AI agent pricing trends 2026: what 38 dated price changes show

AI agent pricing in 2026 has settled on usage meters, not seats. Of the 41 agent products in my pricing census, 23 (56%) bill per credit, per run or per minute, the median entry plan is $20 a month, and pricing per resolved ticket has spread only in customer support. Underneath those plans, the tokens agents run on keep collapsing in price: GPT-4-level capability costs 99.8% less than it did in October 2023. Product prices have not followed. Vendors change the unit they bill in instead of the number on the page.

I keep three datasets on this, all free to copy. The AI agent pricing change log and census holds 38 dated pricing changes from 23 vendors since August 2024. The LLM API price history rebuilds what a million tokens has cost, every day, back to October 2023. The AI agent pricing index re-reads 16 platforms' pricing pages every month. Every figure below comes from one of those pages or from the source linked beside it, with the date it was read, because these numbers move.

AI agent pricing in 2026: the numbers

Here is the market in one table. The census classifies each product by the unit its bill scales with on the entry paid plan; the change log counts every pricing change I could date and source.

Two of those rows pull in opposite directions, and the rest of this piece is about the gap between them: the raw material gets cheaper by orders of magnitude, while the finished product sits at $20 a month.

Trend 1: inference keeps getting cheaper, fast

The cost of running a model at a fixed level of capability has fallen faster than almost any price in software. Stanford HAI's 2025 AI Index found that the inference cost for a system performing at the level of GPT-3.5 (64.8 on MMLU) dropped from $20 per million tokens in November 2022 to $0.07 by October 2024, using Gemini-1.5-Flash-8B: a more than 280-fold reduction in about a year and a half. For models scoring above 50% on the much harder GPQA benchmark, the same report puts the fall at $15 per million tokens in May 2024 to $0.12 by December 2024 (Phi 4).

Other trackers agree on the direction and disagree on the speed. Epoch AI found that prices for a fixed performance level fall anywhere from 9x to 900x per year, depending on the milestone (March 2025). Guido Appenzeller at a16z put it at 10x every year for a model of equivalent performance, a trend he named LLMflation (November 2024).

What the 2026 AI Index says about inference cost

If you came looking for the 2026 version of the 280-fold figure: the 2026 AI Index, published in April 2026, does not carry an updated per-token price series in its text. Its inference section measures energy per query instead. Citing Jegham et al. (2025), it reports DeepSeek V3.2 at about 23 Wh per medium-length prompt (roughly 1,000 tokens in and 1,000 out), GPT-5 (high) at 21.9 Wh, and Claude 4 Opus and GPT-5 mini at 5 to 6 Wh. Its technical chapter notes that six labs now share the top tier of Arena ratings, "shifting competitive pressure toward cost, reliability, and domain-specific performance."

For a running price series, the LLM API price history holds three capability levels fixed and records the cheapest way to buy each one, month by month since October 2023. On 9 October 2026:

• GPT-4 class: $0.0575 per million tokens blended (gpt-oss-20b on DeepInfra, $0.03 input and $0.14 output), against $37.50 for the original GPT-4 on 1 October 2023. That is 99.8% cheaper, about 652 times, in three years.
• GPT-4o class: $0.0703, down 36% in 12 months.
• Gemini 2.5 Pro class: $0.152, down 94% in 12 months from $2.41.

Blended means three input tokens to one output token. Two things stand out. The biggest single step, in August 2024, came from a small model clearing a bar that only frontier models had cleared before. And since September 2025 the floor has belonged to inference hosts serving open-weight models, not to the labs; it has even ticked up, from $0.025 to $0.0575, as host prices and tier membership shifted. The frontier itself gets cheaper slowly: the median new model from 11 major labs still lists at $0.75 input and $2.55 output per million tokens.

Trend 2: cheaper tokens did not make agent plans cheaper

If tokens are hundreds of times cheaper, why does the typical agent plan still cost $20 a month? Because an agent spends far more tokens per task than a chat reply, and the newest models spend more still. Cursor said it plainly when it rewrote its Pro plan in June 2025: "New models can spend more tokens per request on longer-horizon tasks," and "the hardest requests cost an order of magnitude more than simple ones." At median usage, its $20 of included usage covered about 225 Sonnet 4 requests, 550 Gemini requests or 650 GPT 4.1 requests.

A worked example: one agent task, two price points

Take an illustrative agent task with 20 steps, where each step re-reads 15,000 tokens of context and writes 800 tokens. That is 300,000 input tokens and 16,000 output tokens per run. At list prices on 9 October 2026:

The same work costs about 24 times more on a current mid-priced model than on the cheapest model that clears the original GPT-4 bar. Prompt caching cuts the input side, since cached reads cost a fraction of fresh input on most APIs. To price your own prompts, run them through the token counter, which compares per-call and per-month cost across GPT, Claude and Gemini; the longer method is in what it costs to run an AI agent around the clock.

