introduced a new premium tier priced at double its previous top model, and its newest models use a redesigned tokeniser that produces up to roughly a third more tokens for the same text, which quietly moves the meter itself. All three providers now also charge premiums for the things production systems need most: priority processing, data residency, and speed.
To be fair to the vendors, the newer models are genuinely more capable, and the budget end keeps getting cheaper: OpenAI added a cheaper production tier as recently as this month, and DeepSeek made a 75% price cut permanent in May. But the direction on the serious tiers is unmistakable, and monetisation is visibly shifting from the headline token price to the tiers, multipliers, and premiums that production workloads cannot avoid. Now multiply the two halves of the story together. A rising meter under a workload that has grown tenfold is not a small problem, it is the whole budget conversation. The floor is a great place to experiment. It is not where your production system lives, and it is not where your invoice is generated.
“A rising meter under a workload that has grown tenfold is not a small problem.”
There is also a reason to expect the turn to continue, and it sits on the labs’ own books. OpenAI expects to spend roughly US$22 billion this year against US$13 billion in sales, about $1.69 out for every dollar in, with losses projected to deepen through 2028 and profitability pushed to the end of the decade. HSBC, modelling generous growth assumptions, still finds a US$207 billion funding gap between OpenAI’s revenue outlook and the compute commitments it has signed, which run to hundreds of billions across Oracle, Microsoft, and Amazon. The counterpoint is real and worth stating: Anthropic expects its first operating profit in the second quarter of 2026, which suggests the enterprise-weighted version of this business can pay its way, though the figure excludes stock compensation and the company has itself trimmed its margin projections as inference costs ran hotter than expected.
What changes next is who the labs answer to. Both are preparing to face public markets: OpenAI has filed confidentially and is reported to be targeting a listing as early as September 2026 at a valuation above a trillion dollars, with Anthropic said to be eyeing October. The extraordinary patience of the past three years was a feature of private mega-investors making a decade-long bet. Public shareholders holding a trillion-dollar stock through years of projected losses are a different constituency, and the pressure they apply has a predictable direction: harvest.
So here is the part of this article that is a forecast rather than a record, stated as one. We think the era of investing ahead of returns gives way to a monetisation phase, and that it will not look like a headline price rise. It will look like what the past year already looks like: premium tiers, paid priority, residency surcharges, meter changes, and successor models that cost more than the ones they replace. The bill for the build-out eventually arrives, and it is presented first to the customers least able to leave.
One more thing about how these increases reach you, because it is easy to miss: vendors rarely raise the price of the model you are on. They retire it, and the successor costs more. Google’s workhorse line was not repriced; it was replaced. The same lever moves whole platforms: OpenAI is winding down its entire fine-tuning platform this year, and its Assistants API shuts down next month, so the surfaces you build on are as impermanent as the price list. A model swap costs evaluation time on any architecture, including ours; that part is symmetric. What is not symmetric is choice. On a single-vendor system, the successor and its price are mandatory. On a model-agnostic one, the successor has to compete for the job.
“a US$207 billion funding gap between OpenAI’s revenue outlook and the compute commitments it has signed”
“The bill for the build-out eventually arrives, and it is presented first to the customers least able to leave.”
It is tempting to make lock-in sound deeply technical, and there is technical residue: fine-tuned models that cannot be exported, embedding databases that must be regenerated on a model change, integrations written against one vendor’s features. All real, all billable. But let’s be honest about what actually keeping most organisations where they are.
Somebody approved a vendor, the tools work well enough, and from that day the single-vendor stack is simply what the company uses. Every individual who might propose something more flexible is volunteering for personal risk to save the organisation money later, and nobody has ever been fired for renewing the incumbent.
That is the true mechanism of lock-in: not chains, but a chair that is comfortable enough. And the trouble with defaults is the timing of them. The default is chosen at the moment the vendor is at its cheapest and friendliest, and it is lived with at the moment the vendor no longer needs to be. Flexibility is not something you can decide to have later, because later the decision belongs to whoever owns the stack you settled into.
