Agent Zero · main guide

Pick your model and budget

Agent Zero is free, so you never pay for the agent itself. The model is what can cost money, because it often runs on a company's computers rather than yours.

Several rows below mention multi-step work, which means a later step needs the result of an earlier step. The 1 to 5 numbers are judgments, not test scores. A 5 means it can feel like a company's strongest current model, and no row that runs only on your computer gets a 5.

Prices and limits were checked on 14 November 2026, and they change.

The full table

The first rows use a provider that runs the model for you, and the next rows run the model on a computer you already have. The last rows are unofficial routes: some sign you into a chat plan, and some send your text through a relay.

Each catch is a ban, a surprise bill, or a privacy risk. A surprise bill is a charge you did not sit down to approve, and a privacy risk means someone else can read or keep your prompt.

This page names the unofficial routes so you can see the risk, but it does not tell you how to turn them on.

Option	What you pay in a normal month	Card or computer	How smart it feels (1-5, a judgment)	The catch
Google Gemini free key	$0	No card. It runs on Google's computers.	4. The strongest free set that can use tools. Google's two strongest models are not on the free tier.	Google may use those free prompts to improve its products.
Groq	No confirmed free model	It runs on Groq's computers.	This guide could not confirm which Groq models are free, so this row has no score.	The models page does not label a model free, and no free cap is published.
Z.AI GLM-4.7-Flash	$0	The price page does not say if a card is needed. It runs on Z.AI's computers.	3. The model page says it can use tools.	The privacy policy skips this kind of program access. The rate limit is not published.
Light hybrid: Claude Opus 5.5 for the hard parts, with the cheap hours on one model. For the cheap hours, the API id is `deepseek-flash`, and the version is DeepSeek-V4.1-Flash-1017.	about $17	You pay for what you use. The models run on their companies' computers.	4. About 1 in 10 steps uses Claude Opus 5.5. DeepSeek says Flash keeps up with its stronger model on simple jobs.	Official posted prices. About $21 if every cheap step falls in DeepSeek's pricier hours.
Heavy hybrid: the same mix, with more work and more hard steps	about $185	You pay for what you use. The models run on their companies' computers.	5. About 1 in 4 steps uses Claude Opus 5.5. Anthropic, the company behind Claude, calls it the model for hard multi-step work.	Official posted prices. About $201 if every cheap step falls in DeepSeek's pricier hours.
A key from Google (AI Studio or Vertex), Claude, or OpenAI	Pay for what you use. The notes do not total a normal month.	Credit card. No extra computer.	5 with that company's strongest current model. Lower with a smaller model.	Surprise bill. A busy agent can spend more than a chat plan feels like it should.
8 GB computer, `qwen3.5:4b` (3.3–4.0 GB; Q4_K_M is 3.3 GB)	$0 plus the computer you already have	The computer you already have.	1. Small local file with pictures, tools, and thinking. Long step chains were not measured here.	It depends on your computer. Apache 2.0. Ollama prints 256K of context, which this computer cannot hold.
16 GB computer, `mimo-v2.6:9b` (6.6–8.7 GB)	$0 plus the computer you already have	The computer you already have.	2. Pictures, tools, and thinking, with more room than 8 GB. Long step chains were not measured here.	It depends on your computer. The Ollama page states an MIT license. Ollama prints 256K, which this page does not promise here.
About 24 GB of memory the model can use, `qwen3.8:27b` (18 GB)	$0 plus the computer you already have	The computer you already have. On a Mac, often shared memory. On Windows, often the graphics card's own memory.	3. Pictures, tools, and thinking, and a 32K context fits. Slower than a company's strong model.	It depends on your computer. Apache 2.0. Local Agent Bench, not a paper, measured a 32K context at 17 GB. Do not promise the printed 256K.
32 GB, `qwen3.8:27b` (the same 18 GB file)	$0 plus the computer you already have	The computer you already have.	3. The same model as the row above, with more memory left for the chat.	It depends on your computer. The extra memory is for context. This page does not promise the printed 256K.
48 GB, `qwen3.8:27b-q8_0` (30 GB)	$0 plus the computer you already have	The computer you already have.	3. The same model in a 30 GB file, with pictures, tools, and thinking.	It depends on your computer. A 30 GB file leaves less spare memory than the 18 GB file.
64–80 GB, `qwen3.8:27b-bf16` (56 GB)	$0 plus the computer you already have	The computer you already have.	3. The same model in a 56 GB file.	It depends on your computer. On 64 GB, if the system and the context do not fit, stay on `qwen3.8:27b-q8_0`.
