600+ AI models. How do you choose?

Anthropic released Claude Fable 5.1 on September 1st. Meta released Muse Spark 1.3 on the 2nd. OpenAI released GPT-6 Astra on the 3rd. Three frontier model launches in three days.

Zoom out and the problem gets silly quite quickly. Artificial Analysis now tracks performance data across 643 models. Its benchmarks span reasoning, coding, agentic work, long context, professional document work, tool use, visual reasoning and more.

So which one should you use?

That question used to be relatively easy. Pick the best model you could afford and get on with it. It isn't anymore.

Models now come in families. OpenAI alone has GPT-5.6 Luna, Terra and Sol, then GPT-6 Astra above them. Within those models you can choose how much compute they spend thinking: none, low, medium, high, xhigh, max. Anthropic has adaptive reasoning and its own effort levels. Gemini has reasoning settings too.

So the thing you are choosing is no longer really a model.

It is a model × reasoning level × workload.

And there are far too many combinations to pick one because somebody on X said it was good.

Fortunately, this is not a new kind of decision. Economists and operators have been dealing with versions of it forever.

Take two things you care about. Plot them against each other. A choice is Pareto efficient when you cannot improve one dimension without giving something up on the other.

For AI models, the obvious starting pair is capability, up, and cost, across. You want to be as far up and to the left as possible.

If another model is both cheaper and more capable than yours, there isn't much to debate. Your model is dominated. If moving to a better model requires paying more, you have reached the frontier. Now there is an actual decision to make.

The useful question isn't “what is the best model?” It is “where does paying more stop buying enough additional capability to matter?”

That produces a much more useful picture of the market.

The exact dots will keep moving, probably before we finish writing this, but the shape is what matters.

Notice the recent Spark 1.3 for example. On Artificial Analysis' previous index version, its publicly available xhigh configuration scored alongside some of the strongest frontier models while its blended token price was just about $0.78 per million tokens. GPT-6 Astra, not far on capability, was around $7.70 on the same blended-price methodology. Fable 5.1 was around $7.17. Gemini 3.8 Flash sat lower on capability but natively multimodal and at roughly $0.58.

That doesn't mean Spark is “better than Astra”.

It means the burden of proof changes. If two models produce similar results on the work you care about and one costs dramatically less, the expensive model needs to earn the difference. Which is how we would make almost any other procurement decision.

The chart is useful. It also lies a little. “Capability” is not actually one thing. And, more practically, models do not all perceive the same world.

Some take text. Some can look at images. Some understand video. Some can listen to audio. Some can work directly across all of them. That matters enormously if you are building an actual product rather than answering benchmark questions.

Here is the rough landscape for some of the models we are currently most interested in:

Gemini, Spark and Qwen are natively multimodal: text, images, video, audio and documents in a single call. The rest are mostly text-and-image; GLM-5.3 is text-only. A cheaper, higher-capability model is only cheap if it can actually see the inputs your product uses.

And suddenly that nice two-dimensional frontier needs some context.

A model that cannot process video is not “cheap” for a video product. It is unusable.

A model with slightly lower general intelligence might be dramatically better for a workflow that needs to reason over an hour of footage. The third dimension is really fitness for the job.

Public benchmarks are useful for narrowing the field. We don't think they should make the final decision though. The better approach we have found when choosing models for a real business is to build a small evaluation around the work the business actually does.

Take the recurring tasks. Make them representative. Give every candidate model exactly the same work. Measure the result and the real cost of producing it. Here is an example step-by-step process:

• ▪Define the jobs. Twenty to a hundred representative tasks is often enough to expose meaningful differences: interpret this strategy problem, reconcile these numbers, understand this image, fix this codebase, work autonomously for an hour.
• ▪Define what good looks like. Objective correctness where possible; a tight rubric where it isn't.
• ▪Run model configurations, not just models. Luna-max may compete with a more expensive model at medium. Gemini-high is a different economic proposition from Gemini-low. Treat them as separate points.
• ▪Measure the whole task. Tokens, reasoning tokens, retries, tool calls, latency and failure rate. Token price alone is not cost.
• ▪Plot capability against cost. Then draw the Pareto frontier.
• ▪Apply the hard constraints. Modalities, data residency, latency, cloud environment, licensing and operational reliability.

The result will be different for every business.

A company extracting information from photographs of shipping documents has one frontier. A company building an AI strategy analyst has another. A company running coding agents continuously against a large repository has another again.

There is rarely one “best model”. There are a handful of efficient models, each sitting on the frontier for a particular job.

Most real jobs are not one task. A document pipeline is extraction, then reasoning over what was extracted, then writing the summary a human will actually read. An agentic workflow is planning, tool use, checking the output and formatting it. Each step is its own task, and each task has its own frontier.

So for most jobs today, the question is nit “which model do we use?” but “which model does each step use?”. A $0.58 model handles the 90% of steps where it sits on the frontier; the expensive frontier model earns its tokens only on the few steps that genuinely need it. The blended cost of the job drops well below any single-model answer, and capability on the hard steps actually goes up.

You don't pick a model for the job. You pick a model per task, and the job becomes a routing problem.

This is why “keep the routing separable from the product” is a critical principle. Routing is the mechanism that turns a set of per-task frontiers into one efficient job.

That being said, we do have slightly unfashionable bias that we are conscious of: "We have a lot of respect for Gemini."

Google established serious multimodality early. Gemini 3.8 Flash is fast, broadly capable, relatively inexpensive, handles text, images, audio and video natively, and sits behind infrastructure from a hyperscaler we are comfortable putting serious production workloads on. That combination matters.

AWS Bedrock is doing a great job bringing the other models closer to that combination, especially open-weight models. They have become much easier to deploy without turning your team into an inference infrastructure company.

So we try not to award points for familiarity. Or country of origin. Or whose launch video was better. Or which model happens to power the chatbot we personally use every day.

We use Gemini where Gemini sits on the frontier. Spark where Spark does. GLM, Qwen, OpenAI, Claude or Grok when one of them does. That sounds obvious written down. In practice, it is surprisingly rare.

There is one final inconvenience. This isn't a decision you make once.

Anthropic, Meta and OpenAI just demonstrated that rather neatly by releasing major models on three consecutive days. The frontier moved three times before the working week was over.

So the model layer should be treated as something replaceable.

• ▪Keep your evaluation set.
• ▪Keep the routing separable from the product.
• ▪Rerun the candidates periodically.
• ▪Switch when something meaningfully crosses the frontier.

The winner is not the business that picks the best model today. It is the business that can keep picking the right ones over and over again.

In summary: There is no best AI model. There is a Pareto frontier for the job you need done that keeps moving. And, frankly, that is a much easier problem to solve.