Directory
Which AI model should your automations run on?
An AI agent is the cheapest hire you will make, and the model is its wage. These pages cover what each provider charges per million tokens, how well it holds up on tool-calling agent work, how you can access it, and what its data policy actually says. Prices are list prices as of August 2026.
What this collection is for
Ten providers, one page each, all answering the same question: should a small business run its automations on this model? Four are American - OpenAI, Anthropic Claude, Google Gemini and xAI Grok. Four are Chinese - DeepSeek, Moonshot Kimi, Z.ai GLM and Alibaba Qwen. The remaining two - Mistral and Meta Llama - matter mostly because you can run the weights yourself.
Skip this directory if you are using a no-code builder that bundles the model and hides the token bill; in that case the decision that matters is the platform, not the provider, and the platforms directory is your page. Skip it too if you have not yet decided what to automate - the model is one of the last choices in a build and one of the cheapest to change, so making it first is wasted effort. Start at the ROI calculator or an agent playbook instead.
Most small businesses never need the most expensive model on the market. A well-scoped automation - triaging an inbox, drafting quotes, tagging leads, summarising calls - spends most of its tokens on routine reading and writing, and the budget tier of almost every provider handles that at a fraction of flagship cost. Where you do want a frontier model is the small share of steps that involve judgement, long multi-step tool use, or anything a customer sees unedited.
The practical pattern is a two-model stack: a cheap model for volume and a strong model for the hard step, with the workflow deciding which one runs. That is how we keep monthly model spend on a typical build in the tens of dollars rather than the hundreds. The AI automation cost page breaks down where the rest of the money goes, and our automation services cover the build and handover.
Every page below answers the same question - should a small business run its automations on this model? - with current pricing, honest weaknesses, hosting options including self-hosting where the weights are open, and a factual read of the provider's data policy.
How to choose, in the order that actually matters
Start with the constraint, not the leaderboard. If the automation touches regulated or client-confidential material, the provider's retention and training policy decides the shortlist before any benchmark does - that is the first thing each page below states. If it does not, cost per run and tool-calling reliability decide it, in that order.
Then price the workload rather than the token. A model at half the headline rate is no cheaper if it needs two attempts to return valid JSON, and a long system prompt resent on every call quietly dominates the bill in a chatty workflow - which is why prompt caching often saves more than switching provider. Work out the monthly figure from your own volume before comparing anything: the cheapest LLM API comparison models three realistic workload shapes end to end.
Reliability beats raw capability for automation. An agent that calls tools correctly 99% of the time and fails loudly is worth more than a smarter one that occasionally invents a parameter, because the second failure mode is silent and reaches your customer. Test with your own awkward cases - the messy forwarded email, the half-filled form - not with a benchmark suite.
Finally, keep the model swappable. Route every call through one place in the workflow so changing provider is a config edit, not a rebuild. Prices and limits in this directory have already moved more than once since we published it, and treating the model as a replaceable part is what stops that from being expensive. If you would rather not maintain that yourself, we pick the stack and hand it over documented on a free automation audit.
One caveat on everything below: the open-weight models can be self-hosted, which changes the arithmetic completely - you trade a per-token bill for a GPU bill and the work of running it. That is worth it at sustained volume and rarely worth it below it. Each provider page says which side of that line it sits on, and the missed-call build guide shows what a real workflow's model spend looks like at small-business volume.
Put concretely, the four questions and where each one lands:
- Is the data sensitive or contractually restricted? Then the data residency and zero-retention lines decide it before any benchmark. In practice that points at the enterprise tiers of OpenAI, Anthropic or Google, at European hosting via Mistral, or at self-hosting Llama. Each page states its provider's policy factually rather than reassuringly, and the DPA entry explains what to ask for.
- Is cost the binding constraint? Then the budget tiers of the Chinese providers - DeepSeek, Z.ai GLM, Qwen, Kimi - and the small models from the American labs are the field. The table below ranks the cheapest listed model per provider, and DeepSeek versus GPT covers what you give up.
