For every agent

Strong models for less

This page shows the cheapest lawful ways to put a strong model behind your agent, and it applies to every agent in this guide. Here, a strong model means one that rates a 4 or a 5 on the 1 to 5 judgment scale in Pick your model and budget. Those numbers are judgments from each company's own wording, not test scores.

Every price below is a light month, using the same formula as the rest of this guide. That is 4,800 steps, each sending 3,000 tokens in and getting 800 tokens back, with no caching unless a row says otherwise. Your own steps may be larger or smaller, so treat each total as a worked example rather than a quote. Everything here was checked on 15 November 2026, and prices and offers change often.

The cheapest strong setups

The rows run from cheapest to dearest. Each model listed here can call tools, according to its company's model page or API reference, or the host's model page. The "Where it fits" column says which agents may use the row, because one row is allowed only in some agents.

Setup Light month How strong it feels (1-5, a judgment) Where it fits The catch
Google Gemini free key, gemini-4-hydrogen $0, within a daily limit Google sets for your project 4. The same score as in the full table. Any agent that accepts a Gemini key. Privacy risk. Google may use free prompts to improve its products. The daily limit is not published, so a long job can stop halfway.
MiniMax-M3 on MiniMax's own API about $9 4. MiniMax calls it a frontier coding model, and its API reference lists tool calls for it. Any agent that accepts an OpenAI-compatible address. MiniMax labels the price a "Permanent 50% off". This guide could not confirm MiniMax's data rules or a spend limit, so start with a small prepaid amount.
DeepSeek V4 Pro, deepseek-v4-pro, in off-peak hours about $17. About $34 if every step falls in DeepSeek's pricier hours. 4. The same score as in the full table. Any agent that accepts an OpenAI-compatible address. Safe official. You prepay a balance, so the balance is your limit.
Light hybrid: deepseek-flash for most steps, Claude Opus 5.5 for 1 step in 10 about $17. About $21 in DeepSeek's pricier hours. 4. The same score as in the full table. The agents that can mix models, as Mix a cheap model with a strong one shows. Safe official. Two companies see your prompts, so you need two spend limits.
GLM-5.3 on DeepInfra, at its discounted price about $18. About $28 at DeepInfra's list price. 4. DeepInfra's model page tags it for function calls. Any agent that accepts an OpenAI-compatible address. DeepInfra marks the price "38% off" and does not say when that ends. You must add a card or prepay first.
Z.AI GLM Coding Plan, Lite $18 flat 4, with GLM-5.3. Only the tools Z.AI lists. Of this guide's agents, that is OpenClaw, Hermes Agent, Goose, and Eigent. Ban risk in any other agent. No refunds.
Ollama's cloud plan, Pro $20 flat, which includes $60 of cloud use 4 with glm-5.3:cloud. A light month on it uses about $37 of the $60. Any agent that can use Ollama. Unused credits do not roll over. Ollama says prompts are never logged or trained on.
OpenAI's gpt-6.1-terra about $34 4. The same score as in the full table. Agents that support OpenAI's newer Responses connection. Surprise bill without a hard spend limit.
Claude Sonnet 5.5, with prompt caching on about $50, as an illustration. About $67 without caching. 4. The same score as in the full table. Any agent that accepts a Claude key. Caching needs the agent to turn it on. Surprise bill without a spend limit. The saving is an illustration, not a measurement.
Claude Opus 5.5 for every step, with prompt caching on about $100, as an illustration. About $134 without caching. 5. The cheapest row that rates a 5 for every step. Any agent that accepts a Claude key. Surprise bill without a spend limit. Claude Max's included API credit can cover some of it, as explained below.

Two strong models are left out on purpose. Qwen3.8-Max and Kimi K3.1 Pro cost about $43 and $64 for a light month on their cheapest official hosts, which is more than several rows above. The pages this guide opened also did not confirm tool calls for Qwen3.8-Max.

OpenRouter's free models rate a 3 in the full table, so they are not in this list, but one cheap step makes them far more useful. Their daily cap depends on the credits you have bought over the account's whole life, so a single $10 purchase lifts it from 50 to 1,000 requests a day for good. The $10 stays in your account as credit you can spend, although a balance below zero blocks the free models. Free models are often busy, so give the agent several :free models, or OpenRouter's openrouter/free router, which falls back only to other free models and accepts tool calls. A fallback list that includes a paid model is billed for it.

