AI agent cost per month: what 24/7 really costs, by model

A light personal AI agent on a cheap model costs a few dollars a month in tokens: about $4 at list price on DeepSeek V4 Flash or GPT-5.6 Luna for the usage we model below. Put the same agent on Claude Opus 5.5 or GPT-5.6 Sol and it's closer to $70. A busy team bot on a premium model lands somewhere between $250 and $550 a month, and an automation-heavy agent can go past $1,300.

The server is the small part. The bill is decided by the model you pick and how much text your agent resends every time it thinks. Below is our cost model with every assumption written down, so you can swap in your own numbers. Prices were checked on each provider's own pricing page on 2 November 2026.

Why does an AI agent use so many more tokens than a chatbot?

In a chat window you type, the model answers, and the conversation grows a little. An agent calls the model over and over, and each call carries far more than your last message:

• The system prompt and persona. OpenClaw rebuilds its system prompt on every run from the tool list, the skills list and workspace files such as AGENTS.md, SOUL.md, USER.md and MEMORY.md. By default it allows up to 20,000 characters per injected file and 60,000 in total, according to its token use docs.
• Memory and skills. Hermes Agent puts MEMORY.md (capped at 2,200 characters, about 800 tokens) and USER.md (1,375 characters, about 500 tokens) into the system prompt at the start of each session, per the Hermes memory docs.
• Tool definitions. Each tool's schema is input too.
• Tool results. A web page, log file or search result the agent reads comes back as input on the next call. OpenClaw caps a single tool result at 16,000 to 64,000 characters, depending on the model's context window.
• Multi-step loops. "Check my calendar and draft replies" might take five model calls, each resending everything above plus the previous step's output.
• Scheduled jobs. On first setup, OpenClaw creates a proactive check that runs every 30 minutes in the agent's main conversation, as its heartbeat docs describe. That's 48 model calls a day before anyone says a word.

Output tokens are priced three to six times higher than input. But an agent reads far more than it writes, so input is where most of the money goes.

How many tokens does one agent turn use?

Here is an illustrative budget for one model call in the middle of a task. It's our assumption, not a measurement, and yours will differ. In OpenClaw, /context list shows what each injected file costs and /usage tokens adds token and cache details to every reply, so you can swap our guesses for real figures quickly.

Your message is a sliver, and input dominates the cost even though output is priced five times higher on Sonnet 5.

Token counts also vary by model. Anthropic's pricing page says Claude 4.7 and later produce roughly 30% more tokens for the same text. We use one set of counts for every model, so read the Claude rows as a floor.

Prompt caching: why resent context gets cheaper

With prompt caching, the provider keeps the start of your prompt for a while. When the next request begins with the same tokens, they're billed at a much lower cache-hit rate. Agents are close to the ideal case, since most of each call repeats the last one.

Three things break it. First, caches expire: Anthropic's standard cache lasts five minutes (a one-hour option costs 2× the input price to write), and Kimi's default is also five minutes. Second, changing anything near the start of the prompt invalidates everything after it, which is why Hermes freezes its memory snapshot for the session (its docs say this "preserves the LLM's prefix cache"). Third, each model has its own cache, so switching models mid-conversation starts cold.

Some providers also charge to write the cache. Anthropic bills five-minute writes at 1.25× input, and OpenAI lists 30-minute writes for GPT-5.6 at 1.25× too. DeepSeek and MiniMax list no write fee, Z.ai lists cache storage as free for a limited time, and Google lists an hourly storage price for context caching.

Here are the list prices we used, per million tokens, as of 2 November 2026. The last column is our own AI API price list, kept apart from the official numbers.

Sources: DeepSeek, Z.ai, Alibaba Cloud Model Studio, MiniMax, Kimi, Google Gemini API, OpenAI, Anthropic, all checked 2 November 2026. ¹ Peak prices; off-peak is half. Peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays. DeepSeek's own API now serves V4 Flash requests with V4.1 Flash (1017); our list carries V4-Flash-0731, so these are different versions. ² Alibaba's docs say the qwen3.8-max cache-hit price isn't the usual 10% and refer to the console; we couldn't verify it, so cached tokens are billed at full price here. ³ Closed to new sign-ups since 31 October 2026, retiring by the end of December. ⁴ Rises to $1.50 / $7.50 (cache $0.15) on 1 January 2027. Our list also has Gemini 3.7 Flash at $0.15 / $0.75.

How to estimate your own monthly bill

You need four numbers about your agent and three prices for the model. The maths is plain multiplication:

monthly input  = calls per day × input tokens per call × 30
monthly output = calls per day × output tokens per call × 30

cost = uncached input × input price
     + cached input   × cache-hit price
     + output         × output price      (prices per 1M tokens)

Worked example, scenario A below on DeepSeek V4 Flash: 60 calls a day at 12,000 input tokens is 21.6M input tokens a month. Half are cache hits. So 10.8M × $0.30 = $3.24, plus 10.8M × $0.006 = $0.06, plus 0.72M output × $1.20 = $0.86. Total: $4.17.

Three adjustments. Anthropic and OpenAI's GPT-5.6 charge 1.25× input to write the cache, so we billed their whole uncached share at that rate, a deliberate slight overestimate. Qwen3.8 Max gets no cache discount, which is why it looks expensive in B and C. DeepSeek is at peak price throughout. Our HostAiHub prices have a single input rate, so that column has no cache discount.

Three usage scenarios, with the assumptions written down

The cache share rises with activity on purpose. An agent that calls the model every few minutes keeps its cache warm, while one that wakes up hourly often finds it expired. Scenario A is meant to resemble a personal Hermes agent on Telegram. Scale the input per call to your own persona and traffic.

Monthly cost by model

US dollars per month (30 days), rounded. Official columns apply caching as described above; HostAiHub columns use our flat prices.

