The right AI model for each task: stop automatically taking the biggest one.

Since AI assistants started offering dropdown menus of templates, a habit has taken hold in open-plan offices: open the tool, select the latest and most powerful template, and then ask it to rewrite three lines of email. It's a bit like taking a 38-ton truck to pick up bread. It works, but it's slow, expensive, and ties up the truck for those who really need it.

This article explains, without jargon, how model ranges work, why the biggest is almost never the best choice, and which model to choose for each major family of tasks at work.

1. A range of models is like a range of cars

Each publisher now offers three or four models ranging from small to very large. The names change every three months, but the logic remains the same.

What these tiers actually change is the response speed (small tokens respond in one second, large tokens can take a minute to think), the depth of reasoning, and the cost. Looking at publishers' public pricing, there's a factor of 10 between small tokens and frontiers at Anthropic (Haiku at $1 per million input tokens, the publishers' billing unit, Fable at $10), and a factor of 100 at OpenAI between GPT-6 Luna ($0,10) and Astra ($10). Google and Mistral show similar price differences.

Updated September 26, 2026. Two releases have shaken up this picture since the article was published. Anthropic launched Claude Opus 5.5, which performs on par with Fable 5.1 on most tasks, responds more than 30% faster than Opus 5, and costs about 40% less to run; at the entry price, it costs $4 per million tokens compared to $10 for Fable 5.1, a 60% discount. OpenAI replaced GPT-5.6 Sol and Luna with GPT-6 Sol and GPT-6 Luna, with prices cut in half. These announcements confirm the point made in this article: what required the frontier model three months ago is now accessible at a lower price point, for a fraction of the cost.

Updated October 12, 2026. At the end of September, OpenAI launched GPT-6.1 Sol, which it presents as being close to GPT-6 Astra, its most powerful model, at a fifth of the price. The trend is confirmed month after month: the power of the frontier model drops a notch, and the "biggest by default" reflex is becoming increasingly expensive for no reason. For code agents, see also our comparison. Claude Code or Codex.

When you have a subscription, you don't pay per request. But your company pays for these discrepancies through licenses, quotas, and usage limits: most plans restrict messages on powerful models (for example, 200 messages per week on GPT-6 Pro in ChatGPT's Pro plan) and are virtually unlimited on fast models. And if your teams use AI through an internal tool connected to an API, each message is billed according to the chosen model.

2. Why bigger isn't always better

Three factual elements, drawn from independent rankings, deserve to be known.

For routine tasks, the difference is invisible. The Arena rating system (formerly LMArena) involves hundreds of thousands of users voting blindly between two answers. In September 2026, the top ten models are all within 13 Elo points, while their prices range from 1 to 50. A versatile model like Gemini Flash or Claude Sonnet sits alongside edge models. When writing, summarizing, translating, or rewording, you wouldn't be able to tell the difference with your eyes closed.

The difference only appears on long and complex tasks. The GDPval benchmark, built by OpenAI using real-world professional deliverables (reports, spreadsheets, presentations, across 44 different industries), clearly demonstrates an advantage for frontier models. However, these are tasks that take several hours and involve dozens of steps. It's not simply rewriting an email.

The price per word doesn't tell the whole story, nor does the time spent thinking. An independent analysis of coding tasks aptly summarizes the pitfall: "A model that's 12 times cheaper per token can end up costing more per task if it takes 20 times longer to process." The right indicator isn't the advertised price or the model size, but the cost to achieve a result you're willing to accept.

The developers themselves say so. Anthropic recommends in its documentation an "efficiency first" approach: start with the basic model and only upgrade if the results are disappointing. OpenAI describes Luna as the model for "everyday chat, summarizing, writing, translation, and practical questions." Google explicitly positions Flash-Lite for "translation and data processing."

3. The printable table: which template for which task

The general rule in one sentence: Short and well-defined task, small model; long, vague or multi-step task, large model.

A helpful guideline: if you can formulate your request in one sentence and will be able to assess the result at a glance, the small template is sufficient. If you need to provide three documents in context and will have to review them extensively, choose the large one.

4. Sliders that nobody uses

Consumer interfaces hide three settings that make more of a difference than changing the model.

The reflection cursor. ChatGPT offers "Instant," then several levels of response up to "Pro." Gemini distinguishes between "Fast" and "Thoughtful." Claude automatically adjusts its response time based on the question. This slider, using the same model, transforms a one-second response into a one-minute response, with the corresponding cost. For email, stick with "Instant."

