AI Plans That Let You Choose the Model: One Subscription, Many LLMs

A model picker has quietly become a product category. Six subscriptions now let you switch between Claude, GPT, Gemini, Grok and the Chinese open-weight models inside a single plan and a single bill: Abacus.AI's ChatLLM Teams at $10 a month, Poe at $19.99, Perplexity Pro at $20, Kagi Ultimate at $25, GitHub Copilot Pro at $10, and Cursor from $20. Every one of them is cheaper than paying Anthropic and OpenAI separately.

And every one of them meters you differently, which is where the interesting part lives. I spent the past few weeks going through these plans for a selfish reason — I have been looking for an alternative to paying for Claude and ChatGPT at the same time — and the research changed what I was looking for. The question is not "which plan has the most models". It is "what currency does this plan charge me in, and how fast does my favourite model burn it".

The short version: what each plan actually gives you

Why I went looking in the first place

I wrote last year about the quiet death of the $20 AI subscription — the observation that the entry tier runs out exactly when you start doing real work, and the next rung up is a $100 conversation with yourself. Nothing about that has changed. What changed is that I got tired of the arithmetic.

Paying Anthropic and OpenAI at the same time is $40 a month for two products that overlap on perhaps 70% of what I ask them. I do not need two chat interfaces. I need the ability to put the same question to a different model when the first one is being stubborn — which, if I am honest about my own usage, is the only reason the second subscription exists.

So the brief was narrow: find something that replaces at least one of those two seats, gives me a real model picker rather than a marketing claim, and does not quietly downgrade me to a cheap model when I am halfway through something. Six plans made the shortlist. None of them answered the brief the way I expected.

These are two different products wearing the same label

The first thing the research clarified is that "one subscription, many models" describes two categories that barely compete with each other.

Aggregator chat — Poe, ChatLLM Teams, Kagi, Perplexity, T3 Chat — is a chat box with a dropdown. You are buying breadth: the ability to run the same prompt past four labs before lunch. Developer tooling with a model picker — Cursor, GitHub Copilot — is an IDE that happens to let you choose the engine. You are buying the harness, and the model choice is a setting inside it.

Confusing the two is the most common mistake I see in the "one plan, all models" comparisons floating around. A $10 ChatLLM seat does not replace Copilot, because it has no idea what is in your repository. A $20 Cursor plan does not replace a chat assistant, because you do not want to open an IDE to ask about a contract clause. If you are choosing, start by deciding which of those two jobs you are actually paying to solve.

The aggregators: cheap breadth, metered in a currency you did not choose

ChatLLM Teams is the value outlier and it is not close. Abacus.AI charges $10 per user per month — $7 for the first — and its FAQ lists access to "all the state-of-the-art LLMs including GPT-5.6 Sol, Sonnet 5, Opus 5, Fable 5, Gemini 3.1 Pro, Gemini 3.7 Flash, Grok 4.6, Qwen 3.8 Max, DeepSeek v4, Kimi K3, GLM 5.2 and 20 other closed and open source models." That is Anthropic's entire current lineup, OpenAI's flagship and the four Chinese models worth caring about, for half the price of one native plan. The company claims you can send thousands of messages without hitting a limit and that it serves "about 10x more tokens than any other LLM service" — a vendor claim, not a measurement, and the same FAQ notes that large attachments do trigger usage limits.

Poe is the oldest answer to this question and the most honest about the trade. Quora sells you compute points, not messages: 660,000 a month at the headline $19.99 tier, scaling to 8.25 million at $249.99, with a $4.99 entry plan and roughly 17% off for annual billing. Points are a real currency — an Opus-class model drains them at a completely different rate from a Flash-class one — which means Poe is the only plan here where you can watch the cost of your model preference in real time. That is either the best feature on this list or the most stressful, depending on your temperament.

Every one of these plans converts your model choice into a currency. The catalog is never the constraint. The exchange rate is.

Kagi Ultimate at $25 is the odd one, because you are really buying a search engine and the Assistant comes with it. Its top tier unlocks Research mode plus flagship Claude, GPT and Gemini versions alongside GLM-5.1, DeepSeek V4 Pro, Mistral Medium 3.5 and Nemotron 3 Ultra. If you were already considering paying for search without ads, the LLM access is close to free. If you were not, $25 is the most expensive way onto this list.

