Chapters
Forecast work and quality separately
Start with a completed workflow you can recognise. A research brief, a code change and a financial analysis each need a stated acceptance standard. Use the same work unit when you estimate demand and capacity.
Record how often each task occurs and how much work each run involves. Include context, tools, retries and review. Longer sessions and parallel agents can increase consumption. State whether a session count already includes those agents so that you count their work once.
Then set the quality floor for each task. Classification needs labels checked against examples. Financial analysis needs reproducible calculations. Fabricated case law can invalidate legal work. A rare task can still require access to your most capable model.
Frequent work and high-stakes work shape the portfolio
Illustrative workloads · Relative frequency has no numerical scale
Routine classification
High frequency
Its quality floor
Labels checked against examples
Test whether a smaller model can meet this bounded rule.High-stakes financial analysis
Low frequency
Its quality floor
Reproducible calculations tied to sources
Retain access to a model that passes this standard.Keep task importance separate from volume. A frequent routine task can account for most consumption while an occasional high-stakes task determines which models you must retain. Image generation also needs its own workload and quality evidence.
Start with a rough forecast. Session records should refine it when the purchase warrants the effort. Measure complete workflows on the models and settings you intend to use.
Check how access is sold
A retail subscription can provide substantial use for a fixed monthly fee. The bill remains fixed within its included use, while access depends on the plan's limits. Buying equivalent use through an application programming interface, or API, can cost more. The difference depends on the work and the commercial terms.Claude Enterprise describes usage-based seat-plus-usage billing and distinguishes legacy arrangements. OpenAI billing separates ChatGPT subscriptions from API charges, including eligible promotional credits. Checked on 15 November 2026. The retail comparison concerns effective API-equivalent use. No provider cost, subsidy or hidden consumer cap is measured here. Full source note 1
Metered access charges for consumption. An enterprise arrangement may add a seat fee before usage charges. Claude's usage-based Enterprise plan is one such arrangement. OpenAI also keeps ChatGPT subscriptions and API billing separate. Check the contract you will buy and any included credits.
Keep three quantities distinct. An API token price is a tariff for a specified kind of token. A benchmark median task cost records what a task cost in that benchmark setting. A subscription pool debit measures how much of an allowance your work consumes. Converting between them needs evidence about the model, workflow and plan.
Cursor's legacy plans provide a concrete example. Their Cursor Models and Other Models pools reset monthly. Third-party models consume the Other Models pool at their API prices. A different allocation can therefore change how much work the same pool supports. Cursor now distinguishes these legacy terms from its newer offering.Legacy Cursor plans identifies the two monthly pools and third-party debits at API prices. Current models and pricing distinguishes the newer offering. Checked on 15 November 2026. Cash fees and pool consumption are separate. The Cursor example in my earlier working notes concerns the legacy arrangement. Its Max Mode reference concerns context and billing, a different setting from a model's maximum reasoning effort. Full source note 2
Check the exact model, effort setting, interface and required tools. Access through one application says nothing about the tools in another. Include required extras in the bill before testing the budget.
For a company, group people by workload and intensity. Aggregate usage under the actual terms. A shared allowance and a per-seat allowance produce different capacity forecasts. Avoid adding allowances that the contract prevents people from sharing.
Compare plans and allocations together
I built the Model Selector to compare purchases against a workload. Its three views support different parts of the decision. Selection tests your work and quality requirements. Token plans shows the catalogue. Plan value compares useful work against the fee.The three views and their assumptions are documented in the interactive model selector and procurement guide. Checked on 15 November 2026. The selector combines recorded task scores, Arena user feedback, benchmark task costs and inferred allowances. Its estimates depend on those inputs and their dates. Arena's task-cost method samples at most the first three tasks in a session. The historical Arena capture was taken on 11 October 2026 at 22:34 HKT, with rankings dated 8 October and median task costs over the preceding 14 days. That historical capture establishes no current subscription capacity. Full source note 3
The purchase needs five connected inputs and outputs. Work I need records demand. Quality I require determines which allocations qualify. Budget and access constrain the purchases. Compare plans and allocations tests those combinations. Recommended mix reports the purchase, allocations, bill and headroom.
