Track · Pricing & Monetization8 min read

What an AI agent costs per feature: a product manager's budget for one sprint

Every rate, quota and plan rule below was read on the vendor's own page on 19 November 2026, and the plan terms on 20 November 2026. Three links are referral links: a qualifying CometAPI top-up, eligible NanoGPT usage and a first Z.ai payment may earn us a commission or credits, at no extra cost to you. Every other link goes straight to the vendor and earns us nothing.

What an AI agent costs per feature: a product manager's budget for one sprint
What an AI agent costs per feature: a product manager's budget for one sprintstevepb / pixabay

"Can we afford an agent on this sprint?" Nobody can quote that from a vendor page: an agent is metered in tokens, not in features. The vendor pages publish the price of a token and the size of the window it must fit into, and that arithmetic can be done once and reused every quarter.

The unit of planning is a run, not a feature. An agent knows the context it was handed, not where one ticket ends and the next begins: a feature that goes through three rounds of revision is three runs of context.

What a run is, and where the budget goes

Three components, counted three ways.

Volume. How much context you repeat. Every vendor prices the same token twice, fresh and read back from cache. Claude Sonnet 5.5 lists $2 per million input tokens and $0.10 per million cached reads, and MiniMax-M3 lists $0.30 input and $0.06 cached. A repeated repository read costs a twentieth of a fresh one at Anthropic and a fifth at MiniMax.

Windows. A coding plan is not a monthly pool. Z.ai's GLM Coding Plan publishes two independent windows: 2,000 credits for Lite, 12,000 for Pro and 28,000 for Max every five hours, plus 10,000, 60,000 and 140,000 a week, resetting five hours after each consumption. OpenCode Go counts allowances over five-hour, weekly and monthly windows. Unused quota expires rather than rolls over, and Ollama resets included usage on the subscription start date.

Waste. Calls that spend tokens and return nothing usable. Openference charges client errors and successes against your allowance but not upstream provider failures, with a hard burst cap of 50 requests a minute. Z.ai sells no overage: when the quota runs out the system waits for the next five-hour cycle instead of billing the account, so a wasted call comes out of the next window. And changing model or effort invalidates the context cache on Kimi Code, which turns a cached run into a fresh one.

The published rates and the cost of one run

Assumed run for every row: 1,000,000 input tokens and 100,000 output tokens, nothing cached. Those are hypothetical inputs, not measurements. The formula is yours to re-run: input tokens times the input rate, output tokens times the output rate. Anything marked (our arithmetic) is ours; the rest is the vendor's.

Model class Input / output per million tokens Cached read per million One hypothetical run (our arithmetic)
Claude Haiku 5.5, prompts up to 100K $0.10 / $0.50 $0.01 $0.15
MiniMax-M3, up to 512K input tokens $0.30 / $1.20 $0.06 $0.42
GLM-5.3 on Ollama's cloud list $1.40 / $4.40 $0.26 $1.84
Claude Sonnet 5.5 $2 / $10 $0.10 $3.00
Claude Opus 5.5 $4 / $20 $0.20 $6.00

The gap between the cheapest and the dearest row is forty times on the same hypothetical run, and that is a routing decision, not a quality difference. If nine tenths of that million input tokens came back from cache, the Sonnet 5.5 row becomes $0.09 cached plus $0.20 fresh plus $1.00 of output, so $1.29 instead of $3.00 (our arithmetic).

What a sprint and a quarter cost

Take the run as the unit and multiply. A sprint of 20 features, one run each: $8.40 on MiniMax-M3, $36.80 on GLM-5.3, $60 on Sonnet 5.5, $120 on Opus 5.5 (our arithmetic). A quarter of four such sprints: $33.60, $147.20, $240 and $480 (our arithmetic).

Read that against the plan prices: OpenCode Go at $10 a month, NanoGPT Pro at $12, Z.ai Lite at $18, Openference Minion at $20 and Claude Max at $100, which brings $100 of API credit a month. A team whose quarterly token bill is $240 needs pay-as-you-go, not a $100 subscription.

So measure one month of your own list-price bill first. That number is the break-even against every credit-for-price plan, and the only figure you need before the planning meeting.