Perplexity Pro at $20 sits in the same family: a model selector bolted onto an answer engine, with a $200 Max tier above it. The models are a means to the citation, not the point of the product. That is a genuine strength for research work and a genuine limitation the moment you want a long, stateful conversation.

T3 Chat deserves a mention as the budget end — $8 a month, model switching mid-conversation, with a monthly message cap in the low thousands. Its catalog moves fast enough that I would check the current model list before subscribing rather than trusting any review, including this one.

Cursor: the clearest example of a picker that is also a nudge

Cursor is the plan most people mean when they say "the one that lets you choose", and its pricing page is the most instructive document in this entire category. Cursor's own docs are explicit: "There are two separate usage pools, each resetting with your monthly billing cycle."

Pool one is Cursor Models — Grok 4.6, Grok 4.5, Composer 2.5 — and every paid tier gets what the docs call generous included usage. Pool two is Other Models: Anthropic, OpenAI, Google, Z.ai, Moonshot. That pool is denominated in dollars, and you get $20 a month on Pro, $70 on Pro+ and $400 on Ultra. The plans themselves run $0 for Hobby, $20 Pro, $60 Pro+ and $200 Ultra. (There is also an India-only Start plan that covers the Cursor Models pool with no third-party access at all.)

Read that structure twice. On a $20 Pro plan you are given $20 of other people's models and an open bar on Grok. The picker is complete — Claude Opus 5, Sonnet 5, Fable 5, the GPT-5.x family, Gemini 3.x, GLM 5.2, Kimi K3 are all in there — but the economics point in exactly one direction.

It is worth remembering who owns the house now. Cursor's own documentation describes the Grok models as coming from a Cursor/xAI partnership, and Musk acquired Cursor outright earlier this year. A free-choice editor whose free choice happens to be the proprietor's model is not a scandal — it is a business, and Cursor is not hiding it. But "you can use any model" and "you can afford to use any model" are different sentences, and only one of them is on the pricing page.

GitHub Copilot: the picker is the paid feature

Copilot's catalog is arguably the broadest of any developer tool right now. GitHub's supported-models documentation lists OpenAI's GPT-5.x family, Anthropic's Claude Opus 5, Sonnet 5, Fable 5, Haiku 4.5 and the 4.5–4.8 Opus line, Google's Gemini 3.x, xAI's Grok 4.5 and 4.6, Moonshot's Kimi K2.7 Code and K3, and Microsoft's own MAI-Code-1-Flash models.

The pricing detail that matters: Copilot Free gets no model choice at all. GitHub's plan documentation describes it as limited access through automatic model selection only. Picking your model starts at Copilot Pro ($10/month), widens on Pro+ ($39/month), and scales through Business ($19 per seat) and Enterprise ($39 per seat). Usage runs on premium requests, and if you want to understand what one of those actually costs you, we broke the conversion down in GitHub Copilot AI credits explained — one credit is $0.01, which makes the arithmetic much less mysterious than the interface suggests.

At $10 a month, Copilot Pro is the cheapest legitimate way to put Claude Opus 5 and GPT-5.6 in the same window. If you spend your day in an editor anyway, it is hard to argue with. We compared the two head to head in Cursor vs GitHub Copilot.

What you actually give up

This is the part the aggregator marketing skips, and it is the part that decided things for me.

You are buying models, not products. The $20 you pay Anthropic or OpenAI does not only buy inference — it buys Projects, Artifacts, memory, Cowork, the mobile and desktop apps, the agentic tooling, the connectors. None of that travels. When you select Opus 5 inside ChatLLM or Poe, you get the model's raw capability inside somebody else's wrapper, and that wrapper is usually a year or two behind the first-party product on everything except the model dropdown.

Three specific things go missing that I would not have predicted:

• Statefulness. Native apps remember your projects, your files, your instructions. Aggregators mostly remember your last thread.
• The long tail of the model. Feature flags, extended thinking controls, context-window options and new capabilities land in the first-party product first and in the aggregators whenever the API exposes them.
• A stable bill. Points and credits mean your cost per month is a function of which model you happened to prefer that week. A native plan is boring, and boring has value.

The escape hatch: stop subscribing entirely

The honest fourth option is to buy no subscription at all. OpenRouter is not a plan — you load credits and pay per token, at the same rates you would pay each provider directly, with a 5.5% fee (minimum $0.80) when you buy credits by card and 5% by crypto. Several hundred models, one API key, no monthly commitment.

