The question almost always arrives in the same shape: “Which one should I get?” And almost always the person asking expects a name.

You won’t get one here. Not out of cowardice, but because for about a year now the honest answer has been a different one: for most tasks, the gap between the three top models is smaller than the gap between a well-asked and a badly-asked question. And where there is a clear difference, it rarely sits where the comparison tables go looking for it.

So this isn’t a ranking. It’s a decision by task — with an as-of date, because this text has an expiry.

A disclosure

The first draft of this text was written with Claude. That belongs on the table before anything here says which model is better at what.

Two things follow. First, every figure has a source you can open yourself — not one of them comes from a comparison site; all of them come from the providers’ own pricing and documentation pages, or from named observers. Second, this post crowns no winner. Where I have an opinion, it is marked as one.

And: there is no affiliate link here. There couldn’t be — OpenAI, Anthropic and Google simply don’t run partner programmes. Which in this case makes things pleasantly simple.

What’s available right now

What's available right now

ChatGPT OpenAI

GPT-5.6 Sol

Top model, coding and agents

$5 / $30 1,050,000 tokens 9 July 2026

Knowledge cutoff 16 February 2026 — anything later it only knows if it searches.

GPT-5.6 Terra

Everyday work

$2.50 / $15 1,050,000 tokens 9 July 2026

The model the free and Go tiers run on.

GPT-5.6 Luna

Fast and cheap

$1 / $6 1,050,000 tokens 9 July 2026

Claude Anthropic

Claude Fable 5

Strongest model, long-horizon work

$10 / $50 1,000,000 tokens 2026

Requires 30-day data retention — not available under zero data retention.

Claude Opus 4.8

Coding and agents

$5 / $25 1,000,000 tokens 2026

Claude Sonnet 5

Everyday work

$3 / $15 1,000,000 tokens 2026

Introductory price 2 / 10 until 31 August 2026.

Claude Haiku 4.5

Fast and cheap

$1 / $5 200,000 tokens 2025

The only model in this field without a million-token context.

Gemini Google

Gemini 3.6 Flash

Workhorse

$1.50 / $7.50 1,048,576 tokens 21 July 2026

17% cheaper on output than 3.5 Flash — Google's current default model.

Gemini 3.5 Flash

Predecessor, still available

$1.50 / $9 1,048,576 tokens 19 May 2026

Gemini 3.5 Flash-Lite

High-volume work

$0.30 / $2.50 1,048,576 tokens 21 July 2026

The cheapest output price in the whole field.

Gemini 3.1 Pro

Google's largest model — still labelled preview

$2–$4 / $12–$18 1,048,576 tokens 2026

Price rises above a 200,000-token prompt. The announced successor 3.5 Pro still hasn't shipped two months after the announcement.

Price per token says less than it promises: the models think for different lengths of time, and those thinking tokens are on the same bill. The cheaper model can be the more expensive job.

As of 24 July 2026. List prices in US dollars, before batch or caching discounts — those cut the bill substantially with all three. Check with the provider before deciding anything.

Two things stand out once you lay it side by side.

The context window is dead as an argument. For two years, “how much text fits in” was the strongest differentiator. Today all three sit at around a million tokens — roughly 700,000 words, several fat novels in one go. Only the smallest Claude model breaks ranks at 200,000. Choosing by context window today means deciding on a criterion the providers cleared off the table a while ago.

Google’s biggest model has been labelled “preview” for months. On 19 May, at its own developer conference, Google announced a Gemini 3.5 Pro — “already in use internally”, shipping “next month”. On 21 July, two months later, three new models arrived instead, and none of them was 3.5 Pro. Google’s strongest available model still carries version number 3.1 and the word “Preview” in its name. Everything else on offer is a Flash model — the fast, cheap class.

That’s not a scandal, but it is a pattern: with all three, what gets announced and what you can actually buy are two different lists.

The decision, by task

Writing code and long autonomous work

This is the most contested one — and the benchmarks aren’t contradicting each other, they’re measuring different things.

Claude Opus 4.8 leads on SWE-Bench Pro (real GitHub issues) at 69.2%. GPT-5.6 Sol sets a new high of 53.6 on “Agents’ Last Exam”, comfortably ahead of Claude Fable 5 there. Gemini 3.5 Flash reports 76.2% on Terminal-Bench 2.1. Three exams, three winners.

Simon Willison, who uses all three daily and has written soberly about them for years, put it this way after the GPT-5.6 launch: despite the record scores, the model had not struck him as better than Claude Fable on complex coding tasks.

My advice: don’t decide by the model here, decide by the tooling around it — how well the thing fits your editor, your terminal, your habits. The models are closer to each other than the environments they run in.

Image, video, sound

The one point with a clear answer: Google.

Google is the only one of the three where image and video generation sit on the same price list as the language models — Veo 3.1 from $0.05 per video second, Imagen 4 from $0.02 per image. On top of that comes Gemini Omni, a video model you can edit across multiple conversational turns, which remembers what you discussed earlier in the session.

