AI Agent Cost Per Month (2026): 5 Models, Real Numbers

Prices checked 21 October 2026.

A support-triage agent handling 6,000 tickets a month costs $9.52 to $150.62 a month in API fees at list price. That's 0.16 to 2.5 cents per ticket. The only thing that changes across that range is the model. Add hosts and service tiers to the mix and it stretches from $4.40 to $903.74. Same prompts, same tokens, same agent.

That's the running cost, the bill that turns up every month. What it costs to build an AI agent in the first place is a different question, and it doesn't fit in a table.

Ask about AI agent pricing and you'll usually get a price per million tokens, which tells you almost nothing. Your bill is set by three levers, and they're worth very different amounts:

Most of the "12× cheaper!" screenshots I've seen are one of those three in a fake moustache.

The agent I priced

One workload, deliberately boring. A support-triage agent: 6,000 runs a month, four turns per run. Each turn sends 2,600 input tokens, 1,800 of them cacheable (system prompt plus tool schemas), and gets 320 tokens back. I assumed a 70% cache hit rate. If you've built a ticket triage agent before, that shape will look familiar.

I ran two other shapes as a sanity check:

Both give the same ranking (DeepSeek, Qwen, Gemini, then Claude and GPT tied) and roughly a 15× gap between the cheapest and dearest model. Everything below uses support triage, because it's the most common shape.

How I billed it, so you can argue with me properly. Cache hits are charged at each route's cached-input price, or at the full input price if the route doesn't publish one, so caching never flatters a route. Cache misses on the cacheable part are charged at the cache-write price wherever the vendor adds a write premium. Claude and GPT-6.1 Sol both do, at 1.25× input. Qwen's own API only charges that premium for explicit caching, so I've priced its implicit mode, where creating the cache costs the normal input price. And where a vendor documents a minimum prompt size for caching, anything below it gets no cache discount at all. That last rule ends up mattering a lot for Gemini.

Prices come from pricing.fru.dev, which reads OpenRouter's public model catalog every morning at 05:00 UTC and records each host's price separately, regions included (the method page has the details). Every route in this post was checked at 05:00:15 UTC on 21 October 2026. I also checked each list price against the vendor's own pricing page. That turned out to matter.

Lever one: the model you pick

At each model's list price (the lab's own API, no tier, no reseller), the same agent costs:

A 15.8× spread for picking a different model. Nothing else in this post comes close.

LLM API pricing pages lead with input and output. Three things in that table don't show up in the headline number, and each one depends on a time or a date.

DeepSeek's list price is its off-peak price. DeepSeek charges double from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Run this agent entirely in peak hours and it costs $19.05, not $9.52. US business hours sit entirely off-peak. A European support desk spends most of its morning in peak, and here in Sydney peak is lunch and dinner. The funny part: nine third-party hosts list this model at exactly $19.05 for this workload, because their only price is DeepSeek's peak rate ($0.30 in, $1.20 out). The model itself is the 17 October snapshot, which DeepSeek's API serves as deepseek-flash. (If you're wondering whether the V4 family is any good, I tested DeepSeek V4 on real code back in May.)

Gemini 3.8 Flash is on a promotional price, and it can't cache this agent at all. Google lists $0.75 in and $3.75 out until 31 December 2026, then $1.50 and $7.50 from 1 January. Its implicit cache only applies from 4,096 input tokens, and this agent sends 2,600 per turn, so every token is billed at full price. Put both together and in January this agent costs $151.20 a month on Gemini 3.8 Flash, 58 cents more than Claude or GPT with caching on. Put it in your calendar.

Claude Sonnet 5.5 and GPT-6.1 Sol now cost exactly the same. Input, output, cached input and cache write are identical on every line this agent touches. That wasn't true until 7 October, when Anthropic halved Sonnet 5.5's cache-read price from $0.20 to $0.10. The feed, and the OpenRouter listing behind it, still shows $0.20 on all eight Claude routes. The stale number makes Claude look $3.02 a month dearer than GPT, and I'd already written a tidy paragraph explaining why. It's gone now. I've used Anthropic's documented rate (0.05× base input) on every Claude route.

The one real difference left is cache lifetime, which matters in lever three.

Lever two: where you buy it

Now hold the model fixed and price every route it's sold through. The left half is every route. The right half drops the flex, priority, fast and ultrafast service tiers, so you're comparing like with like. Each model name links to its full route list.

The gap between the two halves is the point. GPT-6.1 Sol's 12× is OpenAI's flex tier (half price, slower, occasionally unavailable) against its ultrafast tier (six times list). That's buying or selling latency, not finding a better deal. Gemini's 3.6× is flex against priority. Strip the tiers and Claude, GPT and Gemini sit within 10% of their list price everywhere they're sold.