For light or bursty use this beats every plan on this page, and it is not close. If you send a few dozen serious prompts a month, a subscription is a rounding error away from being pure waste — the case we made in are you overpaying for AI? Before you commit to anything here, put your real usage through the token calculator and see what the same month would have cost at API rates. For me the answer was unambiguous, and not in the API's favour — I use these things for hours a day. Yours may well go the other way.

Where I landed

I set out to replace one of my two subscriptions and ended up rejecting the framing. Here is the reasoning, in case it saves you the same three weekends.

The aggregators do not replace a native seat — they replace the second one. That distinction is the whole finding. My daily driver is deep, stateful, hours-long work in one product, and no $10 wrapper does that job. But my second subscription was never doing that job either. It existed so I could ask a different model the same question. That is precisely what a $10 ChatLLM seat or a $19.99 Poe tier does, at half the price and with roughly thirty models instead of one.

So the shape that survives scrutiny is: one native plan for the work, one cheap aggregator for the second opinion, and nothing else. That is $30 a month against $40, and the $30 version reaches Opus 5, GPT-5.6, Gemini 3.1 Pro, Grok 4.6 and the Chinese models instead of two American ones. The second-opinion seat is also where the interesting comparisons happen — it is how I ended up writing Kimi vs Claude, which would not have existed if I had needed a third subscription to run the test.

On the developer side I would not switch to Cursor for the model picker specifically. Not because the picker is bad — it is excellent — but because a $20 plan that includes $20 of the models I actually want is a plan I will overspend on every single month. Copilot Pro at $10 with a transparent credit conversion is the more predictable purchase, and predictability turns out to be the feature I was underrating.

One caveat I will state plainly: every number on this page was verified in August 2026, and this category has re-priced repeatedly this year — Cursor alone changed its billing model three times. Check the current pricing page before you put a card in. The structural argument will outlive the numbers; the numbers will not outlive the quarter.

Frequently asked questions

Which AI subscriptions let you choose between multiple LLMs?

Six plans do it well. Abacus.AI's ChatLLM Teams ($10/user/month) and Poe ($19.99/month for 660,000 compute points) are chat aggregators with the widest catalogs. Kagi Ultimate ($25/month) bundles flagship models into a search assistant. Perplexity Pro ($20/month) offers a model selector on top of search. On the developer side, Cursor (from $20/month) and GitHub Copilot Pro ($10/month) both ship a model picker covering Anthropic, OpenAI, Google, xAI and Moonshot models.

Does Cursor let you choose which AI model to use?

Yes, but the choice is metered in two separate pools. Cursor's own models — Grok 4.6, Grok 4.5 and Composer 2.5 — come with generous included usage on every paid plan. Third-party models from Anthropic, OpenAI, Google, Z.ai and Moonshot draw on a separate dollar-denominated pool: $20/month on Pro, $70 on Pro+ and $400 on Ultra. You can pick any model; you can only pick the expensive ones for so long.

Is one multi-model subscription cheaper than paying for ChatGPT and Claude separately?

On sticker price, yes: ChatLLM Teams at $10/month or Poe at $19.99/month costs less than two $20 subscriptions. On heavy use, often no. Aggregators meter you in points or credits and rarely include the native product features — Projects, Artifacts, agentic coding, mobile apps — that the first-party plans bundle for free. The saving is real for breadth and disappears for depth.

Can you use Claude Opus 5 without an Anthropic subscription?

Yes. Claude Opus 5 is selectable inside Cursor, GitHub Copilot and Abacus.AI's ChatLLM Teams, and available per-token through OpenRouter, all without an Anthropic plan. What you do not get through those routes is Anthropic's own product layer — Projects, Cowork and the Claude apps are tied to the first-party subscription.

What is the catch with multi-model AI subscriptions?

The metering, not the catalog. Every one of these plans converts your model choice into a currency — Poe uses compute points, Cursor uses two separate dollar pools, GitHub Copilot uses premium requests where one credit equals $0.01. Expensive models drain that currency faster, so an unlimited-looking catalog is really a budget you spend at different rates depending on which model you pick.

Does GitHub Copilot Free let you pick a model?

No. GitHub's plan documentation states that Copilot Free offers limited access through automatic model selection only. Model choice starts at Copilot Pro ($10/month), which unlocks a selection of models, and widens on Pro+ ($39/month), Business ($19/seat) and Enterprise ($39/seat).