Copilot Studio credits: 17 per grounded answer, and the open-source meter

What a Copilot Credit replaced

Copilot Studio used to count messages. Since 1 September 2025 it counts Copilot Credits. Microsoft says the swap changed neither the pack nor the rate: "There's no change in the quantity per prepaid pack or to the pay-as-you-go rate" (Microsoft's Copilot Studio billing page). What moved is the shape of the unit. A message was one reply. A credit measures work, which Microsoft defines as "a measure of the time and effort required for an agent to retrieve information, respond to prompts, and perform tasks through actions" (offer details).

Two agents can answer the same 3,000 questions a month, for the same people, and land an order of magnitude apart on the invoice. The meter counts how each answer was assembled. A forecast built on one reply per person estimates the wrong thing.

The rate card behind the bill

Every row in that card is Microsoft's, from the billing rates page, checked October 2026.

Classic answer                              1 Copilot CreditGenerative answer                           2 Copilot CreditsAgent action                                5 Copilot CreditsTenant graph grounding for messages        10 Copilot CreditsAgent flow actions, per 100 actions        13 Copilot CreditsText and generative AI tools, basic         0.1 credit per 1K tokensText and generative AI tools, standard      1.5 credits per 1K tokensText and generative AI tools, premium      10 credits per 1K tokensContent processing, per page                8 Copilot Credits

Reasoning models put a second meter beside that card. Microsoft writes the rule out: "Total cost = feature rate for the operation + text and generative AI tools (premium) for the reasoning model's token usage." An answer from a reasoning model pays the feature rate for being an answer. Then it pays again for the tokens it thought with.

Two ways to pay sit under the same card. The prepaid pack is "$200 per pack/month (*Billed annually)" and holds "25,000 Copilot Credits". The pay-as-you-go meter is "$0.01/Copilot Credit", billed in arrears at the end of the month (offer details).

3,000 answers a month, priced on both meters

The method is one line: multiply the credits one answer draws by the monthly answer count. Then compare the ceiling of packs at $200 against the same credits at a cent.

Workload A is a support agent. It generates an answer, grounds it in the tenant graph and takes one agent action, 3,000 times a month.

credits per answer = 2 generative + 10 grounding + 5 action = 17monthly credits    = 17 x 3,000 = 51,000pay-as-you-go      = 51,000 x $0.01 = $510prepaid packs      = 51,000 / 25,000 = 2.04, so 3 packs = $600two packs + 1,000 credits of pay-as-you-go = $400 + $10 = $410

Workload B keeps the agent and the 3,000 answers. It moves the same questions to a reasoning model that reads and writes 2,000 tokens an answer. At 10 credits per 1,000 tokens that adds 20 credits to every reply.

credits per answer = 17 + (2 x 10) = 37monthly credits    = 37 x 3,000 = 111,000pay-as-you-go      = 111,000 x $0.01 = $1,110prepaid packs      = 111,000 / 25,000 = 4.44, so 5 packs = $1,000four packs + 11,000 credits of pay-as-you-go = $800 + $110 = $910

Same questions, same traffic, $510 in one build and $1,110 in the other. The agent thinks before it answers. Strip the reasoning model, write the fifty most common questions as scripted answers, and 3,000 answers cost 3,000 credits, which is $30. The same month runs from $30 to $1,110, depending on how the agent was built.

Grounding, actions and reasoning are the levers

Microsoft's own worked example prices "a single complex prompt" at 12 credits, 10 for tenant graph grounding and 2 for the generative answer. Grounding is the largest single line on the card, and it is a per-agent switch: "This capability is optional, and you can turn it on or off for each agent." A bot that answers from one uploaded manual does not need the tenant graph. Switching it off removes 10 of the 17 credits in workload A.

Agent actions are the second lever. Microsoft counts the "steps such as triggers, deep reasoning, and topic transitions" that make up an agent's route through a task. Every one is 5 credits. An agent that chains three steps per reply pays more for navigation than for prose.

Reasoning tokens are the third lever, and the one that scales worst. Ten credits per 1,000 tokens costs little at 2,000 and a great deal at 20,000. That is what a long retrieved context plus multi-step inference produces. The premium line is the one token rate that sits on top of a feature rate, so it grows with the length of what the agent reads and writes rather than with the number of requests.

None of the three is headcount. People set the multiplier, the number of answers a month. The design sets whether each answer costs a cent or thirty-seven.

The pack does not roll over

Microsoft states plainly that "Copilot Credits do not roll over to the next month." Money spent on a pack for a quiet month is gone. The rate turns against the buyer as soon as the pack half fills. A pack that fills costs 0.8 cents a credit. A half-filled pack costs 1.6 cents, more than the cent on the meter. Workload A shows the same effect at three packs: $600 for the 51,000 credits actually used works out at 1.18 cents each. Two packs with 1,000 credits of overflow comes to $410, which is $190 cheaper than sizing to the ceiling.

Under-buying carries a sharper cost. Microsoft's enforcement triggers "when a tenant reaches 125% of their prepaid capacity". Once it does, "Custom agents are disabled". A team that sizes packs to its quietest month meets a switched-off agent in its busiest one.

What an open-source alternative meters instead

Kortix, the open-source AI Operating System, prices the same nominal unit differently. Its pricing page reads "1 credit = $0.01" and "11 credits = 1 hour" of agent computer, "Billed per second. Nothing while it is stopped." The page puts two goods on that one meter: "Credits pay for the agent computer and for Kortix models." A Kortix-managed model bills "at the provider's base price plus a 5.5% platform fee", paid from credits, and the app shows base prices.

A key from Anthropic, OpenAI or another provider moves the model half. Model usage "goes straight to your provider": the first $1,000 a month is free, then 5% of it is paid from credits. A connected ChatGPT subscription "uses no credits". The paid plan starts at $50 a month for 5,000 credits, at 100 credits per dollar, and running it on your own network is free.

The two credits buy different things, so the figures are not interchangeable. A Copilot Studio credit carries the model that wrote the answer. A Kortix credit buys machine seconds, and on your own key the model bill goes to the provider, outside the meter. That is why 17 credits of grounded Copilot Studio answer costs 17 cents, while 17 credits of Kortix computer time is about an hour and a half. An agent can run a batch job, a test suite and a script in that time.

Past the price of the model, the two meters differ in shape. One charges per answer assembled, the other per second of work. A stopped machine costs nothing.

When Copilot Studio is the right buy

For an agent inside a Microsoft tenant, answering people who already hold a Microsoft 365 Copilot seat, the credit bill is usually nothing at all. Employee-facing use is included, and agents built for Teams, SharePoint and Microsoft 365 Copilot arrive with the seat at no additional charge (offer details). The Microsoft 365 Copilot add-on is the purchase that unlocks it. Governance is already in place, in the tenant boundary, Purview and Entra.

Paying credits for that stack is a fair price, and moving off it has a cost of its own. That arithmetic belongs to the agents outside it: customer-facing ones, ones used by people without a seat, and any agent that chains reasoning tokens or several actions per answer. For those, the bill is a build decision. Pricing the same 3,000 answers on Kortix is the comparison to run before the next pack is bought. If the question is whether any part of Copilot is free at all, the free tiers are worth a look first.