- Does the agent call tools? Then tool-calling reliability outranks price, because a retry costs more than the token saving and a silently invented parameter costs more than both. Test with your own awkward inputs, not a benchmark, and read MCP versus function calling for how the tools get exposed in the first place.
- Is the volume high and steady? Only then does self-hosting an open-weight model beat a per-token bill, because you are trading it for a GPU bill and the work of running it. Below sustained volume it rarely pays. The self-hosting entry and the context window entry between them explain most of what drives the number.
What every provider charges, read on 6 August 2026
One row per provider, showing the cheapest tier we could verify from that provider's own published pricing page during this pass. It is deliberately not the flagship: the flagship is a rounding error away from irrelevant for the kind of work most businesses automate, and the cheap tier is what a well-built workflow actually runs on. Every number below was fetched on 6 August 2026, and the source list under the table lets you re-check any single figure without taking our word for it.
Two health warnings before you read it. First, these are list prices, so they are the ceiling rather than the bill - caching and batching, covered further down, move the real figure more than switching provider usually does. Second, the tiers are not equivalent in capability. gpt-5-nano and Claude Sonnet 5 are in the same table because businesses genuinely choose between them, not because they will produce the same output on a hard tool-calling step. Price is one axis. Read the provider page for the other one.
| Provider | Cheapest tier verified | Input / 1M | Output / 1M | Cached input / 1M | Based in |
|---|---|---|---|---|---|
| OpenAI | gpt-5-nano | $0.05 | $0.40 | $0.005 | United States |
| Anthropic | Claude Sonnet 5 (promotional rate to 31 Aug 2026) | $2.00 | $10.00 | Not re-verified | United States |
| gemini-3.1-flash-lite | $0.25 | $1.50 | $0.025 | United States | |
| xAI | grok-build-0.1 | $1.00 | $2.00 | $0.20 | United States |
| DeepSeek | deepseek-v4-flash | $0.14 | $0.28 | $0.0028 | China |
| Moonshot | kimi-k3 | $3.00 | $15.00 | $0.30 | China |
| Z.ai | GLM-4.7-FlashX | $0.07 | $0.40 | Not published as a rate | China |
| Alibaba Qwen | qwen-flash, shortest input band | $0.05 | $0.40 | 10% of input | China |
| Mistral | Ministral 3 - 3B | $0.10 | $0.10 | 90% off input | Europe |
| Meta Llama | llama-3.1-8b-instant, hosted by Groq | $0.05 | $0.08 | Not published | Open weights |
Read on 6 August 2026 from: OpenAI (developers.openai.com/api/docs/pricing), Google (ai.google.dev/gemini-api/docs/pricing), xAI (docs.x.ai/docs/models), DeepSeek (api-docs.deepseek.com/quick_start/pricing), Moonshot (platform.kimi.ai/docs/pricing), Z.ai (docs.z.ai/guides/overview/pricing), Alibaba Cloud Model Studio billing documentation, Mistral (mistral.ai/pricing/api), and Groq's model list for the hosted Llama rate. The Anthropic row is the Claude Sonnet 5 rate published on the AWS Bedrock pricing page, which also carries Anthropic's note that the promotional $2.00 / $10.00 rate runs to 31 August 2026 and reverts to $3.00 / $15.00 after it. Anthropic's own pricing pages would not return a valid certificate from this environment during the pass, so the cheaper Claude Haiku tier is not re-verified here and is deliberately left out rather than restated from memory.
The same workload, priced on all ten
Headline rates are useless on their own because input and output are priced differently and every workload has a different ratio between them. So this site uses one reference workload everywhere it talks about model cost: an inbox triage agent handling 1,000 items a month, at roughly 3,000 input and 400 output tokens per run. That is a real shape - it is what the email assistant playbook and the customer support agent both look like once you count the system prompt and the retrieved context.
