On rainy days, we take our boys bowling in Brooklyn. It’s old school, the type of place where the beer is Bud Light and the kitchen is a deep fryer. A good time is had by all, but my five-year-old enjoys it the most for one reason, the arcade.
Last time, we paid for one game. He climbed onto an oversized motorcycle, barely big enough to reach the controls, and started racing. I expected two dollars to buy about thirty seconds. But he kept winning. Another race started. Then another. Eventually, our bowling lane was ready, and I had to pull him away while the game was still going.
Our AI subscription has felt like that. We haven’t figured out how to lower our monthly AI costs to $2, but we continue to experiment to get the most from every dollar we spend on AI.
We signed up for access to useful tools. We weren’t thinking about tokens, cache reads, or compute. We started using the tools and built it into how we do everything. Our cost has grown, but has remained reasonable.
Then we started hearing very different stories: people describing AI bills running into tens of thousands of dollars for single users. A core piece of our business has become integrating client data with AI, and advising on AI set-up. The contrast between our experience and others made us want a clear answer.
What were we actually buying? The first thing we did was ask the AI. There have been many times this year where AI has blown me away with a clear, thoughtful answer. This was not one of them. It gave vague feedback that did not make the answer clear or easy.
So we decided to gather the data ourselves. We wanted to explain why our subscription feels like such good value, understand the cost of recognizable work, and test the choices that change it. That is what we cover in this paper.
compare the pricing between an enterprise and team subscription
AIwhat even is money
Illustrative exchange; this is a joke, not a recorded model response.
For businesses buying Claude, the choice is Team or Enterprise. Team accounts are for groups with between two and 150 users. Enterprise is for companies with at least 20 users.
The guidance from Claude is that Enterprise is for bigger companies with more complex needs. Typically, you would expect the bigger companies to pay less. In this instance, they bill completely differently. Team bundles an allowance into the seat price, so you pay one amount for both access and a pre-set amount of usage. Enterprise charges for consumption on top of a seat fee. What that hides is a massive cost difference.
$25 Standard / seat / month
$125 Premium / seat / month
Prices shown are billed monthly. Annual billing is $20 Standard / $100 Premium. Premium is advertised at 5× Standard usage. Per-person session and weekly limits apply.
Team can also reduce overage costs through discounted usage bundles. The organization calculator below includes these purchases and their shared cap.
$20 / seat / month + usage
Annual billing. Usage across Claude, Code, and Cowork is charged at API rates, with no included token allowance. Organization and user spend limits can control the budget.
We tested the same defined task on Team and Enterprise with identical settings. One Enterprise run was recorded at about $4. The equivalent Team run used about 1.25 percentage points of the weekly allowance.
Five runs to compare the costs between Team and Enterprise.
About $4 for one run, according to the experiment’s billing note.
Results workbook · Test 3, H18Four repetitions of the same workflow moved the weekly usage bar from 0% to 5%, about 1.25 points each.
Results workbook · Test 3, B24:E27The task started with a synthetic general ledger of 2,621 rows across 12 months and five departments. It required removing duplicates, standardizing dates and vendors, building a monthly P&L and data-quality log, flagging operating-income changes above 20%, and writing and running a script to recreate the workbook.
The experiment summary records the runs as passed. The generated workbook, script, and invoice export were not available for this review.
When you extrapolate this over a month, a Standard seat gets you about $1,000 of equivalent Enterprise usage. Not bad for the $5 difference in seat cost when paying monthly for Team.
Four identical Team runs moved the weekly usage bar from 0% to 5%, about 1.25 points per run. A full week’s allowance therefore covers about 80 runs.
The same run on Enterprise billed about $3.75.
80 runs × $3.75 × 4.3 weeks ≈ $1,290 a month.
We rounded down to $1,000 for two reasons. The usage bar shows whole percentages and rounds down, so the four runs may have used closer to 6 points than 5. At 6 points, a week covers about 67 runs, or about $1,075 a month. The figure also assumes someone uses the full allowance every week.
Premium is advertised at 5× Standard usage, so we use $5,000.
But that is not all. On a Team plan, you can upgrade to a Premium seat for 5× the consumption of a Standard seat. The upgrade costs an additional $100 per month with monthly billing, or $80 per month with annual billing. What do you get in exchange for $80? Premium provides about $5,000 of equivalent consumption each month.
