Perplexity: The Running Meter

A currency-pegged wallet with world-class lifecycle plumbing — and no rate card.

Editorial disclosure. This case study was researched and drafted with Perplexity Computer — the product whose credit system it audits. The analysis relies exclusively on public documents, published prices, press coverage, and user commentary, all cited in the source register; the agent had no access to non-public pricing information.

Case Facts

CNI: Credit Native Index

CGVS: Credit Governance Visibility Score

In One Breath

When Perplexity launched Computer — an agent that orchestrates nineteen models to run workflows for hours or months — it bolted its first consumption meter onto a subscription business: credits pegged at 100 to the dollar, included monthly on the $200 Max plan, purchasable à la carte on the $20 Pro plan. The lifecycle engineering is among the best in the field: tasks pause instead of dying, negative balances are forgiven, auto-refill is opt-in with user-set monthly limits, draw order is published, and error refunds are policy. But there is no rate card. Credit burn is a dynamic pass-through of compute cost — "examples, not guarantees" — so users discover the price of a task only by running it. The market's complaints target exactly that gap, and the company's own refund policy states the philosophy in one line: credits pay for computing power, not outcomes.

1. The Business and the Buyer

Perplexity built its brand on answer-engine search, then moved up the autonomy spectrum: the Max tier ($200/month or $2,000/year) arrived in July 2025 as a classic premium-access play. On February 25, 2026 the company launched Computer, "a general-purpose digital worker" that decomposes goals into subtasks and delegates them across 19 models — Claude Opus for orchestration, Gemini for research, GPT-5.2 for long context, Veo for video. Computer is the company's first consumption-billed consumer product — the moment an all-you-can-eat search subscription met workloads whose marginal cost varies a hundredfold by task.

The buyer is a prosumer or professional individual: researchers, analysts, content creators, "business strategists conducting competitive intelligence." The reference value is genuinely dual, and that duality drives the whole value test. Against human alternatives — a $500+ contractor deliverable, a $30,000/year terminal — a $10–$25 task is absurdly cheap. Against the flat-rate agent subscriptions of Anthropic and OpenAI, a metered $200/month that one task can exhaust looks expensive, and users say so explicitly.

The archetype is Orchestration Platform — the archetype the book warns faces the hardest weight-design problem, because the action space is unbounded and the underlying cost structure is a portfolio of other vendors' meters.

2. The Model As Built

Foundation (C1–C4). The unit is an abstract credit pegged to currency: 100 credits = $1, everywhere. There is no weight table. Burn is a function of task complexity, published only as indicative bands:

The help center is candid: "These are examples, not guarantees. Your actual usage may differ." Regular Ask searches remain unmetered — like valueIQ's unmetered co-pilot, the conversational front door is free and the agentic work behind it is charged.

Plan structure splits into two different Tanso patterns. Max ($200/month) includes 10,000 monthly credits — $100 of face value, half the plan price — a true credit-subscription hybrid. Pro ($20/month) includes no monthly credits at all: it is a seat-led gate with pay-as-you-go credit purchases attached, seeded by a one-time 4,000-credit bonus that expires in 30 days (Max's one-time bonus is 35,000, up from 20,000 at launch). Enterprise Pro and Enterprise Max include 500 and 15,000 monthly credits respectively. One anomaly: EU credit packs show a mild inverse volume curve — €4.99 buys 500 credits while €259.99 buys only 25,000, so the per-credit price rises with pack size.

Lifecycle (C5–C13). This is where the model shines. When credits run out, active tasks pause — they do not cancel — and resume automatically when the balance is restored; work in progress is never destroyed. Balances may briefly go negative without penalty. Auto-refill is off by default, opt-in, with a user-set refill amount and a hard monthly auto-refill limit; every Max account carries a monthly spending cap (default $200, can be raised to $5,000). Draw order is published: Bonus, then Plan, then Purchased. Monthly credits do not roll over; purchased credits persist until used, expiring only after a year of inactivity; bonus credits expire on their stated date. Credits are personal and non-transferable — there is no pooling. Refund policy is unusually explicit: credits are restored when a task fails from a platform error; they are not refunded when a completed task disappoints, because — in the policy's own words — "the credits paid for the computing power that ran the task."

Governance and GTM (C14–C17). A usage dashboard shows balances by credit class, days to refresh, per-thread and per-media-type breakdowns, and per-task usage; five distinct email notifications cover purchases, low balance, exhaustion, spending-cap hits, and payment failures. What is missing is the front end of the loop — no pre-task cost estimate, no simulator, no machine-readable pricing publication — and the back end of trust: the change protocol is one sentence ("we'll update this page... if there are any significant changes"), and practice has been worse than the promise. In February 2026 the $5 monthly API credit bundled with Pro was removed silently, with no announcement — users discovered it through 401 errors.