The flat-rate rows are cheap only while your agent stays inside each plan's rules, so read Flat-rate plans and their rules before you pay for one. Whatever you pick, set a limit first, as Set a spend limit first explains.

Free credits and trials

A trial or credit lets you try a strong model for nothing, but most of them end after a set time. Each one also comes with conditions that are easy to miss.

Skip the Google Cloud $300 trial for models. Google's page says the $300 credit cannot pay for the Gemini API in AI Studio. Anyone who signs up for the trial or upgrades on or after 16 November 2026 also cannot spend it on Google's Gemini platform, formerly Vertex AI. It cannot pay for partner models such as Claude on Google Cloud either.

If you are a student, researcher, startup, or open-source maintainer

Most readers will not qualify for these programs, because each one checks who you are. Apply only if you honestly fit the rules.

Cut the bill on any model

Two levers work inside a normal agent: prompt caching and DeepSeek's off-peak hours. Batch processing saves more, but only for jobs that can wait.

Prompt caching

An agent sends the same long opening, its instructions and its tool list, with almost every step. When that opening is stored, the provider charges a small fraction for it on the next step. Each company handles it differently, and caching never lowers the price of the reply.

Here is an illustration, not a measurement. Suppose 2,000 of each step's 3,000 input tokens repeat. A light month on Claude Opus 5.5 drops from about $134 to about $100, and Claude Sonnet 5.5 drops from about $67 to about $50. DeepSeek V4 Pro off-peak drops from about $17 to about $11, and DeepSeek Flash from about $4 to about $3.

Caching has a catch in the light hybrid. Claude Opus 5.5 handles only 1 step in 10, so its 5-minute store can expire between Opus steps. An Opus step that pays to store its opening every time costs about $0.030, which is more than the $0.028 it costs with no caching. This guide could also not confirm that any of its agents keeps the opening exactly the same each time, which caching needs. Check your provider's usage page after a week to see whether cached reads show up.

DeepSeek's off-peak hours

DeepSeek charges half price outside its peak hours, which are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. This needs no setting, so it is the only discount an ordinary agent gets for free. A job you schedule for the evening or the weekend in your time zone often falls off-peak, depending on where you live.

Batch processing, only for jobs that can wait

Anthropic, OpenAI, and Google charge half price for requests sent as a batch, and xAI takes 20% off some Grok models. The catch is time, because a batch can take up to 24 hours. An agent waiting for each reply cannot use this, so it fits only a nightly job that you build outside the agent.

OpenAI's Flex option charges the same half price for slower normal requests, which could suit an unattended overnight run. A light month on gpt-6.1-terra would be about $17 on Flex. OpenAI may refuse a Flex request when it is busy, and this guide could not confirm that any of its agents can turn Flex on.

Flat-rate plans and their rules

A flat-rate plan gives you a fixed allowance for a fixed monthly price. Each plan also has rules about which programs may use it, and those rules decide whether it is safe for your agent.