But wait, there’s more. On a Team plan, you can buy overage credits so that when people hit their plan limit they can continue to work. At this point you are paying API rates, but you can get a 30% discount by pre-buying $1,000 credit bundles for $700 each, up to $3,000 of credit value per organization per month. Enterprise has no similar bundle discount.
The chart and organization calculator hold the usage estimates fixed at $1,000 for Standard and $5,000 for Premium. Both choose the lowest-cost mix of seats and shared credit bundles. The chart divides the organization bill by the number of active users.
20 people · 50% usage variation · Team billed monthly. Adjust organization settings ↓
Annualized totals = monthly estimates × 12, assuming the same usage and settings each month.
A balanced starting curve, with lighter and heavier users around the average. Total usage stays the same.
The calculator chooses the lowest-cost mix of Standard seats, Premium seats, and optional credit bundles for the whole organization. Included seat allowances stay individual; purchased credits cover the combined overage.
$1,000 of credits costs $700; up to $3,000 of discounted credit value per Team organization each month. Smaller bundles: $50 for $45; $250 for $200. Credits cover overage after included usage. No bundle discount is applied to Enterprise. Published terms ↗
Team: $25 Standard / $125 Premium per seat per month, billed monthly. Enterprise: $20 per seat per month, billed annually, plus usage.
Estimated equivalent usage included in a Standard seat. Premium uses $5,000 (5×). From the original finance test ↗
Choose a normal starting curve or a mostly-light-user profile: most people use less than average, while a few heavy users pull the average up. This is a right-skewed distribution. It uses a lognormal shape (log-scale spread 0.7), standardized before applying the spread control. The spread control sets the starting standard deviation as a percentage of the selected average; the default is 50%. Both shapes are planning assumptions, not measured corporate usage.
For the selected number of people, use evenly spaced percentiles of the chosen shape. Floor negative usage at zero, then rescale the values so their average exactly matches the slider. These adjustments can reduce the realized spread and change the shape near zero. Very small teams cannot fully represent a distribution; one person has no between-person variation. Zero spread restores the equal-usage scenario.
The same settings always produce the same people and costs. These are representative profiles, not random forecasts or confidence intervals. Usage above $10,000 is possible for an individual even though the average slider stops there.
Seat choices and bundle purchases are optimized together. The purchase cap applies once to the whole organization. Unused seats required by the minimum are charged at Standard rates. Enterprise uses the same total usage plus its seat fees, so changing the spread or distribution alone does not change its modeled bill. Team savings can change because heavier users may need Premium seats or overage, while lighter users retain unused individual allowances. If everyone remains within their allowance, the bill can stay unchanged.
The monthly estimate starts with no existing credit balance. It checks the published bundle sizes against ordinary overage, charges the full purchase price this month, and reports any unspent credits separately. Future use of that balance is not deducted from this month’s cost. The model can choose ordinary overage when buying a bundle would cost more. Savings compare the best setup with bundles against the best setup without them.
It costs a lot more, but surely it must come with more features? And it does. The biggest additions address governance needs: role-based access and restrictions, audit logs, custom data retention, and a Compliance API. Some organizations need these features; most do not.
Enterprise capabilities · Team capabilities
In almost every case, a Team plan is far less expensive, especially once consumption increases. Unless you really value tracking for compliance or monitoring reasons, the incremental cost is hard to justify. There are also to plug some of the visibility gaps that exist on a Team plan so that you can still see into your organization in a similar way to what exists on Enterprise.
Interested in closing the gap between Team and Enterprise features? Send us a note. We are happy to discuss.
If you are in a regulated industry where the compliance needs require an Enterprise plan, then it is a great option. But the cost differential is significant enough that that question should be thoughtfully examined. And as the team learns how to use the tools and moves up the adoption curve, your usage is likely to increase. The decision you make today could have a multi-hundred thousand dollar impact over the next year.
Consuming AI today can feel like buying electricity. A token, like a kilowatt-hour, is not an intuitive unit. Over time, we learn that a light and an air conditioner put very different demands on the meter.
Today, for many users of enterprise AI, there is a visibility gap. They know the team is using the tool. They get and pay the bill. But it is hard to understand the cost of the things being done. What does it cost to review a financial package, mark up an agreement, or build something from a few spreadsheets? We ran a wide variety of daily tasks in Claude on Opus 5.5, our workhorse model at the moment, and checked the cost of each. The results are summarized in the workflow exhibit below.
