TL;DR
Anthropic, OpenAI, and Google all released new frontier or flagship models in September 2026 alone. Any model choice you make today is a snapshot.
In June 2026, a US export control directive led Anthropic to switch off its two newest models for every customer worldwide for 19 days. Systems that could fall back to another model kept running.
Single-vendor risk goes beyond regulation. Prices move, platform policies change, and models get retired.
Match models to the work: the most capable frontier models for high-stakes reasoning, cheaper models for volume, specialist models for images, and local open-weight models for confidential or offline work.
Put a routing layer between your business and the models. That can be a gateway such as Vercel AI Gateway, an aggregator such as Abacus.AI, or your own orchestrator.
Own your prompts, workflows, and data. They are the parts of your AI operating system that should outlive any model.
What model-agnostic means in practice
A model-agnostic setup keeps everything that makes your AI useful, meaning your instructions, workflows, business context, and data, in a layer you own. The model is a component that plugs into that layer. Switching providers becomes a configuration change rather than a rebuild.
The alternative is a system where prompts are tuned to one model's quirks, workflows live inside one vendor's product, and company knowledge sits in that vendor's storage. That can work well for a while. It also means a decision made by someone else, whether a price change, a policy change, or a government order, lands directly on your operations.
What happened when two frontier models were switched off
On 9 June 2026, Anthropic released Claude Fable 5 and Claude Mythos 5, two versions of the same underlying frontier model, meaning the most capable class of model a lab offers. Three days later, the US government, citing national security authorities, issued an export control directive suspending access to both models by any foreign national, inside or outside the United States, including Anthropic's own foreign-national staff. Because the order took effect immediately, Anthropic disabled both models for all customers, US customers included. Its other models were not affected.
Anthropic complied and said publicly that it disagreed with the basis for the order. The Department of Commerce lifted the controls on 30 June, and Fable 5 returned globally on 1 July, 19 days after it went dark. Mythos 5 came back only for a set of US organisations approved by the government.
I will leave the merits of the decision to others. The operational lesson is what matters for a leadership team. Companies that had built workflows around Fable 5 with no fallback stopped. Companies whose systems could route the same request to another model carried on. Running the model through a second cloud would not have helped either, since access through AWS, Google Cloud, and Microsoft Foundry also had to be restored afterwards. The restriction followed the model wherever it ran.
Nobody I know of predicted that event in May, and the next disruption will be just as hard to see coming.
Models change faster than procurement cycles
Even without government action, the ground moves quickly. In September 2026 alone, Anthropic released Claude Fable 5.1, OpenAI released GPT-6 Astra, Google released Gemini 3.8 Flash, and on 22 September Anthropic shipped Claude Opus 5.5 while OpenAI shipped the cheaper GPT-6 Sol and GPT-6 Luna. Published rankings of the best coding model disagree with each other depending on the benchmark, and one widely read ranking reported that its answer changed three times in September.
Models also retire. The original Nano Banana image model, which went viral in August 2025, is now listed by Google as its legacy model, with the faster and cheaper Nano Banana 2 Lite at the entry tier. Any marketing team whose image workflow was hard-wired to the original now has a migration project.
Prices spread wide as well. At reported list prices, output from Fable 5.1 costs around $50 per million tokens (a token is roughly three-quarters of a word), while GPT-6 Luna costs around $0.50. That is a 100-fold difference. Sending every request to the most capable model is the most expensive way to run AI, and for a lot of work it produces no better result.
Other ways a single provider can move under you
Platforms change their terms. From 15 January 2026, Meta barred general-purpose AI assistants from the WhatsApp Business API, removing a distribution channel several AI companies had built on. Businesses using AI for defined tasks such as customer support could continue, but it showed how quickly a platform owner can redraw the line.
Regulation moves too. The EU's Digital Omnibus, in force since 27 July 2026, pushed back the AI Act's high-risk obligations for standalone systems, which include recruitment tools, from August 2026 to December 2027. That gives HR teams in Europe more time, and it also shows the rules are still being written. A provider that suits your compliance position today may not suit it next year.
Then there is ordinary commercial risk. Providers change pricing, rate limits, data retention terms, and which models come with which plan. None of this is unusual. It lands harder when you have no alternative.
How to match models to the work
Being model-agnostic comes down to choosing deliberately for each type of work and keeping the option to change. This is roughly how I break it down with clients:
High-stakes reasoning. Strategy analysis, financial review, and architecture decisions justify a frontier model such as Fable 5.1, GPT-6 Astra, or Opus 5.5. A second-best answer here is expensive.
Everyday knowledge work. Drafting, summarising, research, and most internal assistants run well on mid-tier models at a fraction of the cost.