That is why flat $20 plans persist. The token price is not the product price; the number of steps, the model choice and the retries are. Cheaper inference shows up as more work per dollar inside the same plan, not as a smaller invoice. I go through the cost side in more depth in AI agent cost optimization.

Trend 3: usage meters replaced seats

Seats fit software a person operates. Agents break the link between people and work, so most agent products now bill on what the agent does. This is how the 42 products in the census charge on their entry paid plan, read between April and October 2026 (most on 15 September or 4 October).

Seats survive where a person drives every action, as with coding seats and Microsoft 365. Credits dominate multi-step workflow agents, because a credit lets a vendor charge different amounts per action without touching the headline price. Per-minute pricing owns voice agents. Hybrids, a plan price with included usage and metered overage, are where the coding agents landed. For the monthly bill each family produces on a real workload, see AI agent pricing explained; for the lowest floors, the cheapest AI agent platforms.

Trend 4: price rises arrive as a new unit, not a new number

Of the 38 changes in the log, only 5 are plain list-price increases and 7 are decreases. The other 26 change what a plan includes or introduce a new unit, so the headline price often stays put while the bill moves. A buyer comparing two pricing pages a year apart could see no change at all.

• Make switched its billing unit from operations to credits on 27 August 2025, converting balances 1:1, so that usage could be priced by processing complexity and AI use (Make Help Center).
• Relevance AI split its single credit into Actions, for what the agent does, and Vendor Credits, for the model cost (Relevance AI).
• GitHub replaced premium request units with GitHub AI Credits on 1 June 2026, consumed by tokens at each model's API rate, while Pro stayed at $10, Pro+ at $39, Business at $19 per user and Enterprise at $39 per user (GitHub, 27 April 2026).
• Cursor replaced a limit of 500 requests a month with $20 of model usage at API pricing on 16 June 2025, then apologised and refunded unexpected charges incurred between 16 June and 4 July (Cursor).

One vendor went the other way. On 19 March 2026 Windsurf dropped credits for daily and weekly quotas, priced Pro at $20 and added a $200 Max plan (Cognition). A quota is easier to understand and harder to budget past, so expect the two approaches to keep trading places. Across the 15 platforms whose pricing pages the index snapshots, the set of printed prices moved 34 times in the last 12 months, and only confirmed changes make it into the log.

Trend 5: pricing per outcome stays in customer support

Charging only when the agent succeeds is the most discussed model and one of the least common. Every outcome-priced product in the census answers support tickets, because a resolved ticket is easy to count and both sides agree on what it is worth.

What the same support month costs on each

Take 10,000 support conversations a month, of which the agent fully resolves 7,500 (75%). At the published rates as of October 2026:

• HubSpot Customer Agent: 7,500 × $0.50 = $3,750
• Intercom Fin: 7,500 × $0.99 = $7,425, more if configured handoffs count as outcomes
• Salesforce Help Agent: 7,500 × $2 = $15,000
• Salesforce Agentforce per conversation: 10,000 × $2 = $20,000

Helpdesk seats come on top in every case, and each vendor defines a resolution slightly differently, so the lowest rate is not automatically the lowest bill. Outside support, most agent work has no unit this clean. What is the billable outcome of "watched the pipeline all week" or "drafted the brief"? That is why credits, not outcomes, are the default everywhere else.

Trend 6: the price war is in coding agents

If there is an AI agent pricing war in 2025 and 2026, it is among coding agents, and it has produced a clear shape: a $20 entry plan, a $100 to $200 tier for heavy users, and usage at API rates beyond that.

The entry prices converged, so the competition moved to how far $20 goes and what happens after it runs out. Cursor's June 2025 refunds and the bills Replit users reported after Agent 3 launched (one wrote that they "blew through $70 in a night") both came from that line, not from the price on the plan card. I log what went wrong in cases like these in AI agent failures: lessons from 63 incidents.

How to read an AI agent price in 2026

Every change in the log teaches the same lesson: the headline price is the least informative number on a pricing page. These six checks take ten minutes and catch most surprise bills.

1. Name the unit. Seat, credit, run, task, minute or outcome. The census above shows how each family behaves.
2. Find what one unit buys for your task. A credit is a vendor currency whose burn per action varies. Lindy, for example, puts an everyday ask at 2 to 250 credits and a research report at up to 1,000, per the pricing index.
3. Price a normal month and a heavy month. Zapier counts every step as a task while n8n counts a whole run as one execution, so the same workflow lands on very different bills. The automation platform picker ranks 16 platforms for a described task, with their prices.
4. Read what happens at the limit. A hard stop, a grace period, automatic overage or a throttle. This one line decides whether a busy month is an annoyance or a surprise invoice.
5. Check the change history. A vendor that has restructured twice in two years will probably do it again. Dated histories exist for Zapier, n8n and the other platforms in the index.
6. For agents you build on an API, price the tokens yourself. Use the token counter, and check the free LLM API tiers before you commit to a model.