The arithmetic is deliberately boring. 1,000 runs at 3,000 input tokens is 3.0 million input tokens a month. 1,000 runs at 400 output tokens is 0.4 million output tokens. So the monthly bill is three times the input rate plus four tenths of the output rate, and the table below is nothing but that sum applied to the row above it. No customer data, no measured deployment, no benchmark - just the published rate times the volume.
| Provider | Model priced | Cost / month | What the number hides |
|---|---|---|---|
| Meta Llama | llama-3.1-8b-instant on Groq | $0.18 | Cheapest of the lot, and the weights are open if you outgrow the host |
| OpenAI | gpt-5-nano | $0.31 | Cheapest US major by a wide margin; cached input takes it lower still |
| Alibaba Qwen | qwen-flash | $0.31 | Only at this rate while the prompt stays inside the shortest input band |
| Mistral | Ministral 3 - 3B | $0.34 | Flat in and out, so long answers do not punish you |
| Z.ai | GLM-4.7-FlashX | $0.37 | GLM-4.7-Flash is listed free, which makes this a test-bench provider |
| DeepSeek | deepseek-v4-flash | $0.53 | DeepSeek's own page warns of a significant price rise coming |
| gemini-2.5-flash-lite | $0.46 | Google's cheapest listed tier, about 1.5x OpenAI's; the newer 3.1 Flash-Lite costs more, not less | |
| xAI | grok-build-0.1 | $3.80 | Input-heavy pricing, so triage is the wrong shape of job for it |
| Anthropic | Claude Sonnet 5 | $10.00 | At the promotional rate; $15.00 at the standard rate from 1 September 2026 |
| Moonshot | kimi-k3 | $15.00 | A flagship priced like one; the cheaper K2.x line was not verifiable this session |
Read the spread rather than the ranking. Eight of the ten providers run this workload for under four dollars a month, and six of them for under a dollar. At that scale the choice between the cheap tiers is not a cost decision at all, because the difference between $0.31 and $0.53 a month is not worth an afternoon of anyone's time. What the difference does tell you is how much room you have when the workload grows: a provider that is four times dearer at 1,000 runs is still four times dearer at 100,000, and that is where it starts to matter.
Step up a tier and the arithmetic changes shape. The same workload on a mid or frontier model costs roughly $1.08 on gpt-5.6-luna, $1.65 on deepseek-v4-pro, $4.28 on GLM-5, $7.50 on gemini-3.6-flash or Mistral Medium 3.5, $8.40 on grok-4.5, and $10.00 on Claude Sonnet 5 at its promotional rate. Those are all still small numbers against a wage, which is the honest headline: the model is almost never the expensive part of an automation. The platform subscription, the integration work and the maintenance are bigger, which is what automation build cost and what AI automation costs both set out with the recurring lines broken up. If you want that modelled against your own headcount rather than ours, the ROI calculator and the n8n workflow cost calculator do the sums.
Two things the table cannot show. Reasoning models bill their internal thinking as output tokens, so a reasoning tier at the same listed rate can cost several times more per run than the row suggests. And retries are pure waste: a model that returns malformed JSON one time in twenty adds five percent to the bill and rather more to your debugging. That is why evals on your own awkward inputs beat a price comparison for deciding anything.
Caching and batch discounts, which move the bill more than list price
If your workflow resends the same system prompt, the same tool definitions and the same few-shot examples on every call - and almost every workflow does - then most of your input tokens are identical every time. Prompt caching bills those repeated tokens at a fraction of the normal rate. The discounts are large enough to reorder the table above: OpenAI publishes cached input at 90 percent off, Mistral publishes the same 90 percent, DeepSeek lists a cache hit on deepseek-v4-flash at $0.0028 against $0.14 on a miss, and Moonshot lists a Kimi K3 cache hit at $0.30 against $3.00. On a chatty agent with a heavy system prompt, caching is worth more than any provider switch you could make.
Batch modes are the other lever and apply whenever the answer is not needed in the next few seconds. OpenAI, Google and Mistral each publish a flat 50 percent discount for batch processing. A great deal of business automation is genuinely batch work built as real-time out of habit: overnight document processing, database enrichment, a month of summaries. If nobody is waiting, halve the bill.