Automatic mode. ChatGPT, Gemini, and Microsoft Copilot can now automatically route a question to the appropriate effort level. Leaving it on "Auto" is a good starting point for users who don't want to think about it. The only time you should manually adjust the difficulty level is for a long task that you know is challenging.

The context provided. A simple template with the correct document attached outperforms a more complex one that simply guesses. Most complaints attributed to "not powerful enough templates" stem from requests lacking context. Before upgrading, include the file, the original email, and the expected example.

5. What this changes at the company level

Studies on "model routing" (sending each request to the least expensive model capable of handling it) are converging. FrugalGPT, a 2023 Stanford study, achieved up to 98% lower costs for the same quality compared to GPT-4 in its tests. RouteLLM, in 2024, reduced costs by a factor of 3,66 while maintaining 95% of the quality. Feedback from practitioners in 2026 estimates that 60 to 80% of common enterprise requests are handled well by a low-cost model.

There's also an internal fairness argument: in most enterprise subscriptions, quotas for powerful templates are shared or capped. Every email rewritten with the frontier template is one less request for the colleague who has to analyze a 200-page request for proposals.

The same logic applies to the choice of the tool itself: Copilot, ChatGPT, Claude, Gemini, and Mistral are not all equal, depending on the business needs and the company's data. Our page Which AI tool should I choose? help to decide, and our AI training by tool take place on the one your IT department has selected.

6. A four-step method

1. Start small. The default fast or versatile model, in instant mode.
2. Provide context before changing the model. Attachment, example, precise instructions.
3. Only move up a level if the result is disappointing. and note down what type of task it is. After two weeks, you will have your own chart.
4. Reserve the boundary model for long tasks In-depth research, a comprehensive deliverable, multi-step reasoning. Don't skimp on this: that's precisely why it exists.

7. And what applies to the budget also applies to the planet

The entire economic argument in this article has an ecological counterpart, and it's probably the strongest argument for widely adopting these practices. For the same response, a small model consumes far less electricity and water than a frontier model. The reason is simple: the energy expended to produce a response depends first on the number of parameters activated for each word generated, and then on the number of words generated. A distilled model literally means less computation for a result you won't even notice.

What matters, at the organizational level, is the mass effect. A single request carries almost no weight, and blaming an employee for asking an AI a question makes no sense. But ten thousand people rewriting three emails a day using the frontier model, reasoning to the max, is a different matter altogether: it's the same work delivered, but with an order of magnitude more effort. That's where the leverage lies, and it's collective, not individual.

The good news is that this approach to efficiency doesn't require any sacrifices. Using the quick email template, letting automatic routing do its job, keeping instant mode as the default, and reserving more complex processes for tasks that truly warrant it: these are precisely the steps that reduce costs and response times. Trade-offs where cost, environmental impact, and user-friendliness all align are rare enough to be noteworthy. And because they rely on habitual practices, they belong in the AI ​​usage guidelines and training programs, rather than as an annual reminder of individual goodwill.

The right reflex is therefore not "the most powerful", but "the most suitable". And in nine out of ten cases in the office, the most suitable is also the fastest, the cheapest and the least energy-intensive.

Explore more!

• Arena, ranking by human preference in a blind test: arena.ai/leaderboard
• Artificial intelligence, speed and price compared: artificialanalysis.ai/models
• GDPval, actual professional tasks: openai.com/index/gdpval and independent monitoring artificialanalysis.ai/evaluations/gdpval-aa
• Anthropic, model selection guide: platform.claude.com and prices of the Claude models, including Opus 5.5: platform.claude.com/pricing
• OpenAI, launch of GPT-6 Sol and GPT-6 Luna: openai.com
• OpenAI, recommended models and uses: developers.openai.com et models available in ChatGPT
• Google, Gemini pricing and positioning: ai.google.dev
• Splunk, reducing costs through template routing (FrugalGPT, RouteLLM, Minions): splunk.com
• Epoch AI, the fall in the price of inference: epoch.ai
• Google, environmental impact of a Gemini query: cloud.google.com
• AI Energy Score, comparative energy consumption of models (Hugging Face, Carnegie Mellon, Cohere, Salesforce): huggingface.github.io/AIEnergyScore
• Energy of inference and cost of extended reasoning, review Joule (2026): cell.com/joule
• Mistral, life cycle analysis of Large 2: mistral.ai