How to choose a model and a plan
Reading benchmarks and rankings without being fooled, knowing when a subscription or the API makes sense, and choosing by profile. With October 2026 prices.
Mattia PanattaUpdated 15 October 202612 minBeginner
In short
- The big three (ChatGPT, Claude, Gemini) are all excellent: the choice depends on what you do and the tools you already use, not on who is “best”.
- Benchmarks measure capability on standard tests; Arena measures what people prefer. Neither measures your work: test on 5 of your own tasks.
- Daily use: one ~€20/month subscription is almost always enough.
- APIs make sense for automations and products; for personal use, a subscription.
How to read the numbers
Artificial Analysis Intelligence Index
A composite score combining several independent benchmarks: reasoning, knowledge, maths, programming, tool use, agentic tasks. Artificial Analysis runs the tests itself, under the same conditions for everyone, and also measures speed and cost per task.
What it tells you: how capable a model is overall, and what that capability costs. As of October 2026 the top is 58 (Claude Opus 5.5); GPT-6.1 Sol scores 52 at an eighth of the cost per task.
What it doesn’t: how it does on your task, how pleasant the style is, how the app behaves.
Arena
On Arena people ask a question, get two answers from two anonymous models and vote for the better one. Millions of votes become an Elo score, as in chess. There are separate rankings for text, web development, agents, images, video.
What it tells you: which model people prefer on real requests. What it doesn’t: whether the preferred answer was also correct (style persuades), and newly released models have few votes (Arena marks them “preliminary”).
A few points don't matter
Between 1,508 and 1,502 on Arena, or 53 and 52 on the index, the difference is within the margin of error. Look at groups (who’s at the top, who’s mid-pack), not exact positions.
All up-to-date data, with an interactive chart and cost calculator, is on the Models & prices page.
Subscription or API?
- Subscription (ChatGPT Plus, Claude Pro, Google AI Pro): a fixed monthly fee to use the app, with usage limits. Includes web search, files, images, voice, projects, agents. The right choice for personal and daily professional use.
- API: pay as you go, per token. For integrating AI into your own programs: automations, chatbots, products. No message limits, but every call costs and you build the interface.
A person using AI daily for work via API might spend more or less than a subscription depending on how much they use reasoning models. For direct use, a subscription is almost always simpler and cheaper.
Plans compared (October 2026)
| Free | Basic | Mid | Top | |
|---|---|---|---|---|
| ChatGPT | Free: GPT-6 Luna with limits | Go €8, Plus €23 | — | Pro from €103 |
| Claude | Free: Sonnet and Haiku | Pro $20 ($17 yearly) | Max 5x $100 | Max 20x $200 |
| Google AI | Free: Flash-Lite | Plus €4.99, Pro €21.99 | Ultra 5x €99.99 | Ultra 20x €219.99 |
Monthly prices from official pages for Italy; Claude in dollars excluding VAT. Top plans include advanced features (always-on agents, Pro models, API credits) and much higher limits.
Which for whom
Occasional use, curiosity. Stay on free plans and use two, say ChatGPT and Claude. When an answer matters, compare them.
You write a lot, analyse documents, work with long texts. Claude Pro. The favourite for writing quality and document reading, and it includes Claude Code if one day you want to build something.
You want everything in one: images, voice, data, charts. ChatGPT Plus. The top-voted image generator on Arena, GPT-6’s new interactive answers, Codex and a huge ecosystem.
You work in Gmail, Docs, Drive. Google AI Pro. Gemini inside Google apps, Gemini 4 Argon, video with Flow and 5 TB of storage: if you already pay for Google One, the step is small.
Student or minimal budget. Google AI Plus (€4.99) or ChatGPT Go (€8), plus free Gemini Notebook for studying.
Heavy use: agents, coding all day. €100+ plans (Claude Max, ChatGPT Pro, Google AI Ultra) pay off only if you often hit the base plan’s limits. Measure for a month first.
Company or professional firm. Team, Business or Enterprise plans: they don’t use data for training, have centralised admin and data-processing agreements. A requirement, not a luxury, if you handle client data.
Choosing a model via API
For automations and products, the rule is: the smallest model that does the job well.
| Task | First choice | Budget alternative |
|---|---|---|
| High-volume classifying, extracting, summarising | Claude Haiku 5.5 | GPT-6 Luna, DeepSeek V4.1 Flash |
| Good-quality chatbot | GPT-6.1 Sol | Gemini 3.8 Flash |
| Complex analysis, agents | Claude Opus 5.5 | GPT-6.1 Sol (max) |
| Code | Claude Opus 5.5 | Claude Sonnet 5.5, GPT-6.1 Sol |
| Data that can’t leave the company | Self-hosted open-weights models (GLM-5.3, MiMo-V2.6-Pro) | Providers with EU servers |
The test worth more than any ranking
Pick five real tasks you do often (a difficult email, summarising a typical document, an analysis, a search, something creative). Run them through two or three assistants with the same prompt and compare. In half an hour you know which works best for you, the only ranking that counts.
Task 1 — [description]. Prompt: [text]. Score ChatGPT /10 · Claude /10 · Gemini /10. Notes. Task 2 — … (Repeat for 5 tasks. Rate: correctness, usefulness, style, how many fixes were needed.)
Three tips to avoid mistakes
- Don’t chase every release. New models come out monthly; the jumps that really change your work are 2-3 a year. Redo your personal test every six months.
- Distrust viral comparisons. An impressive answer on X isn’t a sample: it’s a selection.
- Check prices and limits on the official page before paying. They change often, and the data here is as of 15 October 2026.
Sources
Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.