“Somebody approved a vendor, the tools work well enough, and from that day the single-vendor stack is simply what the company uses.”
If the AI market ever does consolidate to a point where a dominant vendor faces investor pressure to harvest its installed base, there is a live demonstration of what that looks like, and it is worth knowing even though it comes from the neighbouring aisle.
After Broadcom acquired VMware, it ended perpetual licences, collapsed the product catalogue into subscription bundles, and repriced the installed base. The results are now working through the courts. Tesco’s £100 million High Court claim alleges increases of 237% on licences it describes as superfluous; the European cloud association CISPE reported member increases of up to 1,500%; AT&T sued, describing demands for a king’s ransom, before settling, and the dispute has not cooled: in mid-July 2026, five European cloud associations were still urging the EU to impose interim measures against Broadcom. Tesco is now migrating some 40,000 server workloads off VMware, which is the quantified cost of unwinding a twenty-year dependency in the middle of a dispute. And commercially, the strategy is working, which is exactly why it is the relevant precedent: when switching costs are high enough, a vendor can reprice dramatically and keep most of the customers anyway.
Nothing in the AI pricing record suggests this is happening today; competition among well-funded providers is doing its job. The Broadcom case matters for a different reason. It establishes what the option to leave is worth, and that the time to secure that option is while you still have it, cheaply, at design time, not later at migration time.
“when switching costs are high enough, a vendor can reprice dramatically and keep most of the customers anyway.”
The protection that follows from all of this is architectural. Keep the orchestration layer, the part of the system that decides what happens, retrieves your knowledge, calls tools, and assembles answers, in your control and model-agnostic, built on open frameworks rather than any one vendor’s managed runtime. Frameworks such as LangChain and LangGraph exist for exactly this: the model becomes a configuration choice behind an interface you own, and observability platforms like LangSmith let you watch every call closely enough to know whether a cheaper model is actually holding up. In the systems we build, a model swap is a re-pointing exercise measured in days, not a re-architecture measured in months. It is also where the money has quietly moved: cost analyses through 2026 consistently find that the model invoice is only a minority of what a production AI system costs, with the larger share sitting in the orchestration, retrieval, evaluation, and observability around it. The layer that carries most of the cost and all of the control is a strange one to hand to somebody else.
Owning the orchestration layer also unlocks the most immediately bankable benefit: routing by task. Not every query needs the most expensive model. Send the simple, high-volume work to fast, cheap models and reserve the frontier for the questions that genuinely need it. The academic results here are striking, with the FrugalGPT research matching top-model quality at up to 98% lower cost in benchmark settings, and practitioner accounts through 2026 consistently report savings of 40 to 85% in production. Treat the specific percentages as self-reported, because audited enterprise case studies remain scarce, but the direction is beyond dispute, and it compounds with the capability trend: for most production tasks, very capable models now exist at a tenth to a hundredth of frontier prices, while the frontier keeps its lead mainly on the hardest agentic work. A router lets you buy each tier of intelligence at its market price. A single-vendor system buys everything at one price, and it is never the low one.
Two honest caveats, because the model-agnostic argument is often oversold. First, abstraction is not free. Frameworks add their own complexity, routing thresholds drift and need tending, and the layer you insert is itself a dependency: in March 2026, compromised versions of a popular open-source model gateway were briefly published to the Python package index carrying a credential stealer, a sharp reminder that the swap layer concentrates your API keys and deserves the same security discipline as anything else in the critical path. Second, abstraction does not eliminate lock-in so much as re-seat it, moving it from the model vendor to a layer you own and understand. That trade is still decisively worth making, because it converts unpredictable forced migrations on a vendor’s calendar into predictable integration work on yours. But it is an engineering discipline, not a product you switch on, and it rewards builders who have done it more than once.
“In the systems we build, a model swap is a re-pointing exercise measured in days, not a re-architecture measured in months.”
