The "which AI should we use?" question lands on my desk every few weeks. The answer is rarely one model, and in 2026 it's different from what it was twelve months ago. This is how I think about it after building production systems on all three families.
The short version
All three major providers, Anthropic (Claude), Google (Gemini) and OpenAI (GPT), now have enterprise-grade models that handle most business use cases competently. The differences are real, but they're at the margins.
The choice usually comes down to four things: the task, the ecosystem you're already in, your data residency and compliance requirements, and the contract terms you can get.
Where each is strongest in 2026
Claude (Anthropic)
Claude is my default for serious reasoning, long-form analysis and anything that needs a model to follow detailed instructions reliably. Claude 4 Opus and Sonnet are especially good at long-document analysis, where the 1M-token window earns its keep. They're strong at writing, reviewing and refactoring code, and they follow complex multi-step instructions without drifting. Claude is also more likely than the others to say "I'm not sure" instead of making something up.
It's weaker on real-time data access, and its tool ecosystem is still catching up with the others. It handles images and other media, but it isn't the most advanced of the three there.
Gemini (Google)
Gemini is the strongest of the three for multimodal work: understanding images, analysing video and handling large documents with mixed content. It's also the obvious choice if you're deep in Google Workspace and Google Cloud, since it plugs into Docs, Sheets and Gmail and Vertex AI gives GCP teams solid MLOps tooling. At the lower end of the range its cost per token is very competitive.
I find it less consistent than Claude on complex reasoning. The product has also moved around a lot (Bard, then Gemini, with model names changing regularly), so it's worth keeping an eye on.
GPT (OpenAI)
OpenAI still has the broadest ecosystem, with the most third-party integrations, the most existing libraries and the biggest community of practitioners. Its function calling has historically been the most polished, its real-time and voice modes are strong, and it offers a wide spread of models at different price and performance points. Through Azure OpenAI it also fits neatly into a Microsoft 365 stack.
It's less differentiated on raw capability than it used to be. The lead it once had across the board has narrowed a lot as Claude and Gemini caught up.
What decides it for an enterprise
For an organisation choosing a provider, model capability usually matters less than the following.
Data residency and contract terms
Where is your data processed, where can the provider promise it stays, and what indemnity do they offer? For New Zealand and Australian organisations this often points to Azure OpenAI (with Australian data regions), or to Vertex AI or Anthropic on AWS in similar regions. The picture changes every quarter as providers open new regions.
Your existing ecosystem
If you're a Microsoft 365 shop, Azure OpenAI fits naturally and your enterprise agreement probably covers procurement. Google Workspace shops will find Vertex AI the easiest path. Anthropic is available through both Amazon Bedrock and Google Vertex AI, so it can fit either way.
Procurement and support
Large organisations care about volume discounts, dedicated support, SLAs and a way to escalate when something breaks. All three offer this at enterprise scale, with varying depth. With $1M+ a year in spend, expect to negotiate.
Avoiding lock-in
For most enterprises the best architectural choice is to put the model behind your own API layer so you can switch providers when prices or capabilities change. It costs little to do, and it's worth a lot when a contract negotiation goes sideways.
What I recommend
For most of the organisations I work with, the answer is two providers rather than one. The primary one matches your existing ecosystem, so Microsoft shops go with Azure OpenAI and Google shops with Vertex. The secondary covers what the primary is weaker at, and doubles as negotiating leverage. Claude is the most common second choice, because it's strong where the others are weaker.
Setup costs a little more. That flexibility has paid off every time a provider changed its pricing, capabilities or terms over the past two years.
The trap
The trap is tying your business logic to one provider's quirks: its tool-calling format, its favourite prompt patterns, its bespoke features. Use whatever helps, but keep a thin abstraction in front so you can switch. Teams that did this barely notice when prices move or a new flagship lands. Teams that didn't tend to spend the next quarter migrating.
How to decide
In 2026 the answer is rarely "use the best model". Each provider is best at different things, and your tasks, your current stack and how much lock-in you'll tolerate usually narrow the field for you. Pick a primary and a secondary, and put both behind an abstraction you control.
If you're choosing a provider for a specific project, or setting up a multi-provider strategy that doesn't get out of hand, I'd be glad to help.