Now the traps, all three of which have caught us or someone we read. The first is prepositions. Moonshot documents batch inference as costing 60 percent of the standard price. That is a 40 percent discount, not 60 percent off, and it is worse than the flat 50 percent the American labs publish. We have restated that one wrong before, so read the sentence, not the number.
The second is stacking. Alibaba's billing documentation states that batch pricing and the cache discount are mutually exclusive on Qwen: you take one or the other, never both. That is not true of the providers where the two multiply together, so a spreadsheet that assumes stacking will overstate Qwen's advantage by a wide margin. For a chatty agent the cache is usually worth more; for an overnight backlog, batch wins.
The third is that a cache costs something to create. Cache writes are typically billed above the standard input rate, so a cache that is hit once loses money and a cache that is hit ten times is close to free input. Short-lived caches expire faster than most people assume, which means a workflow that fires four times an hour may never get a hit at all. Check the time-to-live on the provider page before you design around it, and read the cheapest LLM API comparison for the same arithmetic run across three workload shapes end to end.
One more constraint worth knowing before you optimise: the cheapest route and the strictest privacy posture can be mutually exclusive. Batch endpoints often retain job data for a fixed window, which puts them outside zero data retention arrangements. If a workflow is regulated, price it without batch and accept the difference.
Pick a tier per step, not per workflow
The single most common mistake we see is choosing one model for a whole automation. A workflow is not one job. Triaging an inbox is five or six distinct steps, and they have wildly different difficulty. Classifying an email into one of eight buckets is trivial and a budget tier does it perfectly. Deciding whether a customer sounds angry enough to escalate is judgement. Drafting the reply that customer will actually read is the only step where a weak model is visibly, embarrassingly worse.
So route by step. A workable default across the builds we ship looks like this, and it is worth writing down before you open a workflow editor.
- Extract and classify: budget tier, always. Pulling a phone number, a date, an address or a category out of unstructured text is a solved problem at $0.05 per million tokens. Constrain the output to a fixed schema and validate it in the workflow rather than trusting the model.
- Summarise: budget tier, with a length limit. Summaries are read by a human who will notice if they are wrong, which makes this the safest place to be cheap.
- Decide and route: mid tier if the decision has consequences. This is the step where a wrong answer sends a quote to the wrong person, and where guardrails and a human in the loop earn their keep more than a better model does.
- Call tools in a loop: whichever tier is most reliable at function calling, because a retry costs more than the token saving and a silently invented parameter costs more than both. Test this yourself with the messy forwarded email and the half-filled form.
- Write anything a customer reads unedited: the best tier you are willing to pay for, or a human. This is the only step where model quality is a brand risk rather than a cost line.
Because the expensive step is usually a small share of runs, a two-model stack costs barely more than the cheap model alone. If one in ten triaged emails escalates to a drafted reply on a mid tier, the reference workload moves from about $0.31 to well under a dollar. That is the whole trick, and it is why we quote model spend on a typical build in tens of dollars a month rather than hundreds. The n8n platform page and the LangGraph page both cover how the routing is actually wired, and MCP versus function calling covers how the tools get exposed in the first place.
Seven ways to keep the token bill down
In rough order of how much they save on a typical build, and none of them require changing provider.
- Stop sending the whole record. Most workflows paste an entire CRM object or email thread into the prompt because it is easier than selecting fields. Sending the four fields the step actually needs often cuts input tokens by more than half, and it improves accuracy because there is less to get distracted by.
- Cache the stable prefix. Put the system prompt, tool definitions and examples first and the variable content last, so the repeated part is cacheable. This is a five-minute change with a permanent discount attached.
- Cap the output. Output tokens cost several times input on nearly every provider in the table. A max-tokens limit plus a prompt that asks for structured fields rather than prose is the cheapest quality improvement available.
- Filter before the model, not with it. If a rule can decide it, a rule should decide it. Auto-replies, out-of-office notices and known-vendor mail never need to reach an LLM at all, and dropping them at the trigger is free.
- Batch anything nobody is waiting for. Half price at three of the providers above, with the caveats in the previous section.
