How to Choose an AI Model: Cost, Speed or Quality?

Choosing an AI model is no longer a question of "which one is best" but of "which one fits this exact task, at what cost, and how fast". That decision rests on three levers that can never all be pushed at once: cost per million tokens, latency, and reasoning quality.

Two announcements on 13 August 2026 make the trade-off concrete. Google DeepMind launched Gemini 3.7 Flash, and OpenAI, working with Cerebras, unveiled a tier called "Ultrafast". Every figure in this article is accurate as of that date, and this market moves quickly.

Why is picking a model no longer about "the most powerful one"?

Behind the phrase "AI model" sits an entire family of variants. What an AI model actually is is worth defining once and for all: every major publisher now offers a fast, cheap model, a mid-tier model, and a slower, pricier reasoning model. Defaulting to the most capable one every time means paying for quality a task simply does not need.

The opposite mistake happens too: pushing a small model onto every task to save money, and losing reliability exactly where a decision is on the line. The right move is to qualify the task first, then pick the model, not the other way round.

What are the three levers that pull against each other?

Three variables move in opposite directions almost every time: cost per million tokens, latency and throughput, and reasoning quality. Pushing one nearly always pulls another one back.

Cost per million tokens

A model bills input and output separately, at different rates. Understanding tokens and context windows explains why feeding a long document into every question runs up the input bill even when the answer is two lines long. That is often where an unexpectedly high AI invoice comes from.

Latency and throughput

Latency is the delay before the first word of a response arrives. Throughput is how fast the rest of it gets written. Each matters differently depending on the job: a voice assistant suffers from high latency, a batch document run suffers more from low throughput.

Reasoning quality

A more capable model chains the steps of a problem together better and errs less often, though never to zero. Understanding and limiting AI hallucinations stays necessary regardless of which model is chosen, including the most advanced one. And when the quality that matters most depends on how well the model knows a company's own documents rather than on the model itself, choosing between RAG and fine-tuning becomes the real question to settle.

What does Gemini 3.7 Flash, released on 13 August 2026, change?

Gemini 3.7 Flash, released on 13 August 2026, is positioned around software engineering and agents. It accepts text, image, audio and video as input, with a context window of up to one million tokens and a maximum output of 64,000 tokens (source: official Google DeepMind model card).

Its introductory price, valid until 31 December 2026, is $0.75 per million input tokens and $3.75 per million output tokens. From 1 January 2027 it rises to $1.50 and $7.50. Worth noting: the model offers three selectable reasoning levels (low, medium, high), which lets a team trade off latency and cost against reasoning depth and tool use explicitly, request by request (source: VentureBeat).

What does OpenAI's "Ultrafast" tier, built with Cerebras, actually mean?

Unveiled the same day, the Ultrafast tier runs a model called GPT-5.6 Sol at a publisher-claimed output speed of 750 tokens per second, up to 14 times the Standard tier. It runs on Cerebras' wafer-scale chips, with announced targets including real-time voice, financial research and incident response (sources: Tech Times, How2Shout).

Two caveats matter here. First, that throughput figure is a claim made by the publisher, not an independently verified benchmark. Second, as of 13 August 2026, Ultrafast is a limited, waitlisted preview: no published price, no general availability date, no official model identifier. For comparison, the Standard-tier prices per million tokens (input / output) are known: Sol at $5 and $30, Terra at $2 and $12, Luna at $0.20 and $1.20. Do not budget for, or promise a client, an Ultrafast price until an official rate card exists.

Why does the latency of a single call become decisive for an agent?

In an ordinary chat, latency is felt once per message. In an agent that chains dozens of calls together, reading a document, querying a tool, drafting, checking, starting again, the same latency repeats at every step. A gap of a few hundred milliseconds per call, trivial on its own, becomes several minutes across a complete task.

That is why the right model choice shifts once a business moves from a chatbot to an agent. An agent automating a specific task usually gains more from a fast, less powerful model called frequently than from a brilliant but slow one called rarely. Good architecture reserves the most capable model for the steps that actually carry a decision, and lets a fast model handle the mechanical ones.

How should an SME choose a model for its own use case?

Three profiles show up most often in an SME, each with its own trade-off.

A data entry, sorting or classification assistant, sorting invoices, pulling fields off a form, summarising a short email, tolerates a small, fast, cheap model perfectly well: the task is repetitive and an error, when it happens, is cheap to fix.

Contract review, a diagnostic, or any summary that feeds into a decision justifies paying for quality and accepting a slower answer: this is where the cost of a mistake dwarfs the cost of the model.

An agent chaining dozens of calls to complete a full task, research, drafting, checking, formatting, should prioritise unit latency above all else, even if that means splitting the work across several models by step. Understanding what an AI agent actually is and what it automates helps frame that architecture before choosing which model runs it.

On that front, the vocabulary of artefacts, skills and hooks gives the reference points for building an agent that applies a stable procedure rather than being re-briefed at every session.

This three-lever framework only earns its keep once it is applied to a real, costed use case. That is exactly what an AI audit for finance does: identify, line by line, where a small model is enough and where quality has to come first, before committing any budget.