And that 10% isn't a reseller markup. It's geography. Every route at 1.1× in that table is a regional endpoint: Azure US and EU, Vertex US and Europe, Bedrock us-east-1. Anthropic and OpenAI both document a 10% uplift for pinning inference to a region. The non-regional routes on Bedrock, Azure and Vertex charge exactly what the lab does. So on frontier models, shopping around saves nothing, and needing data residency costs you 10%.

Open-weights models are different. DeepSeek spreads 4.5× and Qwen 6.0× across hosts that genuinely compete on price. If you run an open-weights model, finding the right host is worth an afternoon. (Or skip hosts entirely: Qwen3.8 27B is the one I ran locally on a 24GB M4.)

Before you go chasing the bottom of those ranges, three things to check.

1. The cheapest route is usually a tie. Five of Claude's eight routes cost exactly $150.62. Gemini's two flex routes tie at the bottom, and once tiers are stripped, so do its list and Vertex global routes. "The cheapest route" is normally a handful of routes, not one clever vendor.
2. Cheap often means quantised, but not always. DeepSeek's cheapest non-tier route ($4.82) runs at fp4, and Qwen's cheapest ($12.20) at fp8. Precision doesn't sort the list, though. Qwen on DeepInfra at bf16, one of only two 16-bit routes, costs $20.36. That's cheaper than nine of the ten fp8 routes and about 28% of the dearest. A route might be cheap because it's quantised, or it might just be cheap, and a price list can't tell you which. Test output quality on your own prompts.
3. A missing cache price is a penalty in my numbers, not necessarily on your invoice. Routes that don't publish a cached-input price get no caching benefit here. Even so, Qwen on Mancer ($31.68) still beats Qwen's own API ($35.82).

One more on Qwen: its dearest route is Cerebras at $73.22, which sells speed and lists cached input at the same price as fresh input. Leave it out and Qwen's spread drops from 6.0× to 5.1×. Still worth the afternoon.

Lever three: prompt caching

Same route (each model's list price), same workload. Only the cache hit rate changes:

Going from no caching to a 70% hit rate saves $4.45 a month on DeepSeek, $10.28 on Qwen and $50.98 on Claude or GPT. Now put that next to what host-shopping is worth on the same models (service tiers excluded): $16.93 on DeepSeek, $61.02 on Qwen, $15.06 on Claude or GPT.

On DeepSeek, picking the right host is worth almost four times as much as caching, and on Qwen nearly six times. On Claude and GPT it flips: caching is worth more than three times the entire route spread.

Gemini 3.8 Flash isn't in the table because its line would be flat at $75.60. A 2,600-token prompt never reaches the 4,096-token minimum for implicit caching. Explicit caching runs through Google's generateContent API and adds an hourly storage fee, and the docs I checked don't give a minimum size for it, so I haven't counted on it.

So which lever wins depends on the model you picked. Cheap open-weights models reward host-shopping. Expensive frontier models reward prompt discipline.

Those numbers assume your agent can actually hold a warm cache. That means a stable system prompt, tool schemas in the same order on every call, and nothing that changes at the top of the prompt. A timestamp in the first line of your system prompt is enough to miss every single time. (I wrote a whole field guide on structuring agent context if you want the long version.)

Before you count on caching, check two numbers for your model: the minimum prompt it will cache, and how long the cache lives.

Here's where Claude and GPT stop being twins. Anthropic's $2.50 cache write lives for 5 minutes, OpenAI's for 30. Claude refreshes the timer on every hit, so a busy queue stays warm. It's the gaps that hurt, and support tickets arrive in bursts with gaps in between. If your traffic is spiky, the identical price list won't produce identical bills. Anthropic sells a 1-hour cache at $4.00 per million tokens, OpenAI a 12-hour one at $5.00. And because OpenAI caches by default, with the breakpoint at the end of the latest eligible message according to its caching guide, check your usage report to see how many tokens you're paying to write.

Retries, the multiplier nobody budgets for

I also priced a 20% failed-and-retried run rate. That's an assumption, not a measurement. On DeepSeek it takes the bill from $9.52 to $11.43; on Claude or GPT, from $150.62 to $180.75. Treat it as a ceiling, because a retry that re-sends the same context can hit the cache. Structured output removes a whole class of retries too: with schema-guaranteed responses there's no malformed JSON to send back. And if your agent really fails one run in five, the API bill isn't your biggest problem.

How to estimate your own AI agent's cost

Here's the arithmetic behind every table in this post. It fits on a napkin:

turn cost    = ( fresh input  × input price
               + cache misses × cache-write price
               + cache hits   × cached-input price
               + output       × output price ) ÷ 1,000,000

monthly cost = turn cost × turns per run × runs per month

Fresh input is the part of the prompt that can't be cached. Your hit rate splits the cacheable part into hits and misses. If the vendor has no write premium, use the plain input price for misses. If the cacheable part is below the vendor's minimum, count all of it as fresh input.