Tokens, API time, and cost describe the consumption. The four workflows below show the work behind those readings.
Reported cost is the work priced at API rates, as shown in Claude’s usage panel. On Enterprise, that is roughly what you pay. On Team, it draws down the seat’s included allowance until that runs out.
We ran these four in Claude Code on Opus 5.5, with memory on.
Toggle a workflow to see what it adds to the meter.
Different tasks, different loads on the meter. Costs are API-rate equivalents. Open a workflow to see what the work involved.
Under $2
6 workflowsReview a nondisclosure agreement.
Read the morning’s call notes and draft a follow-up.
Review six client questions in a feedback log, query Snowflake to identify data gaps, and develop a fix plan.
Review a nondisclosure agreement and produce a markup.
Send a Microsoft Teams invitation.
Review Asana tasks, calendar commitments and recent email. Update 16 tasks, including revised due dates, and create one call-preparation task.
$2 to under $10
12 workflowsReview and summarize the most recent financial package.
Gather context from project and memory files, code, commit history, Asana, and a read-only NetSuite query, then produce a call-preparation PDF.
Investigate whether a billing-frequency issue overstates prepaid multi-year contracts. Read the brief and source code, run about 30 read-only warehouse queries, and write the findings.
Prepare for a meeting, then compose the follow-up email.
Prepare a two-page interview briefing from recruiting records, calendar, email and research. Include assessment scores and a profile link, and stage a short email to three teammates.
Compare internal job costing with a third-party vendor, using live Snowflake data to examine job counts, print volumes, and potential savings.
Create a new artifact from three Excel files, incorporate two rounds of feedback, and draft follow-up emails to the team.
Change the data in an HTML dashboard, then create a shareable version.
Check 66 target accounts for news using browser research and search. Draft an approximately 1,100-word digest in Outlook and make 21 workbook edits across 12 rows.
Research a prospect, its sponsor and people. Adapt six slides in a 16-slide overview deck, write a roughly 2,000-word call-preparation sheet, and refresh a prospect context file.
Sweep meeting notes, update open items in the project-management system, and organize project files: read 42 Asana tasks and 13 meeting notes, create 7 tasks, update 4, close 6, and create 10 files.
Turn a synthetic general ledger into a nine-sheet workbook, seven-slide deck and PDF, one-page memo, rebuild scripts and an unsent cover email. Work through four scripted review rounds, including a cost reclassification and a reconciliation.
$10+
5 workflowsPlan a research launch, develop three page-design options, prepare an audience workbook and launch documents, and edit a draft research page through the browser.
Review the past two months of emails, identify items not yet logged, stage them for logging, and flag information needing follow-up.
Iterate on an existing reconciliation workbook (18 sheets, about 23,000 rows and 73,000 formulas), then create a PDF summary.
Continue the research launch in a new session: iterate on pages, set up a newsletter, prepare launch assets, update Asana and draft an Outlook email.
Develop and build five management tools: a 13-week cash forecast, collections cockpit, weekly flash, owned-exception feed, and value-creation tracker.
These are 23 sessions we ran on Opus 5.5: 21 in Claude Code and 2 in Cowork. Illustrative only.
The pattern behind the numbers is intuitive. Longer tasks cost more. Tasks take longer because they have to do more: read more files, carry more context, and go back and forth more times. Tools also appear to add a fixed cost before any work starts. An NDA review, where Claude read one document and answered in the chat window, cost $0.43. Sending a Microsoft Teams invitation, which needs a connector, cost $1.69, about four times as much.
None of this is surprising. What it does not show is how much each choice moves the meter. So we pulled apart the factors a user controls and tested them one at a time.
When you are using Claude or ChatGPT, there are many different choices you make along the way. We wanted to test how each of these impacts the cost of running the same task.
Claude: Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 4.5 · GPT-6: Luna, Sol, Astra
Fable cost about 3.5× as much as Opus.
$2.97 versus $0.86 estimated for the same three finance tasks. This compares the task set, not quality or each task individually.
Compare the combined finance task set, or explore Task A, B, and C separately. All costs use the workbook’s rates.