128 GB and up, `qwen3.5:122b` (81 GB)	$0 plus the computer you already have	The computer you already have.	3. A much larger local file, with pictures, tools, and thinking.	It depends on your computer. Apache 2.0. The tag to pull is this 81 GB file.
Gemini CLI or Antigravity sign-in inside someone else's tool	No price to list. Google stopped the consumer "Login with Google" path for Gemini CLI on 18 June 2026.	A credit card on the Google plan, if a login still exists. OpenClaw, a different agent, will not create it.	4, if a login still answers with the named Google models.	Ban risk. Google's rules allow closing the account. February 2026 bans included proxies. A second violation is described as permanent.
Claude plan sign-in inside someone else's app	The plan's price while the 15 June 2026 pause holds, plus any extra-usage credits. Credit bundles: $45 buys $50, $200 buys $250, $700 buys $1,000.	Credit card.	5, when Anthropic is the one answering.	Surprise bill and ban risk. Anthropic's rules forbid passing a personal plan through an outside program, and say they may enforce that.
Sign in with ChatGPT, inside OpenAI's published rules	Your ChatGPT plan, up to its included limit. Extra credits only if you turn them on.	Credit card. The saved sign-in must stay on a computer you control.	4. ChatGPT plan models. The pages did not list how much is included.	Surprise bill, if automatic credit buying is on. OpenAI says that charge can arrive with no separate notice.
Several Claude, Gemini, or ChatGPT logins behind one local proxy, the setup the CLIProxyAPI project page describes	The subscriptions on those accounts, until one is closed. No separate proxy fee was listed.	A credit card on each account. The project page says the proxy runs on your computer.	4, as smart as the account it uses, while that account still works.	Ban risk. Google names suspension. Anthropic may bill extra credits and act against programs posing as others. OpenAI may cut off apps sharing one person's plan.
GitHub Copilot Free, device sign-in	$0. 2,000 code completions a month (times it fills in code). Unused ones do not roll over.	No card. No extra computer.	2. The free plan picks the model, and the allowance is small.	The pages checked show no OpenClaw-specific ban. The free plan chooses the model for you.
GitHub Copilot Pro, Pro+, or Max, device sign-in	Pro: $10 for 1,500 credits. Pro+: $39 for 7,000. Max: $100 for 20,000. One credit is $0.01. Credits reset monthly and do not roll over.	Credit card.	4. Paid Copilot models, until the credits run out.	Surprise bill, once extra use is allowed. GitHub bills it afterward. With a $0 budget, you upgrade or wait for the monthly reset.
gpt4free public addresses	$0 on the addresses marked as needing no key. No amount of use is promised.	No card. No extra computer.	2. Whoever answers. Auto means a working provider, not one company's best model.	Privacy risk. The operator may keep excerpts for 14 days. The provider may store your prompt to improve models. Ban risk if it uses a logged-in browser.
NanoGPT	$12 a month covers 60 million input tokens a week on open-source models. Or pay as you go.	Card (at least $1) or crypto, meaning digital coins (at least $0.10). No extra computer.	3 on the $12 open-source allowance. A closed model such as Claude feels higher only if it is really what answers.	Privacy risk. NanoGPT and the provider it picks can read the prompt. A provider may serve a smaller, less exact copy under the same model name.
RelaySwitch	No monthly fee. Prepay a set amount from $10 to $500, then pay per token. They claim rates under the maker's list price.	Card or crypto. They claim no identity check. No extra computer.	3. The names look like top models. Their own check that day marked 20 models changed and 13 not reached.	Privacy risk. They claim not to store messages, but the machine running the model still gets the prompt. Their daily test is their own.
Chutes	No free tier. Add money, then pay per token, or per second for their graphics chip. One chip: $2.25 an hour plus $6.75 setup.	Card, crypto, a bank transfer through Stripe, or a coin called TAO.	3. The models opened that day were open ones such as Gemma and Qwen, not Claude or GPT.	Privacy risk on ordinary models. A third-party platform's staff can read that content. A locked mode claims more, but an automatic route can leave it.
A discount relay. OhMyGPT is one example, not a recommendation.	Prepaid credits, not a flat monthly fee. 250,000 credits per dollar. Their example: a tiny request cost 3 credits against a list price of 4.	A prepaid balance. No extra computer.	3. The shop picks the supplier unless you do. Some cheap rows put the discount in the model name.	Ban risk. Reselling can get the shop's account closed, leaving your balance stuck. Refunds are generally refused. Privacy risk. They may store and review what you send.