For the wider set of figures behind these trends, AI agent pricing statistics for 2026 collects them in one place.

Where Gravity sits, for disclosure

Gravity is my company, so read this as a disclosure, not a recommendation. Gravity is in public beta and sells a flat subscription with usage included: Free is $0 (no card), Mini Pilot is $15 a month (one automated agent, usage included), and Auto Pilot is $20 a month. If a month's usage runs out, you can buy more from your account. In the census that puts Gravity among the 3 flat-subscription products, and its $5 Pro plan is the lowest published entry price there, which says more about alpha pricing than about where the market is heading.

The trade-off of any flat plan applies to ours as well: it is easy to budget, and the included usage is where the real limit sits, so check it against the work you have in mind. The plans are on the pricing page.

Frequently asked questions

How are AI agents priced in 2026?

Mostly on a usage meter. As of 4 October 2026, 23 of the 41 agent products in our pricing census bill on usage (13 per credit, 6 per run or task, 4 per minute), 7 combine a plan price with metered overage, 4 charge per outcome, 4 per seat and 3 sell a flat subscription. The median entry paid plan is $20 a month.

Are AI agent prices going up or down?

Both, and the direction is often hidden in the unit. Of 38 dated changes logged from August 2024 to October 2026, 5 are plain increases and 7 are decreases; the other 26 restructure what a plan includes or introduce a new unit. Devin's entry price fell from $500 to $20 a month in April 2025, while GitHub moved every Copilot plan to usage-based billing in June 2026 at unchanged plan prices.

What did the Stanford AI Index find about inference cost?

The 2025 AI Index reported that the inference cost of GPT-3.5-level performance fell more than 280-fold, from $20 per million tokens in November 2022 to $0.07 in October 2024. The 2026 edition, published in April 2026, measures inference by energy per query rather than an updated price series. Gravity's LLM API price history continues the series: GPT-4-class tokens cost $0.0575 per million on 9 October 2026, against $37.50 in October 2023.

Why are AI agent subscriptions not getting cheaper if tokens are?

An agent spends many more tokens per task than a chat reply: each step re-reads its context, calls tools and sometimes retries, and newer models spend more tokens on hard tasks. Falling token prices therefore show up as more work inside the same plan, usually around $20 a month, rather than as a lower price.

Which AI agents charge per resolution or per outcome?

As of October 2026: Intercom Fin at $0.99 per outcome, HubSpot's Customer Agent at $0.50 per resolved conversation since 14 April 2026, Salesforce's Help Agent at $2 per resolution (announced June 2026), Zendesk per automated resolution since August 2024, and Sierra at undisclosed rates. All of them are customer-service agents.

How should I compare AI agent pricing?

Start from one real workload. Name the billing unit, work out how many units one run uses, and price a normal and a heavy month on each platform. Then check what happens at the plan limit and how often the vendor has changed its pricing. Gravity's pricing index and change log list current prices and dated changes, free to reuse.

Sources

• Gravity, AI agent pricing change log and pricing-model census, updated 5 October 2026 (census read 4 October 2026).
• Gravity, LLM API pricing history, updated 9 October 2026.
• Gravity, AI agent pricing index, verified 15 September 2026.
• Stanford HAI, 2025 AI Index Report, inference cost analysis, 2025.
• Stanford HAI, 2026 AI Index Report, April 2026.
• Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks, 12 March 2025.
• Guido Appenzeller, a16z, Welcome to LLMflation: LLM inference cost is going down fast, 12 November 2024.
• Cursor, Clarifying our pricing, July 2025, and the Ultra plan announcement, 16 June 2025.
• Cognition, Introducing our new Windsurf pricing plans, March 2026.
• TechCrunch, Devin, the viral coding AI agent, gets a new pay-as-you-go plan, 3 April 2025.
• Anthropic, Claude pricing, read October 2026.
• GitHub, GitHub Copilot is moving to usage-based billing, 27 April 2026.
• Make, Credits as the new billing unit in Make, August 2025.
• Relevance AI, Splitting the bill, 2025.
• Intercom, From resolutions to outcomes, 12 March 2026, and Fin pricing, read October 2026.
• HubSpot, Customer Agent and Prospecting Agent: now you pay when the task is complete, April 2026.
• Constellation Research, Salesforce takes a run at outcome-based Help Agent pricing, 26 June 2026, and Salesforce, Agentforce pricing, read October 2026.
• CX Today, Zendesk unveils industry-first outcome-based pricing model, 29 August 2024.
• The Register, Replit Agent 3 pricing complaints, 18 September 2025.