If media is your subject, this isn’t a close call.

Reading and summarising long documents

Doesn’t matter. Genuinely doesn’t. See above — all three swallow several hundred pages in one go. Use whichever you already have open.

Lots of small jobs, automated

Gemini 3.5 Flash-Lite has the cheapest output price in the entire field at $0.30 in and $2.50 out per million tokens — a factor of two to three against the small models from the other two.

But be careful with that arithmetic; see the next section.

If you already live inside Google

The cheapest entry in the field is Google AI Plus at $4.99 — with 400 GB of storage on top. And Google’s $19.99 tier includes YouTube Premium Lite.

That’s the real reason Google wins with private users, and it has nothing to do with model quality. If you’re paying for storage and YouTube anyway, the AI is close to free on paper.

If data protection matters professionally

A detail that shows up in almost no comparison: Anthropic’s strongest model, Fable 5, requires 30-day data retention and isn’t available under zero data retention at all. If your firm, practice or client mandates exactly that, the flagship is off the table — and you end up on Opus 4.8, not Fable.

Lines like that live in the developer docs, not on the product pages.

What the benchmarks leave out

Price per token is broken as a measure. The models “think” before they answer, and those thinking tokens are on the same bill as the answer. A model at half the token price that reasons three times as long is more expensive at the end of the month. Willison says exactly this: raw price-per-token comparison has become misleading. Anyone who really wants to know what something costs them has to push their own typical task through all three and look at the bill — not at the price list.

The exams are graded by the examinees. After OpenAI’s models trailed Anthropic’s on SWE-Bench Pro, OpenAI published a critique of that benchmark — estimating that around 30% of its tasks are broken. That may be entirely correct on the facts. It is still a witness testifying in its own case. And the same caution runs the other way: anyone celebrating a benchmark they lead is equally interested.

And since July 2026, the provider no longer decides alone when a model ships. GPT-5.6 was cleared for broad release only after a completed review by a US government body — the Center for AI Standards and Innovation at the Department of Commerce. That says more about the next few years than any percentage in this post.

What it costs

What it costs

ChatGPT OpenAI

ChatGPT Go

€8 / month

The cheap middle tier. Runs on Terra.

ChatGPT Plus

€23 / month

The plan most people mean. Free choice of model.

ChatGPT Pro

€229 / month

For constant use and Sol reasoning without a tight cap.

Claude Anthropic

Claude Pro

$20 / month ($17 billed annually)

Includes Claude Code. No cheaper middle tier exists.

Claude Max

from $100 / month ($200 for 20×)

5× or 20× Pro's usage limits. Same models.

Gemini Google

Google AI Plus

$4.99 / month

The cheapest entry in the field, incl. 400 GB storage.

Google AI Pro

$19.99 / month

Plus YouTube Premium Lite and storage — the bundle is the argument.

Google AI Ultra

$99.99 / month ($199.99 for 20×)

5× or 20× Pro's limits, 20 TB storage.

As of 24 July 2026. OpenAI lists euro prices for Germany, Anthropic and Google list dollars — so both currencies appear here exactly as each provider states them.

The gap that stands out: Claude has no cheap middle tier. It goes straight from free to $20. ChatGPT has Go at €8, Google AI Plus sits at $4.99. If you use AI regularly but not professionally, Anthropic simply doesn’t have an offer for you.

The other way round: Claude Pro includes Claude Code, the programming tool, at no extra cost. Whoever that matters to already knows it.

Why this text has an expiry date

Look at the pace.

OpenAI: GPT-5.1 in November 2025, 5.2 in December, 5.4 in March 2026, 5.5 on 23 April, 5.6 on 9 July. Five generations in eight months.

Google: Gemini 3.5 Flash on 19 May, Gemini 3.6 Flash on 21 July. Two months.

What that means in practice: an annual plan isn’t a bet on a model, it’s a bet on a provider — the model you picked won’t exist in that form a year from now. Anthropic advertises $17 instead of $20 for annual billing; that’s $36 saved in exchange for not being able to change your mind for twelve months. In a field rolling out a new flagship every eight weeks, that’s the worse trade.

Pay monthly. That’s the most practical advice in this entire post.

My take

If you want one and would rather stop thinking about it: take the one whose interface gets in your way the least. The quality difference will affect your work less than whether you enjoy opening the thing.

If you can pay for two, the sensible combination isn’t “the best plus the second best” — it’s one for text and code, one for media — and in that second slot, Google is close to unopposed.

And if in three months you’re wondering again whether you chose right: this post will have been updated by then. The version history at the bottom shows you what changed.

For the tools around the models — image, video, music, editing, automation — I keep a separate, continuously maintained index of 137 AI tools.

Sources

All checked on 24 July 2026.

If something here looks outdated or wrong to you: get in touch. That’s exactly what the date at the top is for.