High-volume, low-risk tasks. Editing an email, classifying tickets, or tagging records belongs on the smallest model that does the job reliably. Frontier prices to fix a comma make little sense.
Specialist output. I still use Google's Nano Banana models for images, which Claude does not generate, and I prefer Claude for writing code. Each vendor leads somewhere, and those areas shift.
Confidential or offline work. Anything that should never leave your environment goes to a model you run yourself.
A practice now common among engineering teams is worth copying: have a model from a different vendor review important output. Models trained in similar ways tend to miss similar things, so a reviewer from another family catches more.
The routing layer: gateways, aggregators, and orchestrators
The practical way to stay flexible is a routing layer between your business systems and the models.
AI gateways are services that sit between your software and the model providers, giving developers one API that reaches many of them. Vercel's AI Gateway passes through provider list prices without a markup and supports model fallbacks, so if the first model fails or is unavailable, the request moves to the next one on your list. Its own documentation example sends a request to Claude Fable 5 with Claude Opus 5 and a Gemini model as fallbacks, which is close to the exact setup that would have kept a system running through June. OpenRouter and the open-source LiteLLM do similar work.
Aggregators package this for end users. Abacus.AI's ChatLLM puts models from OpenAI, Anthropic, Google, and others behind one interface, and its RouteLLM API picks a model for each prompt by default or lets you name one. For a team that wants several models without juggling subscriptions, that is a sensible starting point.
Both come with a caveat. A gateway or aggregator is also a vendor. If your prompts, workflows, and knowledge live inside it, you have moved the lock-in one layer up. Keep your instructions, business context, and data in systems you control, and treat the router as replaceable too.
It is also the problem my colleagues and I at Praxes are working on: a single interface that looks like the chat tools teams already use, with a layer underneath that picks a model for each request based on the task and the cost. Many executives I work with do not want their teams switching between four tools, and they want everything drawing on the same company knowledge. Routing handles both.
Local models for confidentiality and offline work
Some work should not go to an external provider at all: board papers, employee records, client data under strict contracts, or anything that must run without an internet connection. Open-weight models, which you download and run on your own hardware, now cover a lot of this.
OpenAI's gpt-oss models are released under the Apache 2.0 licence, and the smaller one runs on a machine with 16GB of memory. Qwen, Mistral, DeepSeek, and Gemma models are also widely used, and tools such as Ollama make running them close to a single install.
Local models come with trade-offs. They still trail the best hosted models on the hardest tasks, and someone has to own the hardware, updates, and security. I usually recommend a hybrid: local models for the sensitive slice of work, hosted models behind a router for everything else.
What you should ask your team to do this week
List every AI tool and integration in use, and note which model sits underneath each one.
For each critical workflow, ask what happens if that model disappears for three weeks. If the answer is that the work stops, start there.
Move prompts, instructions, and business context out of vendor products and into documents and systems you own.
Put a routing layer in front of new builds, with at least one fallback model from a different provider.
Decide which data must never leave your environment, and test a local model for that work.
Review model choices every quarter.
Most of this is a design decision made early, which costs far less than a rebuild made in a hurry. If you want a structured way to check where your organisation stands, our guide to running an AI readiness assessment walks through it, and if you would rather have a model-agnostic system designed and built for you, that is what our AI-native operating systems work does.
I have not worked all of this out, and nobody has. What I am confident about is that it is too early in the move to human and AI work to bet the business on one provider staying ahead on quality and price, and staying available to you at all.
Frequently asked questions
What does model-agnostic mean in AI?
A model-agnostic system keeps its instructions, workflows, and data independent of any single AI provider, so the underlying model can be swapped with a configuration change. The model becomes a replaceable component rather than the foundation.
How do you avoid vendor lock-in with AI?
Keep your prompts, business context, and data in systems you own, and put a routing layer such as an AI gateway between your applications and the models. Configure at least one fallback model from a different provider for every critical workflow.
Which AI model is best for business?
No single model is best for every task, and the leader changes often. Use frontier models for high-stakes reasoning, cheaper models for high-volume work, specialist models for images, and local models for confidential data.
How do I pick the right AI model for a task?
Start with the cost of a wrong answer. If a mistake is expensive, use a frontier model. If the task is routine and repeated thousands of times, test the smallest model that does it reliably. Then check data sensitivity, which may rule out hosted models entirely.
Should my company run AI models locally?
For confidential or offline work, often yes. Open-weight models such as OpenAI's gpt-oss can run on your own hardware, so data never leaves your environment. Most companies do best with a hybrid: local models for sensitive work, hosted models for the rest.
