How Eazy Chooses Which AI Model to Use
Eazy doesn't bet the product on a single AI lab. Every step of turning your writing into a finished deck is matched to the model best suited for it, judged against real criteria instead of brand loyalty. Here is the selection philosophy behind that — in plain terms, with no invented routing tables.
Updated Jul 22, 2026·Published Jul 22, 2026
Eazy evaluates frontier models from labs including Anthropic, Google, and OpenAI against four criteria — content quality, design capability, speed, and token cost — and routes the best fit to each step of building a deck. The result is a provider-flexible engine built on frontier models plus a dedicated image model, not a single-vendor bet.
The Four Things We Judge Every Model On
Eazy judges candidate models on four axes: content quality, design capability, velocity, and token cost. No model leads on all four at once, so using one model everywhere means accepting a weak spot somewhere. Matching the axis that matters most to each step is the reason to evaluate models instead of committing to one.
It would be simpler to pick the single 'smartest' model available and route every step through it. But smartest is not one thing. A model that reasons brilliantly about long documents is not automatically the one you want handling a fast, repetitive step, and a model that is fast and cheap is not automatically the one you want writing your opening argument.
Content quality asks how well a model writes, structures an argument, and holds onto nuance across a long document. Design capability asks how well it reasons about layout, hierarchy, and visual judgment when a step calls for those decisions. These are the two axes people assume matter most, and they do — but they are only half the picture.
Velocity is how quickly a model responds, which matters more for steps that happen many times in a single editing session than for a one-time draft. Token cost is what it costs to run a given step at the volume Eazy actually runs it, which matters because a tool used by thousands of people cannot treat every call as free. Ignoring either axis eventually shows up as a slow product or an unsustainable one.
Why We Don't Bet the Product on One Lab
Committing permanently to one AI lab means inheriting its weaknesses too, with no way out if a competitor ships something better. Eazy treats model choice as an ongoing evaluation, not a one-time decision, so new releases from Anthropic, Google, and OpenAI get tested on the same four axes before anything changes for users.
Frontier labs ship new model versions constantly, and the gap between them changes with almost every release — one update narrows a reasoning gap, the next improves speed or cost at a different lab. A product that locks itself to a single vendor's roadmap is betting that one lab stays ahead on every axis, indefinitely. That is not a bet worth making on your behalf.
So Eazy treats every model as a candidate, not a commitment. When a lab ships something new, it gets evaluated the same way the current model was: content quality, design capability, speed, and cost, tested against the actual steps Eazy runs. If it wins on the axis that step needs, it can be adopted for that step. If it does not, nothing changes and nobody notices.
This is also why we won't publish a fixed table of exactly which model handles which step — that table would be out of date within a release cycle, and it invites exactly the kind of brand-loyalty thinking we're trying to avoid. What stays constant is the process, not a locked-in vendor.
How the Selection Actually Works
Eazy's model selection follows a repeatable process, not a one-time setup: evaluate new candidates against the four axes, match each step in the deck-building pipeline to whichever model actually fits it best, route automatically behind the scenes, and keep a fallback so a single provider hiccup never breaks your editing session.
1. Evaluate. Before a new model version changes anything in production, it gets tested against the same four axes — content quality, design capability, speed, and cost — using the actual kinds of steps Eazy runs, not generic benchmarks.
2. Match the step to the model. Not every step needs the same strengths. Drafting an argument from your document calls for content quality; producing a chart configuration or laying out a slide calls for design and structural reasoning; a quick reformat calls mostly for speed.
3. Route automatically, with cost in mind. Heavier steps that need deep reasoning get a model built for that. Lighter, high-volume steps get a lighter model sized to the job, because running a maximal frontier model on a trivial step wastes speed and money for no quality gain you would notice.
4. Keep a fallback. If a provider is slow or briefly unavailable, Eazy can fall back to another capable model so your edit still completes. You experience a working editor; you don't experience — or need to know about — which model handled which request.
Cost-Effective by Design, Not by Accident
Keeping Eazy affordable is an engineering discipline, not a side effect: lighter models get routed to lighter steps instead of running everything through the most expensive option, repeated or predictable calls get cached instead of recomputed, and fallbacks absorb provider issues without passing the cost or the failure on to you.
Running every single step through the largest, most expensive frontier model would be the easy default and the wrong one. Most of what happens while you edit a deck is not a hard reasoning problem — it is formatting, small rewrites, structural bookkeeping. Sizing the model to the actual difficulty of the step is the single biggest lever for keeping the product affordable to run, which is part of how Eazy can offer free early access with credits included.
Caching is the second lever. When the same input would produce the same or a very similar result, there is no reason to pay for a fresh model call every time. Reusing prior work where it is safe to do so cuts cost without touching quality, because the answer was already good.
Fallbacks exist for reliability, but they also protect cost. If a preferred model is temporarily degraded or unavailable, routing to a capable alternative avoids both a broken editing session and the kind of retry storms that quietly rack up spend. None of this is exotic — it's the same generic playbook (right-size the model, cache what repeats, fall back gracefully) that any provider-flexible engine needs, applied specifically to the steps of writing, designing, and refining a deck.
Why Image Generation Gets Its Own Model
Writing an outline and generating an image are different capabilities, so Eazy uses a dedicated image model alongside its text-and-reasoning models rather than asking one general model to do both. That specialization is part of why images you generate inline can carry a specific look instead of a generic AI-art average.
Text and reasoning models are built to read, argue, and structure. Image generation is a different discipline with its own trade-offs around style, composition, and speed. Treating the two as one job and picking a single model to cover both would mean compromising on at least one of them.
That's why, when you generate an image inline while building a slide, that request goes to a model chosen specifically for image generation rather than whatever model happened to draft your outline. It is evaluated on the same kind of criteria — quality of the result, speed, and cost — but against image-specific competitors, not general-purpose ones.
The practical effect for you is simple: drop in your own images or generate one inline, and either way it lands in a slide already designed for you and on-brand by default. The model choice behind that step is invisible; the result in your deck is what matters.
A Chat Model Drafts Text. Eazy Designs the Deck.
Open a raw chat window with Claude, Gemini, or GPT and ask for a presentation, and you get well-written text back — an outline, maybe some bullet points. What you don't get is a designed, exportable deck: real layouts, a design system, working charts, or a file you can present. That gap is Eazy's job.
This distinction matters because it is easy to conflate "a good AI model" with "a good presentation tool," and they are not the same thing. A frontier chat model is genuinely excellent at drafting and reasoning about content in plain text. It has no native concept of a slide canvas, a theme, a chart that renders as real interactive SVG, or a PPTX file your audience can open.
That is the layer Eazy adds on top of model selection: taking whatever content quality the underlying models produce and turning it into slides that are designed for you and on-brand by default, restyleable in one click by applying a theme, with real charts and layouts, exportable to PDF and PPTX. The model writes and reasons; Eazy is the presentation tool that turns that into a deck.
So when this page says Eazy evaluates Claude, Gemini, and GPT, that is about which model drafts and reasons well behind the scenes — not about which one designs your deck. No raw model does that part. Eazy's editor, its design system, and its export pipeline do, regardless of which model handled a given text step that day.
Ready to write your next deck?
Bring a doc, a link, or a prompt. Watch it become a deck you're proud of.