Plan Price a month Strong model Which agents its terms allow The catch
Z.AI GLM Coding Plan Lite $18. Pro and Max prices were not published on the pages that loaded. GLM-5.3 Only listed tools. Of this guide's agents: Goose and Eigent as coding tools, and OpenClaw and Hermes Agent as general-purpose agents. Agent Zero, Atomic Agent, NanoClaw, QwenPaw, Everfern, and OpenManus are not listed. Ban risk outside the list. Z.AI says accounts "with more than three violations may be banned". No refunds and no sharing.
Alibaba Model Studio Token Plan (Personal, Singapore) Lite $6, Essential $10, Standard $18, Pro $43 qwen3.8-max, glm-5.3, deepseek-v4-pro It names coding and agent tools, but for interactive use only. Ban risk. It bans automation and "non-interactive batch call scenarios", so a scheduled or unattended agent run breaks its terms. The credits each call uses are not published.
Kimi Code, on a Kimi plan Plus is ¥99 in China. Dollar prices were not published on the pages that loaded. Kimi K3.1 Pro on Plus and above No limit to certain tools was found. Kimi says third-party tools count toward the same quota. A rolling 5-hour limit plus a monthly total. This guide could not confirm tool calls for Kimi K3.1 Pro.
MiniMax M Plan Not published on the pages that loaded MiniMax M3.1 Flash. M3 being included was not confirmed. Any tool that speaks the OpenAI or Anthropic format. MiniMax names Hermes Agent and OpenClaw. MiniMax says the plan is for individual, interactive use. Pay-as-you-go MiniMax-M3 is already about $9 a light month.
Ollama Pro $20, with $60 of cloud use glm-5.3, deepseek-v4-pro, minimax-m3 No limit to certain tools was printed. Credits reset monthly. Kimi K3.1 Pro would cost about $67 of credits, more than the $60.
NanoGPT $12 Open models, but the exact versions are not named It is a relay, so any agent that accepts an OpenAI-compatible address. Privacy risk. Cached input counts in full against the 60 million tokens a week. It excludes Claude, GPT, Gemini, and Grok.
Cerebras Code Pro $50, Max $200 GLM 4.7, an older model Not checked, because both plans are sold out. Not a current option.

The GLM Coding Plan is the one most likely to save money, and it is within Z.AI's terms in OpenClaw, Hermes Agent, Goose, or Eigent. In any other agent, pay-as-you-go GLM-5.3 on DeepInfra costs about the same and breaks no terms. Kimi's membership page lists Hermes Agent among its integrations, and this guide found no Kimi wording that limits the plan to certain tools. Relays such as NanoGPT carry their own risks, which Relays and aggregators explains.

Set a spend limit first

A stuck agent can repeat a step all night, so set a limit before you paste a key anywhere. Pick the lowest limit that covers a normal week, and raise it only after you have seen real use.

Provider Control What happens at the limit Catch
Anthropic A hard Monthly spend limit you set below your account tier's cap Requests fail with an error until you raise it or the month resets. Safe official.
Claude Max or Ultra API credit The included credit is the limit when you have no other credits Requests stop until next month's credit arrives. Safe official, as long as auto-reload stays off.
OpenAI An optional hard limit, on top of alerts Requests fail with error 429. Surprise bill. OpenAI says enforcement is not instantaneous, and alerts alone do not stop spending.
Google Gemini A project spend cap that Google calls experimental, or a prepaid balance from $5 Paid requests stop at the cap, or at a $0 balance. Surprise bill. Billing lags about 10 minutes, and Google says agent sessions may run over.
DeepSeek A prepaid balance only Requests fail when the balance reaches zero. Safe official. The balance you top up is your limit.
OpenRouter A limit on each key, with a daily, weekly, or monthly reset Requests on that key are refused. Safe official. A key with no limit can spend your whole balance.

The step-by-step menus for each provider are in Set a spend limit at every provider.

Cheaper routes that break a provider's terms

Each route below is generally a contract matter between you and the provider rather than a crime, but each one is a potential breach of that provider's terms. The provider can respond by limiting or closing the account, keeping any balance on it, and refusing a refund. This section says how each route works and what the provider says it can do, so you can make the call yourself.

Where a step asks for a custom provider, the agent's own address field is the same one a relay uses. The general steps and the per-agent notes on the relays page show where that field lives in each agent.

The GLM Coding Plan in an agent Z.AI does not list

How it works. The plan uses its own address, and any agent with a custom provider field can point at it. In OpenClaw, Hermes Agent, Goose, and Eigent the same steps stay within Z.AI's terms. In an unlisted agent such as Agent Zero, they break them.

  1. Subscribe to the GLM Coding Plan and copy your Z.AI API key. Z.AI's setup page asks only for "your Z.AI API Key", and this guide could not confirm whether the plan uses a separate key.
  2. In the agent, add a custom OpenAI-compatible provider with the base URL https://api.z.ai/api/coding/paas/v4. For an Anthropic-compatible provider, Z.AI lists https://api.z.ai/api/anthropic instead.
  3. Enter the GLM model code you want to use. Z.AI's setup page gives glm-5.2 as its example, and this guide could not confirm the exact code for GLM-5.3 on that page.
  4. Send a short test message, then check that the plan's usage moved. Z.AI warns that a wrong address means the request cannot use the plan's quota.