Build a purchase from the workload
Declare complete workflows
Forecast days, sessions, tasks and task mix. Record the size of each workflow, including context, tools and retries.
Count parallel agents once. Compare demand and capacity in the same work unit and time window.
Set a floor for each task
Keep importance independent of frequency. A rare high-stakes task may require access to a stronger model.
Retain only allocations with the exact model, effort, tools and quality evidence the task needs.
Keep eligible purchases
Add plan fees and required extras before checking the monthly budget. Check which models, interfaces and tools each purchase permits.
Record provider requirements and the maximum number of plans. Test each shared pool, model cap and reset window separately.
Test capacity and reserve separately
Usable capacity − Required work = Headroom
Assign the declared task mix to eligible model and effort pairs. The largest fraction of a governing limit determines coverage.
Check usable capacity against 1.5 times demand in this provisional example. Check each quality floor and the budget independently.
Report the purchase and its assumptions
Return the plans, model allocations, monthly bill and estimated headroom. Keep the objective explicit.
In the synthetic example, the quality-first rule selects Plan B with Q High. Minimum cost at acceptable quality selects Plan A with E Medium. Both cover the reserve.
For each candidate, assign the forecast work to eligible model and effort pairs. Sum consumption within each governing pool. Check shared allowances, model caps and every reset window. Changing an allocation can change both quality and usable capacity.
Usable capacity minus required work equals headroom. A positive result covers average demand. A reserve check asks whether capacity also covers the chosen buffer. A budget check tests the bill, including required extras. Each task must still meet its quality floor.
The selector's capacity estimates use benchmark task costs as proxies and inferred allowances. Unknown limits remain unknown. Use the estimates to narrow the purchase, then compare the forecast with actual usage. A monthly allowance alone cannot establish capacity during a busy short window.
Work through a synthetic purchase
The example below uses synthetic plans, models, prices, quality scores, consumption rates and limits. It shows the method. It describes no current provider offer.All example plans, models, scores, fees, usage rates and limits are synthetic inputs from my earlier working notes. Workflow counts, total consumption, capacity and headroom are derived from them. Capacity preserves the declared mix and timing and is rounded to one decimal place. Quality uses importance weights independently of volume. The quality-first rule differs from minimum cost at acceptable quality. Required extras are zero. The four supplied allocations illustrate the rule; they do not represent a full catalogue search or enforce provider redundancy. Full source note 4
Assume five active days, four sessions a day, one agent and one task per session. That is 20 completed workflows a week. The monthly budget is US$200 with no required extras. A provisional reserve of 1.5 times average demand requires capacity for 30 workflows.
| Task | Weekly workflows | Volume share | Importance weight |
|---|---|---|---|
| Research | 8 | 40% | 7 |
| Coding | 6 | 30% | 6 |
| Analytics | 4 | 20% | 4 |
| Slides | 2 | 10% | 3 |
Model Q High and Model E Medium each receive the same forecast. The quality floor is a synthetic score of 75 on every task.
| Task | Q High units per workflow | E Medium units per workflow | Q High score | E Medium score |
|---|---|---|---|---|
| Research | 6 | 3 | 90 | 86 |
| Coding | 4 | 2 | 88 | 82 |
| Analytics | 3 | 1.5 | 87 | 80 |
| Slides | 2 | 1 | 85 | 78 |
At this volume, Q High consumes 88 allowance units per week. E Medium consumes 44. The importance-weighted quality score is the sum of each task score times its importance, divided by total importance. Volume has no part in that quality weight. The resulting scores are 88.1 for Q High, rounded, and 82.4 for E Medium.
Plans A and B represent separate providers. Their pools remain independent. Research use counts against the weekly pool. It also has its own weekly cap. Each plan also has a peak limit.
| Synthetic plan | Monthly fee | Weekly pool, units | Research cap, units per week | Peak cap, units per minute |
|---|---|---|---|---|
| Plan A | US$80 | 96 | 80 | 10 |
| Plan B | US$140 | 144 | 80 | 16 |
Q High has assumed peak consumption of 8 units per minute. E Medium has 4. The example scales the declared workload and timing together when estimating capacity. A different timing pattern needs a fresh check.