“Not every query needs the most expensive model. Send the simple, high-volume work to fast, cheap models and reserve the frontier for the questions that genuinely need it.”
If this all sounds like engineering paranoia, it may be worth knowing that the people whose job is systemic risk arrived at the same conclusion before most of the industry did.
The Reserve Bank of New Zealand flagged concentration risk from dependence on a few third-party AI providers in its May 2025 Financial Stability Report, and its November 2025 survey found that 22% of New Zealand financial entities lack an established exit strategy for critical third-party arrangements. Across the Tasman, the prudential regulator APRA wrote to industry in April 2026 noting that some regulated entities were heavily dependent on a single provider across multiple AI use cases, with few having tested exit and substitution strategies. And the Australian government has gone a step further than commentary: its GovAI platform for the public service is being built as a model brokerage, offering agencies vetted access to multiple models behind one governed layer, which is the model-agnostic architecture implemented as national policy.
When central banks and regulators start asking whether you could leave your AI vendor, tested exit paths stop being a nice-to-have and start being what good governance looks like.
“22% of New Zealand financial entities lack an established exit strategy for critical third-party arrangements.”
Pull it together and the picture is clear-eyed rather than alarmist. The systems everyone is building are hungrier than ever, the flagship meters have turned upward, and the labs supplying them are carrying vast losses toward public markets whose shareholders will want a return.
The performance gap between the big names and the challengers keeps narrowing on most work. The price gap between them keeps widening. An architecture that lets you pick and mix is the only one positioned to collect on both trends, and the habit of defaulting to one approved vendor is the only thing standing between most organisations and that position.
So the question to ask anyone proposing to build your AI system, ours included, is not which model they will use. Models are the most replaceable part of the whole stack, and the answer will be different in six months anyway. The question is: who owns the orchestration layer, and what exactly happens when we want to change models, or providers, or both? Ask them to walk you through the swap. What gets re-pointed, what gets rebuilt, what gets re-tested, and how long does it take?
A builder with a real answer will describe a configuration change, an evaluation run, and a rollout, because the system was designed for that day from the start. A builder without one will tell you the model they have chosen is so good you will never want to leave. Which is, when you think about it, exactly what a landlord would say.
What is AI vendor lock-in?
The accumulated cost of leaving an AI provider once your systems depend on it: fine-tuned models that cannot be exported, embeddings that must be regenerated, integrations built on proprietary features, and team expertise tuned to one platform. The model itself is usually the easiest part to replace; the structure around it, and the organisational habit of defaulting to one approved vendor, are what lock you in.
Will AI API prices go up?
Prices for a fixed level of capability have fallen dramatically since 2023 and may keep falling. But the tiers production systems rely on turned upward from late 2025, providers now charge premiums for priority, residency, and speed, and total spend rises with usage even when unit prices fall. Budget for the system, not the token price. Our read: as the leading labs move toward public listings while carrying heavy losses, pricing pressure grows from here, and it will land first on the customers least able to switch.
What is a model-agnostic AI architecture?
A system where the orchestration layer, the part that retrieves knowledge, applies rules, and assembles answers, is independent of any one AI provider, typically built on open frameworks such as LangChain or LangGraph. The model becomes a swappable component, so changing providers is a configuration and testing exercise rather than a rebuild.
What is model routing and how much does it save?
Routing sends each task to the cheapest model that can do it well, reserving expensive frontier models for genuinely hard work. Peer-reviewed research has shown up to 98% cost reduction in benchmark settings, and practitioners commonly report 40 to 85% in production. It requires a model-agnostic architecture and ongoing evaluation to do safely.
Linkki builds BrandRAG systems: retrieval-first AI on a model-agnostic orchestration layer built with LangChain, LangGraph, and LangSmith, where models are components you choose, not foundations you are stuck on.
Or give us a call on +64 21 280 2773
Or give us a call on +64 21 280 2773