- Retrieve less, better. If the workflow uses retrieval, a reranking step that cuts twenty retrieved chunks to four usually improves the answer and cuts the input bill at the same time.
- Log spend per run from day one. You cannot manage what you do not measure, and a runaway loop is a four-figure surprise. Basic observability with a hard spend cap is the difference between a cost problem you notice on Tuesday and one you notice on the invoice.
Data privacy and residency, provider by provider
This section is factual and deliberately unexciting. The question is not whether a provider is trustworthy in the abstract, it is what its published terms say about retention, training and where the processing happens - and whether that is compatible with what your workflow touches. Read it as a shortlist filter, not as a verdict.
The American providers - OpenAI, Anthropic, Google and xAI - are the default answer for regulated work in the US, mostly because their API terms, retention windows and data processing agreements are published, negotiable at the top end, and familiar to the lawyers reviewing them. Zero-retention arrangements exist on three of the four, usually gated behind an application and sometimes incompatible with batch endpoints and certain premium models. We could not verify xAI's current API data terms from public sources during this pass, which is a statement about what is published rather than an allegation about their handling.
The Chinese providers - DeepSeek, Moonshot Kimi, Z.ai GLM and Alibaba Qwen - are the cheapest serious options on the market and they are not equivalent to each other on this axis. Alibaba is the easiest of the four to justify to a cautious stakeholder: Model Studio publishes an explicit pledge not to train on your data and runs international deployments in Singapore, Hong Kong, Germany, the US and Japan, so the processing location is a configuration choice rather than an assumption. The others vary in how much they publish and where they process. Several are also available through Western hosting providers running the open weights, which changes the answer entirely, because then the model is Chinese and the processor is not.
The rule we apply on client work is simple and has nothing to do with geography. If the material is regulated, client-confidential or covered by a contract that names sub-processors, the data residency and DPA position decides the shortlist before any benchmark or price does. That rules a lot of providers out of a law firm or a medical clinic build regardless of price, and rules almost nothing out for a cleaning company classifying inbound enquiries. Most workflows are the second kind.
The open-weight route - Mistral and Meta Llama - is the one that removes the question rather than answering it, because the weights run on infrastructure you control. Mistral is European and matters most where the processing has to stay in the EU. Meta does not publish per-token API pricing for Llama at all, so every Llama figure on this site is a third-party host's rate and is labelled as such. Neither is free: self-hosting swaps a per-token bill for a GPU bill plus the work of running it, and the true cost of self-hosted n8n guide prices exactly that trade on the platform side.
Whichever way it lands, write the answer down before you build. Which provider, which region, what retention, and who signed off. That paragraph is the thing a procurement review asks for, and producing it after the fact costs far more than producing it first. It is also part of what a free automation audit hands back in writing.
Where this sits next to platforms and compare
Three directories get confused with each other. This one is which model does the reasoning, and what does it cost. Platforms is what you build the automation on - n8n, Zapier, Vapi, LangGraph and the rest, reviewed one at a time. Compare is A or B, once you already have two candidates of either kind. The two axes are independent: you choose a platform, then point it at a model, and the whole point of routing every call through one place is that the second choice stays cheap to revisit.
Upstream of all three sits the question of whether to automate the workflow at all, which is what the free tools and the agent playbooks are for. Model spend is usually the smallest line in a build; the platform subscription, the integration work and the maintenance are larger, which is what automation build cost sets out.
What to do first
Take one real workflow and estimate its shape before you look at a single price: how many calls a week, how long the input is, how long the output is, and how much of the prompt repeats every time. That last figure is the one that surprises people, and it is why token counts beat headline rates. Then run those numbers through the cheapest LLM API comparison, which models three realistic workload shapes end to end.
Then pick two providers, not one: a cheap model for volume and a stronger one for the step that needs judgement. Wire both behind a single call site so switching is a config edit. Prove it on something small - the quote follow-up build is a good first test because its failures are visible and cheap. If you would rather not maintain any of this, the free automation audit returns the stack and model recommendation in writing, and our automation services cover the build and handover.