3. The Design Audit

The central staircase chain: the C1 decision not to publish weights — defensible for an orchestration platform whose costs are genuinely stochastic — pushes all variance management downstream. And to the company's credit, downstream is where the engineering went: C5's pause-and-resume, spending caps, negative-balance grace, and C9's error refunds are precisely the mechanisms a non-deterministic meter needs. This is a coherent package: manage variance at the wallet, since you cannot eliminate it at the meter. But the DSM's logic still bites. A wallet-layer fix cannot give a buyer — human or agent — a forward price. The result is documented burn shock: a Max subscriber's single codebase task consumed 15,000 credits (150% of the monthly allowance, $150 face value) in 40 minutes, then another 2,000 attempting to push results to GitHub; another user burned 15,000 in under an hour of experimentation.

There is a second-order agency problem the book's autonomy chapters predict: when the metered party is the agent itself, the agent's inefficiency is the customer's bill. The GitHub-push episode — subagents spawned, strategies retried, 2,000 credits spent on delivery mechanics — is the customer paying for the vendor's agent to struggle. No published policy addresses burn caused by agent thrash short of a "task URL to support" review.

4. The Value Test

EVE reconstruction, prosumer segment. The reference value is dual. Against human labor: a heavy task ($8.75–$22.75) substitutes for deliverables the market prices at hundreds of dollars; press coverage seriously entertained a $30,000/year terminal comparison. On that axis VtB is comfortably above 10 — the pricing is a bargain, and the MDNRC CIO testimonial ("we saved more than the cost of the annual licensing in the first week") sits on the pricing page as evidence. Against the competing reference — flat-rate agent subscriptions from Anthropic and OpenAI — the same meter reads differently: the codebase-repair user reported that rival subscriptions would run the equivalent workload "likely several hundred times... for the same $200/month.” For high-volume coding workloads, VtB dips toward or below 1. The meter is calibrated to cost, so value capture is accidental: light-but-valuable tasks are underpriced, heavy-but-marginal tasks are overpriced, and the customer self-selects accordingly — adverse selection that flat-rate rivals absorb structurally.

The two dials confirm it. Dial one is welded to the peg: $0.01 per credit, no volume curve (or a slightly inverted one in the EU). Dial two — the weights — does not exist as a design surface; it is whatever the orchestration stack happens to spend. Nothing in the system can express "a pitch deck for a funded startup is worth more than a pitch deck for a hobbyist." The autonomy risk discount, meanwhile, is charged rather than absorbed: outcome risk sits entirely with the buyer, per the refund policy's computing-power language. OAR is not low — it is explicitly disclaimed.

5. What the Market Did Next

Adoption came fast and loud — Computer was the company's biggest launch, covered across the tech press — and the pricing discourse organized itself around four themes, none of which is "the price is too high":

1. Burn shock and illegibility. The dominant complaint: users cannot forecast a task's cost before running it — "you don't know the cost until the project is running"; credit-versus-token confusion recurs.

2. Change-management wounds. The silent removal of Pro's API credit produced cancellations framed in trust language: "a simple heads-up would have been helpful... it makes them seem unreliable."

3. Expiry and forfeiture friction. Bonus expiries and cancellation forfeitures read as "deceptive" to users who did not find the terms; one user's credits were restored only after weeks of support contact.

4. Genuine delight at the value pole. The same market produced the Bloomberg-terminal comparison and enterprise ROI testimony — evidence that when the task lands, the meter is nearly invisible.

Note what is absent: no meaningful complaint about the peg, the credit abstraction, or the wallet mechanics. The lifecycle engineering worked. The grievances cluster precisely on the two cells the audit marks Dodged — C2 predictability and C16 change management.

Scorecard

6. Lessons for the Practitioner

1. A currency peg is not a rate card. 100 credits = $1 makes the wallet legible and the work illegible. Legibility must exist at the level where buyers decide — the task — or the meter reads as a taxi meter no matter how honest the peg.

2. World-class lifecycle plumbing can contain, but not cure, a non-deterministic meter. Pause-don't-cancel, negative-balance grace, opt-in auto-refill, and spending caps are the best variance-management stack observed in this chapter — and burn shock still dominates the discourse, because variance management is not price discovery. 

3. When the agent spends the credits, agent inefficiency is a pricing problem. Two thousand credits spent retrying a GitHub push is not a support ticket; it is a metering-design question the industry has not answered: who pays when the digital worker struggles?

4. Change management is judged by practice, not policy. One silent perk removal did more trust damage than every expiry rule combined — and the users it burned were integrators, exactly the constituency an agent platform needs.

5. Cost-calibrated credits invite adverse selection against flat-rate rivals. When burn tracks compute, customers route heavy workloads to competitors' buffets and keep light ones on your meter. The escape is the second dial — weights calibrated to value, by segment — which requires admitting you have a weight table at all.