For each allocation, divide full-forecast consumption by each matched limit. The largest fraction is the binding constraint. Its reciprocal is the coverage multiple. Multiply the 20 required workflows by that multiple to estimate usable work, then subtract 20 for headroom.
Both Plan A with E and Plan B with Q cover the reserve
Synthetic comparison · Reserve marker at 30 workflows per week · Bar scale 0 to 45
| Candidate allocation | Monthly bill | Usable work per week | Quality score | Result |
|---|---|---|---|---|
| Plan A · Model Q High | US$80 | 21.8 | 88.1 | Covers average work; reserve falls short |
| Plan A · Model E Medium | US$80 | 43.6 | 82.4 | Covers work and reserve |
| Plan B · Model Q High | US$140 | 32.7 | 88.1 | Recommended under the quality-first rule |
| Plans A + B · Model Q High | US$220 | Excluded | 88.1 | Over the US$200 budget |
Plan A with Q High uses 88 of its 96 weekly units. That limits capacity to 21.8 workflows. E Medium halves consumption and raises estimated capacity to 43.6. Both allocations pass the quality floor, but only E Medium reaches the 30-workflow reserve target on Plan A.
Plan B with Q High provides capacity for 32.7 workflows a week. Headroom is 12.7 workflows. It covers the 30-workflow reserve target and retains the higher quality score. Research and peak use also fit their limits.
Plan B with E Medium is omitted on purpose. It matches Plan A with E Medium on quality, adds capacity the reserve does not require and costs US$60 more a month.
The example's default rule prefers purchases that cover work and reserve within budget, then higher quality. It prefers the cheaper purchase within 0.5 quality points. That rule recommends Plan B at US$140 a month.
Plan A with E Medium also meets the quality floor and reserve for US$80. Choose it if your objective is minimum cost at acceptable quality. The extra US$60 for Plan B buys 5.7 synthetic quality points. Better accepted output or less review and rework would need to justify that premium.
Buying both providers costs US$220 and fails the budget check. The example therefore provides no funded redundancy. Requiring a second provider changes the eligible purchases or the budget.
Check capacity within each window
Average demand needs headroom. My starting capacity target is 1.5 times estimated average use, giving a 50% buffer. It is a provisional assumption that I would test against observed variation and failures at limits.The 1.5 times target adds 50% to average demand. It is the author's provisional starting assumption. It gives no quantified probability of sufficient capacity. Test it within each governing window using the observed variation and cost of running short. Full source note 5
A weekly total can conceal a short-window shortage. Check how work clusters around deadlines, how long sessions last and how many agents run together. Apply the reserve once within the forecast's declared scope. Keep stress cases and peak checks visible alongside the average.
Refine expected consumption with measured mean usage of complete workflows. The mean helps forecast a total across repeated work. A benchmark median describes the middle observed task and can miss expensive workflows. The original essay's 11 October 2026 Arena capture is historical evidence, with a different task boundary from your own work.
If usage limits remain undisclosed, record the observation and its date. A capacity estimate based on observed use remains conditional on that workload and access arrangement. Check the intended application and reset window before renewal.
Give smaller models a tested role
Include open-weight and small language models where they meet the work's standard. Open-weight means you can obtain the model's weights. It still needs an exact release, licence and deployment record. A small model needs a bounded task and evidence that its output passes.
Local operation can be a purchasing requirement. Compare hardware, hosting, operation and review costs with hosted access. Test the actual runtime, tools, context and concurrent workload. Owning the weights leaves those questions open.
The selector holds no weekly cost figure I can use for a locally run model, so this essay makes no claim for its capacity or saving. The open-weight guide and small language models guide set out the records needed before including them in the portfolio.
Buy continuity across providers
I also need to trust the provider. Anthropic has shut me off three times in 2026 without explaining why. That experience changes what I am willing to depend on.
My answer is multiple models and redundancy across providers. A second subscription through the same underlying supplier can leave the dependency intact. Check which provider supplies the model and whether the alternative can complete the work with its required tools.