The risk. This breaks Z.AI's terms in any unlisted agent. Z.AI says users "may not use their subscription benefits for tools or scenarios outside of this scope". Its usage policy lists rate limiting and account freezing, says accounts "with more than three violations may be banned", and says refunds are not supported.

Alibaba's Token Plan in an agent that runs on its own

How it works. The Token Plan has its own key and its own address, which any agent with a custom provider field can use.

  1. Subscribe to Token Plan Personal Edition, then use Create or view API Key on the plan's console page.
  2. Copy the base URL from that console page. Alibaba's tools page lists https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1 for the OpenAI format, but its plan overview says the only region is Singapore. This guide could not confirm the Singapore address on a page that loaded, so use the one your console shows.
  3. Add a custom OpenAI-compatible provider in the agent, and enter a model ID from the plan's list, such as qwen3.8-max, glm-5.3, or deepseek-v4-pro. Alibaba says its auto model does not support tool_choice, so pick a named model.

The risk. Chatting with an agent while you watch may fit the plan, because Alibaba names agent tools. A scheduled job or a long unattended run breaks its terms, because Alibaba says the plan "must not be used for automation scripts, custom application backends, or any non-interactive batch call scenarios". Out-of-scope use "may result in subscription suspension or API Key banning". This guide could not confirm where Alibaba draws the line for a long agent job.

MiniMax's M Plan for unattended work

How it works. MiniMax says any tool that speaks the OpenAI or Anthropic format works, and it names Hermes Agent and OpenClaw.

  1. Subscribe, and copy the M Plan API key, which starts with sk-cp-, from Plan Details in the MiniMax console.
  2. Add a custom provider with the base URL https://api.minimax.io/v1 for the OpenAI format, or https://api.minimax.io/anthropic for the Anthropic format.
  3. Enter the model ID MiniMax-M3.1-Flash. That is MiniMax's cheaper model, and this guide could not confirm that the plan includes MiniMax-M3.

The risk. Using it interactively in any agent fits MiniMax's wording. MiniMax says the plan "is designed for individual, interactive developer use" and that "Pay-as-you-go is recommended for production use". So scheduled or unattended runs sit outside what the plan is designed for, although the FAQ names no specific penalty.

A personal Claude or ChatGPT plan in a third-party agent

How it works. The steps are already on Use a chat plan you already pay for, so this page does not repeat them. See OpenClaw reusing a Claude Code sign-in, Hermes Agent with Claude Max, and CLIProxyAPI.

The risk. Anthropic's Usage Policy forbids services that route requests through consumer plans. Its login page also says it may bill third-party use as usage credits that you pay. OpenAI's Sign in with ChatGPT terms forbid general API access for other tools, which is the shape a local proxy such as CLIProxyAPI takes. Google's Gemini CLI and Antigravity sign-ins are a separate case, with their own steps and Google's ban history, on that page's Google section.

Sharing one plan between several people

How it works. One person signs up and shares the sign-in or the key with others.

The risk. Z.AI says "Account sharing or multi-user access is prohibited". Alibaba says sharing "may be considered a violation and result in suspension", and OpenAI forbids pooling plan usage. You also lose control of the bill, because everyone spends the same allowance.

Discount relays

Some relays resell access below the maker's list price, and their own accounts can be closed, leaving your balance stuck. Relays and aggregators has the steps and the warnings.

Not covered here

Some cheap routes go beyond a broken contract into fraud or unauthorized access, so this guide leaves them out entirely. That includes fake student or identity details, and making extra accounts to claim new-user credit again or to multiply free limits, which means getting something of value by pretending to be a new customer. It also includes stolen, leaked, or resold keys, scrapers such as gpt4free that use other people's services without permission, and chargebacks or card tricks.

Trade-offs: safety versus capability

A flat-rate plan or free credit, or pay-as-you-go. This is the choice between a fixed monthly price and paying only for what your agent uses.

For the three questions behind these choices, see Safety versus capability.

Sources

Checked 15 November 2026. Each page and what it said is recorded in the cheap access research notes, and the light-month prices come from the free and cheap APIs chapter.

Next