Where a person stays in the loop, and where an agent is not worth the spend

Two published rules decide where an agent may run unattended. Z.ai's terms bar using the coding quota for your own applications, bots or SaaS products, and MiniMax says its M Plan is for individual, interactive developer use and points production at pay-as-you-go. Openference's terms go further: allowances do not cover commercial production, background agent loops, batch pipelines or headless daemons, and it rejects non-compliant requests by inspecting request metadata. A scheduled job on a subscription plan is not a cost question. It is a terms question.

Keep the human where the cost is not in tokens. Anthropic charges $0.08 per session-hour for Managed Agents runtime and $10 per 1,000 web searches, so a research agent searching 200 times in one run spends $2 on search alone before the model sees a token (our arithmetic). An output a person reads line by line is already priced at that reader's hourly rate.

And skip the agent where it does not pay: a one-off task a person finishes in ten minutes, a task whose context cannot be reused, a feature that needs three approval rounds before anyone looks at it.

Frequently asked questions

How much does one agent run cost? A run costs tokens multiplied by the published rate. On the hypothetical run above, 1,000,000 input tokens and 100,000 output tokens, nothing cached, Sonnet 5.5 at $2 input and $10 output per million comes to $2.00 plus $1.00, so $3.00 (our arithmetic). Haiku 5.5 does the same work for $0.15 (our arithmetic).

What does an AI agent cost per feature? No vendor publishes tokens per feature, so only your own measurements answer that. The cost per feature is the cost per run times the number of runs it took: three rounds of revision on a $3.00 run is $9.00 (our arithmetic). Compare that with the engineering time you would spend instead.

Is a coding plan cheaper than pay-as-you-go for a team? Only above your own bill. Claude Max at $100 brings $100 of API credit, so it wins once your monthly list-price bill passes $100 and wastes money below that; Z.ai Lite starts at $18. Measure one month of your own list-price bill first: that is the break-even.

How long does a subscription quota last? Five hours and a week, not a month. Z.ai's GLM Coding Plan publishes 2,000 credits for Lite every five hours and 10,000 a week. OpenCode Go counts over five-hour, weekly and monthly windows; Openference's five-hour window is 20 % of the monthly budget. Unused quota expires.

Can we run an unattended agent on a coding plan? Usually not, and the terms say so rather than the invoice. Z.ai bars the coding quota for your own applications, bots or SaaS products. MiniMax says its M Plan is for individual, interactive developer use and points production at pay-as-you-go. Openference excludes commercial production, background agent loops and headless daemons.

Which model should a product team use for an agent? Route by task, not by prestige. The spread on the same hypothetical run is forty times, from $0.15 on Haiku 5.5 to $6.00 on Opus 5.5 (our arithmetic). Drafting, summarising and classification can run on a cheap class; reserve a frontier class for the step where a human reads the output.

Sources

Vendor pages read 19 November 2026, plan terms read 20 November 2026: Claude pricing · Claude credits · MiniMax pricing · Z.ai plan · Z.ai terms · Ollama pricing · OpenCode Go · Openference pricing · Openference terms · NanoGPT pricing · Kimi membership

Referral links

CometAPI pays the inviter 20 % of every tracked top-up in cash, with a 120-day attribution window and a $50 minimum payout; the reader pays the same price and gets the published rule that customer price is 0.8 of the official rate on official models. NanoGPT pays 10 % of eligible usage and 5 % of subscription purchases into the NanoGPT balance, and the reader gets 5 % off eligible usage. Z.ai pays credits worth 10 % of the invitee's first payment after three eligible invitees have paid, and the reader gets 10 % off the first GLM Coding subscription. Every other link points straight at the vendor and earns us nothing.

Affiliate disclosure: the CometAPI, NanoGPT and Z.ai links above are referral links. If a reader tops up CometAPI, uses NanoGPT or buys a GLM Coding subscription through them, we may earn a commission or credits. The reader's price is unchanged.

More from Marcus Bennett

Marcus Bennett

Show full bio

Market editor covering regulation, policy and their impact on product management at Roadmap File.

56 articles