A supplier can also enter its customers' market. Cursor uses Anthropic models while Anthropic has its own coding product. Chegg announced a GPT-4 study companion before OpenAI introduced ChatGPT study mode. I treat that overlap as a supplier-competition risk.Cursor's model documentation lists Anthropic models. Anthropic introduced Claude Code in February 2025. Chegg announced CheggMate with GPT-4 in April 2023. OpenAI introduced study mode in July 2025. Checked on 15 November 2026. These establish product overlap. Supplier-competition risk is my inference; the sources establish no copying or intent. The access interruptions in the text are my reported experience. Full source note 6
Continuity has a cost. Include the alternative's fee and the effort of keeping it usable. A single plan may be sufficient for a workload that can tolerate interruption. Where interruption would prevent important work, test the alternative before depending on it and include its capacity in the purchase decision.
Review before each renewal
Compare the forecast with accepted work, total spend, review time, failed attempts and unused capacity. Record where limits stopped the work and whether the alternative provider could take over.
Revisit the purchase when the workload, model capability or commercial terms change. Better methods can reduce consumption. Greater skill at delegation can also create more useful work worth funding. A higher total bill can be justified by the additional accepted output.
For an individual, you own that review. In an organisation, name someone to maintain the evidence and revisit renewals across workload groups. Count the cost of this process against the value it produces. An elaborate portfolio has to justify the effort of maintaining it.
The purchase leaves a daily allocation decision. Model routing, the next companion essay, explains how I choose an exact model and effort for each task.
Key takeaways
- Forecast complete workflows and set quality floors independently of task volume.
- Compare purchases and model allocations together under the actual pools, caps and reset windows.
- Keep cash fees, API token prices, benchmark task costs and subscription consumption distinct.
- Test headroom, reserve, quality, budget and continuity as separate conditions.
- Reassess the portfolio before renewal using accepted work and actual usage.
Sources and notes
-
Claude Enterprise describes usage-based seat-plus-usage billing and distinguishes legacy arrangements. OpenAI billing separates ChatGPT subscriptions from API charges, including eligible promotional credits. Checked on 15 November 2026. The retail comparison concerns effective API-equivalent use. No provider cost, subsidy or hidden consumer cap is measured here. ↩
-
Legacy Cursor plans identifies the two monthly pools and third-party debits at API prices. Current models and pricing distinguishes the newer offering. Checked on 15 November 2026. Cash fees and pool consumption are separate. The Cursor example in my earlier working notes concerns the legacy arrangement. Its Max Mode reference concerns context and billing, a different setting from a model's maximum reasoning effort. ↩
-
The three views and their assumptions are documented in the interactive model selector and procurement guide. Checked on 15 November 2026. The selector combines recorded task scores, Arena user feedback, benchmark task costs and inferred allowances. Its estimates depend on those inputs and their dates. Arena's task-cost method samples at most the first three tasks in a session. The historical Arena capture was taken on 11 October 2026 at 22:34 HKT, with rankings dated 8 October and median task costs over the preceding 14 days. That historical capture establishes no current subscription capacity. ↩
-
All example plans, models, scores, fees, usage rates and limits are synthetic inputs from my earlier working notes. Workflow counts, total consumption, capacity and headroom are derived from them. Capacity preserves the declared mix and timing and is rounded to one decimal place. Quality uses importance weights independently of volume. The quality-first rule differs from minimum cost at acceptable quality. Required extras are zero. The four supplied allocations illustrate the rule; they do not represent a full catalogue search or enforce provider redundancy. ↩
-
The 1.5 times target adds 50% to average demand. It is the author's provisional starting assumption. It gives no quantified probability of sufficient capacity. Test it within each governing window using the observed variation and cost of running short. ↩
-
Cursor's model documentation lists Anthropic models. Anthropic introduced Claude Code in February 2025. Chegg announced CheggMate with GPT-4 in April 2023. OpenAI introduced study mode in July 2025. Checked on 15 November 2026. These establish product overlap. Supplier-competition risk is my inference; the sources establish no copying or intent. The access interruptions in the text are my reported experience. ↩