United States
OpenAI
The default API for business automation, with a budget tier cheap enough for high-volume work.
Budget tier from $0.20 in / $1.20 out per 1M tokens
Anthropic Claude
The model teams reach for when the automation touches long documents, code, or anything that must not go off the rails.
Cheapest tier from $1.00 in / $5.00 out per 1M tokens
Google Gemini
The cheapest of the US majors at the low end, and the obvious pick if your business already runs on Google Workspace.
From $0.10 in / $0.40 out per 1M tokens on older Flash-Lite
xAI Grok
Very cheap long-context pricing and a dedicated agent model, held back by thin public documentation on data handling.
Grok 4.3 at $1.25 in / $2.50 out per 1M tokens, 1M context
China
DeepSeek
The price floor for capable agent models, with MIT-licensed weights and a China-only data policy you have to read before you commit.
V4-Flash from $0.14 in / $0.28 out per 1M tokens
Moonshot AI (Kimi)
Long-horizon agent models on a Singapore endpoint, priced well for coding work but training on your data by default.
K2.5 from $0.60 in / $3.00 out per 1M tokens
Z.ai (GLM)
MIT-licensed models, a Singapore endpoint, and a published no-storage claim that is stronger than most Western providers offer.
GLM-4.7-FlashX at $0.07 in / $0.40 out per 1M tokens
Alibaba Qwen
The clearest no-training pledge of any Chinese provider, a Singapore endpoint, and genuinely self-hostable open models.
qwen-flash from $0.05 in / $0.40 out per 1M tokens
Questions people ask
What is the difference between a model and a platform?
A platform is what you build the automation on, such as n8n, Zapier, Vapi or LangGraph. A model is the LLM that platform calls to do the reasoning, billed per million tokens. You pick a platform first, then point it at a model, and the model should stay swappable behind one call site.
Which model should a small business start with?
A budget-tier model from any major provider handles the bulk of routine automation work - triaging an inbox, drafting replies, tagging leads, summarising calls. Reserve a frontier model for the small share of steps involving judgement or anything a customer sees unedited.
Is the cheapest listed price the price I will pay?
No. The table shows the cheapest tier we could verify per provider at list price. Cache hits, batch modes and discounted tiers go lower, while long system prompts resent on every call, retries and reasoning tokens push the real figure higher. Estimate from your own call volume and prompt length rather than the headline rate.
What does the reference workload cost per provider?
The workload is an inbox triage agent doing 1,000 runs a month at about 3,000 input and 400 output tokens each, so 3.0M input and 0.4M output tokens. On list prices verified on 6 August 2026 that is $0.18 a month on Llama 3.1 8B hosted by Groq, $0.31 on gpt-5-nano, $0.53 on deepseek-v4-flash, $1.35 on gemini-3.1-flash-lite and $10.00 on Claude Sonnet 5 at its promotional rate. The spread across the cheap tiers is cents.
Do batch and caching discounts stack?
Not everywhere, and the wording is where people get caught. OpenAI, Anthropic, Google and Mistral all publish a flat 50 percent batch discount. Alibaba documents batch pricing and the cache discount as mutually exclusive, so you take one. Moonshot documents batch as costing 60 percent of standard price, which is a 40 percent saving and not 60 percent off. Read the preposition, not the number.
Are the Chinese providers safe to use for business data?
That is a data-policy question rather than a quality one, and the answer depends on what the workflow touches. Each provider page states the retention and training policy factually. Where the material is regulated or client-confidential, that line decides the shortlist before any benchmark does.
Should I self-host an open-weight model?
Only at sustained volume. Self-hosting trades a per-token bill for a GPU bill plus the work of running it, and below steady load it rarely pays. The Mistral and Meta Llama pages say which side of that line each sits on.
How current is the pricing here?
Every figure in the two tables was read from the provider's own published pricing page on 6 August 2026, and the source URLs are printed under the table so you can re-check any single number. Rates have already moved more than once since we started publishing, and DeepSeek's page currently warns of a significant rise. Treat the model as a replaceable part and check the provider page before you commit.