Fable has recorded token counts for each task. The other six models have combined token totals. The workbook allocates those totals using Fable’s task proportions, with an extra 10 percentage points assigned to Task C for budget models and five for mid-tier models. These are modeled splits, not separate task measurements.
Costs use the workbook’s historical rates and token-mix assumptions. Fable’s rate-based total is $2.97; its app-reported total is $5.81. That difference remains unreconciled. The task allocations cannot establish whether observed rankings are consistent across simple and complex tasks.
Chat · Cowork · Code
Cowork was cheapest in both tests.
Across the two workflows combined, Cowork cost 46% less than Code and 23% less than Chat.
When you use Claude Desktop, you have three ways to work: Chat, Cowork, and Code. You can do many of the same things in each. Our question was whether where you choose to work affects cost.
We ran two tasks in Chat, Cowork and Code on Enterprise with Opus 5.5. Cowork had the lowest reported cost on both.
Reported usage cost
$1.21Reported usage cost
$1.03Reported usage cost
$1.91Task B: clean the synthetic ledger, build the workbook and supply a rebuild script. Chat and Code each required one additional user input; Cowork required none.
Memory on (rules in memory) · Memory off (same rules pasted into the prompt)
Memory on cost about 35% more.
$0.58 versus $0.43 per run for the same KPI brief, and about a third faster. Two runs per side, so treat this as directional.
Memory lets Claude carry what it knows about you and your clients from one chat to the next. We wanted to know what that costs compared with giving Claude the same instructions in the prompt.
We asked Claude.ai (Opus 5.5) to write a weekly KPI brief for a fictional client from an 11-row sales table, following 10 reporting conventions. With memory off, we pasted the conventions into the prompt. With memory on, the conventions lived only in memory. Two runs each, 27 October 2026.
Claude.ai · Opus 5.5 · read from each chat’s session panel
The two memory-on runs differed by 360,000 input tokens, more than the gap between the two sides. Past-chat search was on, so memory-on runs could also find the rules in an earlier chat, and reply quality was not scored. A rerun with five runs per side is planned.
Clean context: a fresh session with only the task · Loaded context: the same task after ~710,000 tokens of unrelated text were loaded into the session
Loaded context cost about 11× clean context.
For a simple task, a lot of preloaded context added about $5 to the cost.
A conversation carries context forward. We wanted to understand how starting with a small or large amount of context changes the cost of doing the same work.
We ran the same task in Cowork twice: once in a fresh session (clean context), and once after loading about 710,000 tokens of unrelated text into the session (loaded context). Claude reads everything in the session each time it answers, so the unrelated text adds cost without helping.
Writing task: $0.47 clean versus $5.37 loaded. Both readings include session setup and the task.
CSV vs. Snowflake connector · Text answer vs. Excel file vs. HTML artifact
An HTML artifact cost about 3× a text answer.
$4.41 versus $1.48 for the same analysis request in Code.
One of the most powerful things about Claude is its ability to connect to other systems and use tools to produce visualizations, Excel files, and other outputs. We wanted to understand how those choices affect cost.
Team · Claude Code · Opus 5.5 · memory on
“Answer in chat” means text returned inside Claude Code. These runs do not compare the Chat, Cowork, and Code environments.
Reported usage cost · each comparison has its own scale. Expand a row for its report.
Analyze 2026 revenue trends for the top 10 customers and their product mix. Return the answer in chat.
CSV export and preparation time are not included.
The reported 110 output tokens needs a coverage check; it is much lower than the CSV run’s 5.7k.
The CSV run excludes manual export time. The connector report’s token coverage needs review; this pair does not isolate the cost of a tool call.
Analyze monthly customer revenue and margin from January to September 2026, with commentary on volume, costs, and product mix.
Output quality and matched-input checks have not been independently assessed.
Output quality and matched-input checks have not been independently assessed.
The reported 287 output tokens needs a coverage check for an HTML artifact.
I still don’t know why my five-year-old kept winning those races. I was happy to let him enjoy it until our lane was ready.
We are happy with the value we are getting from AI, too. Even at full API prices, we would consume in roughly the same pattern. But as we help clients, we have already come up with ways to get more efficient in consumption and decrease cost. Our overarching recommendation is that prioritizing adoption and learning the tools is far more important than containing the costs for most organizations at this point in time. We are going to keep experimenting and building a toolkit to help organizations navigate this new arcade.