What's the cheapest LLM for an AI agent in 2026?

At list price, GPT-5.6 Luna and DeepSeek V4 Flash are neck and neck at $29 to $30 for the team bot, with MiniMax M3 at $38 and Gemini 3.6 Flash about three times that. The flagship tier (Gemini 3.1 Pro, Claude Sonnet 5, GPT-5.6 Terra) clusters between $250 and $290, and Claude Opus 5.5 and GPT-5.6 Sol sit above $500.

On "DeepSeek vs GLM vs Claude": for the same workload, DeepSeek V4 Pro comes to $125, GLM-5.3 to $167 and Claude Sonnet 5 to $272. DeepSeek's tiny cache-hit price does much of the work. Opus 5.5's cache-hit price is $0.20, the same as Sonnet 5's, so on heavily cached traffic the Opus premium is mostly output and fresh input.

Things that move these numbers

• DeepSeek off-peak. Off-peak is half price. Traffic spread evenly over the week blends to about 0.6× peak; entirely off-peak, the team bot costs $15 on V4 Flash and $62 on V4 Pro.
• Scheduled changes. Gemini 3.6 Flash doubles on 1 January 2027, and Kimi's K2 models are due to retire by the end of December.
• Newer OpenAI models. OpenAI's page now leads with a GPT-6 family, and its Luna and Terra tiers are cheaper than GPT-5.6's. Same method: the team bot is about $14 on gpt-6-luna and $128 on gpt-6.1-terra. We modelled GPT-5.6 because it's on our own list.

How the HostAiHub column works

Our AI API is prepaid credit behind one OpenAI-compatible endpoint (https://hostaihub.com/v1), with one input price per model and no separate cache rate. On these assumptions it comes out below list price for every model we both carry. The gap narrows as your cache share rises, though, because provider cache discounts grow while our input price stays flat. One counter-example: an automation-heavy agent on DeepSeek V4 Pro running entirely off-peak would pay DeepSeek about $155 a month, against $209 with us. Run your own numbers.

The bill nobody sees: heartbeats and scheduled jobs

An agent with a scheduled check runs the model whether or not anyone talks to it. Take OpenClaw's default proactive check, every 30 minutes in the main conversation, and assume each one carries 20,000 tokens of context and replies with 100. That's 48 calls a day and 28.8M input tokens a month from checks alone.

A 30-minute gap is longer than a five-minute cache, so on many providers each check pays full input price. Here's what that costs per month with no cache hits:

Moving that check from every 30 minutes to every two hours drops the Sonnet 5 figure from $73 to $18. DeepSeek lists no cache-write fee, and with 90% of each check served from its cache, the V4 Flash figure falls to $1.19. OpenClaw's docs also suggest returning NO_REPLY when nothing has changed, which skips an extra final-answer request, and using activeHours so checks don't run overnight. They also suggest a 55-minute interval to keep a one-hour cache warm, then warn that every turn costs tokens and you should compare total usage rather than assume a cache hit.

Price is the second filter, not the first

A cheap model that fumbles tool calls isn't cheap. A failed call usually means another attempt with the error in context, which is another full-price call, and a confused agent can loop several times. Check first that the model calls tools reliably, follows instructions over a long session, and handles the context length your agent actually uses. Price comes after that.

The practical test is boring and it works. Run 20 or 30 real tasks from your agent's history on two or three candidate models and read the transcripts. Count the failures, then the token usage. A mid-priced model that finishes in fewer calls can beat a cheaper one that wanders. If you're still choosing between agent frameworks, our OpenClaw vs Hermes comparison covers how each one handles memory and tools.

How to cut the bill without hurting quality

1. Route by difficulty. Most agent calls are routine: short answers, lookups, checks, summaries. Send those to a cheap model and keep a strong one for multi-step tool work and judgement calls. With 80% of calls on DeepSeek V4 Flash and 20% on Claude Sonnet 5, our team-bot estimate drops from $272 to $78 a month at list price. Keep each conversation on one model where you can, since a switch means a cold cache.
2. Keep the persona and memory lean. OpenClaw's 60,000-character bootstrap limit is a ceiling, not a target. Every 1,000 tokens cut from the system prompt removes 9M tokens a month from the team bot: about $8 on Sonnet 5 and $15 on Opus 5.5 at list price.
3. Cap output and reasoning. Set a max output limit and skip the highest reasoning effort on routine turns. Google's pricing page says its output price includes thinking tokens, so a long reasoning trace bills like a long reply.
4. Help the cache. Keep the start of the prompt stable and put changing content at the end. If your agent goes quiet for longer than the cache lives, weigh a longer cache against its higher write price.
5. Space out scheduled jobs. Lengthen heartbeat intervals, use active hours, and give each job a short prompt so "nothing new" is a cheap answer.
6. Watch tool outputs. A whole web page or log file dumped into context is billed again on every later call in that session. Trim results before they reach the model, and use /compact in OpenClaw to summarise long sessions.

Hosting vs model cost: which one matters?

Hosting is a small, fixed cost. On HostAiHub, a managed Hermes or OpenClaw agent is $10 a month ($100 a year, or $250 lifetime), and n8n starts at $7 a month. A $5 to $10 VPS you run yourself costs about the same, plus your time for updates and backups.

The model bill is the variable, set mostly by which model you use, how much context each call carries and how often the agent wakes up. With us, each agent runs in its own private container on a private SSL subdomain, deploys automatically in a few minutes, gets daily backups kept for seven days and updates in one click with automatic rollback. Bring any provider's key or use our AI API. Plans are on the pricing page.

Every number here is an estimate built on stated assumptions and prices as of 2 November 2026. Prices change often, so check the provider pages above, and measure your own agent before trusting anyone's averages, ours included.