The full notes behind these cells are in the company options, the unofficial routes, and the local options.

The scores in What a local model can and cannot do are publisher claims. They come from the Qwen3.5-4B card, the MiMo page on Ollama, and the Qwen3.8-27B card. The 17 GB fit note is from Local Agent Bench, and that site is not a paper. The memory limits come from Ollama's context length page and its FAQ. The speeds come from llama.cpp's Apple Silicon tests and the CPU-only laptop paper.

Start with Google Gemini's free key

You do not need a card for it, but Google may use those free prompts to improve its products. Z.AI GLM-4.7-Flash is another confirmed $0 model, and this guide could not confirm which Groq models are free.

If you pay a little

The light hybrid is about $17 a month: about 1 in 10 steps uses Claude Opus 5.5 for the hard parts, and the rest use DeepSeek Flash. A heavy month of the same mix is about $185, with about 1 in 4 steps on Claude. If every cheap step falls in DeepSeek's pricier hours, those totals become about $21 and $201.

A $20 ChatGPT or Claude chat plan is a different product. ChatGPT Plus is $20, and Claude Pro is $20 billed monthly, or $17 a month when $200 is billed for the year. Those plans pay for chat inside that company's own app, so the per-use bill for this agent is separate.

On a computer you already own

An 8 GB computer runs qwen3.5:4b (3.3–4.0 GB), and a 16 GB computer runs mimo-v2.6:9b (6.6–8.7 GB). Both of these files can take a picture, call a tool, and use a thinking mode. Ollama also prints 256K for the 4B tag, and an 8 GB computer cannot hold that window. For the 9B tag, Ollama prints 256K too, and this page does not promise that window on 16 GB.

About 24 GB is the first size with room for qwen3.8:27b, an 18 GB file. Local Agent Bench, which is not a paper, measured qwen3.8:27b at 17 GB with a 32K context. Ollama prints 256K for that tag, and this page does not promise that window on 24 GB. A smaller llama.cpp file for this same model is covered on the Local models page. At 32 GB the pick is still qwen3.8:27b, and qwen3.8:27b-q8_0 is the 30 GB file for 48 GB. A 64 to 80 GB computer can run qwen3.8:27b-bf16, which is a 56 GB file. On 64 GB, stay on qwen3.8:27b-q8_0 when the system and the context do not fit. At 128 GB and up, the file to run is qwen3.5:122b, which is about 81 GB.

The cost is $0 plus the computer you already have. Electricity was not priced, although the chip does use more power while it writes a reply than when the computer is idle.

A local model keeps your prompt on your computer. The prompt leaves if you pick a cloud model, sign in so the app can call a cloud model, or let the agent use the web.

What a local model can and cannot do

With 8 or 16 GB of memory, use the Google Gemini free key for work that takes several steps. At those sizes a local model can chat, take a picture, and call a tool. Qwen's own card for the 4B model lists LiveCodeBench v6 at 55.8, TAU2-Bench at 79.9, and BFCL-V4 at 50.3. Those three numbers are the publisher's claim, and this guide did not run that test. MiMo's own table, not an independent paper, puts its 9B model ahead of Qwen3.5-9B on three agent scores. On that table, SWE Verified is 61.1 against 60.0, and SWE Pro is 44.6 against 32.0. The Terminal Bench 2.1 row on that table is 37.1 against 27.0 for Qwen3.5-9B.

Multi-step tool use, browsing, and long documents stay hard, and this page did not measure them here. On a computer with less than 24 GiB of graphics memory, Ollama lets the model hold only 4,096 tokens at once by default. That is about 3,000 words. Ollama's docs say agents and web search need at least 64,000 tokens, and a longer context needs more memory. Qwen's card for Qwen3.8-27B lists Terminal Bench 2.1 at 73.0, SWE-bench Pro at 61.7, and LiveCodeBench v6 at 90.3. That card is the publisher's claim, and Muse Glimmer-30B scores 51.7 and 51.2 on the first two.

Speed depends mostly on your chip. The llama.cpp project tested a 7B model in a file of about 4 GB. It wrote about 14 tokens a second on Apple's base M1 chip, and 21 to 32 on the base M2 to M5 chips. A 2026 paper timed an Intel Core i7-13700H laptop using only its processor. It wrote about 15 tokens a second with a 7B model and 9 with a 12B model, or roughly 7 to 11 words a second. Before writing, the model reads the prompt, which took about 3 seconds for a short 256-token prompt on that laptop. An agent's prompt can be far longer.

The Google Gemini free key also costs $0. The table's judgment is 4 for it, against 1 or 2 for these computers, but Google may use those prompts. The light hybrid costs about $17 a month and sends the hard steps to Claude Opus 5.5. A local model still makes sense for routine steps on private files, with a cloud model doing the planning. The AgentFloor authors suggest the same split, and Mix a cheap model with a strong one shows how each agent does it. To set one up, start with Pick a model for your memory.

Uncensored models: a costly special case

An abliterated model is an open model edited to cut back its built-in refusals, though sellers say some prompts can still be refused. Abliteration charges $1 to $3 per million input tokens and $2 to $5 per million output tokens, while Claude Haiku 5.5 is $0.10 and $0.50 on prompts up to 100,000 tokens and Claude Sonnet 5.5 is $2 and $10. Most readers should not put one in an agent. A hidden instruction in a file can be treated as the request, tool use is only what each seller claims, and some providers keep the prompts. Skip this unless you have a real need, and read Uncensored models before you connect one.

The hybrid above pays Claude's per-use price. A Claude plan sign-in is something else: it is one of the unofficial routes in the table, and this page does not tell you how to turn it on.

On 15 June 2026 Anthropic paused a change and said third-party apps (programs Anthropic did not make) still use the subscription allowance, the amount included in the plan. The login page still says they may charge extra-usage credits, and that you pay those charges.

The next page is the install: The first 10 minutes.