Worked through for Claude Sonnet 5.5 on the support agent:

Multiply by 4 turns and 6,000 runs and you get $150.62 a month. Swap in your own token counts (the APIs report them in the usage block of every response) and you'll have a better estimate than any calculator, mine included.

What I'd do with this

Price your own shape before you look at anyone's comparison, including this one. Turns, tokens per turn, runs per month, through the formula above. The ranking held across all three shapes I tried. The absolute numbers didn't.

Only shop for hosts on open-weights models. On Claude, GPT and Gemini the non-tier routes are within 10% of each other, and that 10% is a region surcharge. On DeepSeek and Qwen there's 4.5× to 6× sitting on the table.

Fix the prompt before you switch vendor. Cache hit rate is engineering you control, and on a frontier model it's worth more each month than any route you can switch to.

Check the vendor's own docs for the lines that move: cached input, caching minimums, peak hours and promo end dates. Each of those changed a number in this post, and the aggregator was two weeks behind on one of them.

Ignore any comparison that mixes service tiers. Flex and ultrafast are different products that happen to share a model name.

And question whether every step needs an LLM. Routing a ticket to one of five queues is a typed decision, and a decision model answers those in milliseconds for a fraction of the cost.

If you'd rather have someone run these numbers on your actual workload before you build anything, or tell you whether you need an agent at all, that's the scoping part of my AI consulting in Sydney. For the "do I need one" question, I've also written up where AI agents actually earn their keep in small businesses.

FAQ

How much does an AI agent cost per month?

In API fees, a support agent running 6,000 four-turn conversations a month costs between $9.52 (DeepSeek V4.1 Flash, off-peak) and $150.62 (Claude Sonnet 5.5 or GPT-6.1 Sol) at list price. That's roughly 0.16 to 2.5 cents per conversation. The cost scales with runs, turns and tokens per turn, so price your own workload rather than borrowing mine.

What's the cheapest LLM API for running an agent?

Of the five models here, DeepSeek V4.1 Flash: $9.52 a month on DeepSeek's own API off-peak, or $4.82 on the cheapest third-party host outside a flex tier (an fp4 deployment). Cheapest isn't the same as good enough, so test it on your real tasks first.

Is prompt caching worth it?

If your system prompt and tool schemas are stable, yes. Going from 0% to 70% cache hits cut this agent's bill by 25% on Claude Sonnet 5.5 and GPT-6.1 Sol ($201.60 to $150.62) and by 32% on DeepSeek ($13.97 to $9.52).

Does prompt caching work on short prompts?

Only above each vendor's minimum. Claude Sonnet 5.5 caches from 512 tokens, GPT-6.1 Sol and Qwen from 1,024, and Gemini 3.8 Flash's implicit cache from 4,096. Below that, you pay full input price for every token.

Is Claude cheaper on Bedrock, Azure or Vertex AI?

No. The global endpoints match Anthropic's own price, and regional or US-only endpoints charge 10% more. Bedrock and Vertex set their own prices, though, so check that yours has passed on Anthropic's October cache-read cut.

What this doesn't cover

Quality, latency, rate limits and terms of service, any of which can make a cheap route the wrong route. I haven't measured any of them here, and a model that fails your task is infinitely expensive. For quality on real work, see my GPT-6.1 Sol review (a close second to Opus 5.5, but slow) and my Claude Sonnet 5.5 review.

"Routes" means hosts listed on OpenRouter, so a host that isn't there isn't here either. Batch discounts and long-context pricing are out of scope, because this agent's prompts are small and it needs answers now. Running the model on your own hardware swaps a monthly token bill for an upfront one, which I looked at for the NVIDIA DGX Spark. The 20% retry rate is an assumption, labelled as one.

Prices: every route checked by the feed at 05:00:15 UTC on 21 October 2026, with list prices, caching minimums and cache lifetimes cross-checked against the vendors' docs the same day. Claude Sonnet 5.5 cache reads corrected from the feed's $0.20 to Anthropic's published $0.10. The arithmetic, assumptions and the script behind every table are mine and reproducible.

Sources

• Per-route prices and check times: pricing.fru.dev (the script reads api/models/<id>) and its method page
• Claude Sonnet 5.5 pricing, cache writes and regional uplift: Anthropic pricing; caching minimum and lifetime: Anthropic prompt caching
• Date of the Sonnet 5.5 cache-read cut: Compute Labs
• GPT-6.1 Sol tiers, cache writes and regional uplift: OpenAI pricing, prompt caching guide and flex processing
• DeepSeek peak and off-peak rates: DeepSeek models and pricing; caching: DeepSeek context caching
• Gemini 3.8 Flash promotional pricing and caching minimum: Gemini API pricing and Gemini context caching
• Qwen caching modes and minimum: Alibaba Cloud Model Studio context cache