Effort selector controls are on their way to Perplexity Computer — at least if a fresh leak is right. On 23 August 2026, the feature-tracking account TestingCatalog spotted Perplexity working on a granular effort selector for its agentic Computer product, including a low effort mode aimed squarely at one thing: the amount of credits the agent burns while it works.
The sighting comes from a TestingCatalog post published on Threads, and its wording is worth quoting exactly: “Perplexity is working on a new granular effort selector for Perplexity Computer. Being able to run Computer on low effort should have a positive impact on credit consumption. This looks familiar 👀.”
That last wink matters. Every major AI lab now ships some form of dial that trades answer quality against compute spent, and an effort selector for a full agentic computer is the logical next step. This article unpacks what the leak actually shows, why credit consumption is the sore point it targets, how rivals already handle effort controls, and what a granular effort selector would mean for businesses that have been watching Perplexity Computer with interest — and watching its bills with alarm.
Table of contents
- What the Effort Selector Leak Actually Shows
- Perplexity Computer: The Product Behind the Effort Selector
- Why Perplexity Computer Needs an Effort Selector
- How Rival AI Platforms Handle Effort Controls
- Perplexity Already Ships an Effort Selector in Its API
- The Bigger Pattern: Perplexity’s Push to Cut Computer Costs
- What an Effort Selector Could Mean for Your Business
- The Stakes: Computer Is Carrying Perplexity’s Growth
- Effort Selector FAQ
- References
What the Effort Selector Leak Actually Shows
Accuracy first: this is a spotted-in-testing feature, not an announcement. Perplexity has said nothing officially, there is no release date, and unreleased features change or vanish. What we can pin down comes from the post itself.
The claims the post confirms
Three things are stated outright. Perplexity is working on an effort selector for Perplexity Computer. The selector is granular rather than a single on/off switch. And at least one of its settings is a low effort mode, framed explicitly as a way to reduce credit consumption.
What is not confirmed
The exact level names are not stated in the post text, so nobody outside Perplexity can yet say whether the effort selector will read low/medium/high, mirror the four-step pickers rivals use, or expose something finer. Treat any specific level list you see elsewhere as speculation.
| Claim | Status | Basis |
|---|---|---|
| A granular effort selector is in the works for Perplexity Computer | Stated in the leak | TestingCatalog post, 23 August 2026 |
| A low effort mode is included | Stated in the leak | TestingCatalog post, 23 August 2026 |
| The goal is lower credit consumption | Stated in the leak | TestingCatalog post, 23 August 2026 |
| Exact effort level names | Not confirmed | Not stated in the post text |
| Release date or rollout plan | Not confirmed | No official Perplexity statement |
Who spotted it, and why that matters
TestingCatalog has a strong record of surfacing real features weeks before launch, from unreleased model options to interface experiments — it is the same account that first flagged the mystery Ox Alpha model before its identity was confirmed. A TestingCatalog sighting is not a guarantee, but it is usually smoke from a real fire.
Perplexity Computer: The Product Behind the Effort Selector
To see why an effort selector is such a pointed addition, you need to know what Perplexity Computer is and what it costs to run.
From answer engine to agentic computer
Perplexity built its name on AI-powered search, then shipped the Comet browser in July 2025 — first as a $200-a-month Max exclusive, then free worldwide from October 2025. Perplexity Computer, launched on 25 February 2026 for Max subscribers, is the next step in that agentic push: a cloud-based AI computer rather than a browser feature. We covered the launch in detail in our Perplexity Computer overview.
What Computer actually does
Computer runs tasks inside an isolated Linux sandbox in Perplexity’s cloud, orchestrating a roster of 19 frontier models — with GPT-5.5 as the default orchestrator since May 2026 — and connecting to more than 400 services through OAuth connectors. It researches, writes code, builds files and websites, and chains multi-step work the way autonomous AI agents are supposed to: you describe the outcome, it plans and executes. A companion Personal Computer product, rolled out to Max subscribers in April 2026, brings a local version to the Mac.
Who can use it today
Access arrived in stages: Max subscribers first, then Pro subscribers from mid-March 2026 with bonus credit allocations. Every tier meters Computer usage in credits, and that metering is where the effort selector story really begins.
Why Perplexity Computer Needs an Effort Selector
Run any agentic workload for a week and the pattern is obvious: capability is rarely the complaint. Cost is. An effort selector attacks the cost side directly, so the credit arithmetic deserves a close look.
What the credits cost
| Plan | Price | Included Computer credits |
|---|---|---|
| Pro | $20/month | None included — bought separately |
| Max | $200/month | 10,000 per month |
| Enterprise Pro | $40/seat | 500 per seat |
| Enterprise Max | $325/seat | 15,000 per seat |
Auto-refill defaults to $200 a month when credits run dry, configurable up to a $2,000 cap — so a busy month does not stop, it invoices.
How fast the credits burn
Published hands-on numbers make the burn rate concrete. A DataCamp reviewer’s single research task, touching eight tools over just under eight minutes, consumed 225.71 credits. A complex due-diligence workflow burns around 500 credits — meaning a 10,000-credit Max month buys roughly twenty of them. One reviewer at Builder.io burned through $200 of credits in two days building a website.
The problem an effort selector solves
Today, Computer decides for itself how hard to work, and every task gets the full agentic treatment whether it needs it or not. Summarising one page and auditing a supplier both draw from the same metered pool at whatever depth the agent chooses. A granular effort selector hands that decision back to the user: spend heavily where the answer must be exhaustive, run on low effort where a quick pass will do.
How Rival AI Platforms Handle Effort Controls
The “this looks familiar” aside in the leak points at an industry-wide pattern. Reasoning models are trained with reinforcement learning to decide how long to think before answering — and every major provider has concluded that users need a dial over that behaviour.
| Provider and product | Control | Levels |
|---|---|---|
| OpenAI API | reasoning_effort | none, minimal, low, medium, high, xhigh, max (model-dependent; GPT-5.5 and 5.6 default to medium) |
| ChatGPT (consumer) | Thinking-time picker | Light, Standard, Extended, Heavy (all four on Pro) |
| Anthropic API | effort | low, medium, high, xhigh, max (default high) |
| Google Gemini API | thinking_level | minimal, low, medium, high (model-dependent) |
| xAI Grok API | reasoning_effort | low, medium, high, xhigh (default high) |
| Perplexity API (sonar-deep-research) | reasoning_effort | low, medium, high |
| Perplexity Computer | Effort selector | “Granular”, at least low — in testing, unannounced |
OpenAI set the template
OpenAI introduced reasoning effort with its o1 models in December 2024 and has kept widening the range since; GPT-5 added a minimal setting in August 2025, and the current API spans seven values. On the consumer side, ChatGPT’s thinking-time picker gives Pro subscribers four steps from Light to Heavy.
Anthropic proved the economics
Anthropic’s effort parameter, introduced with Claude Opus 4.5 in November 2025, came with the clearest numbers in the industry. At medium effort, Opus 4.5 matched Sonnet 4.5’s benchmark score while using 76% fewer output tokens; at highest effort it beat that mark by 4.3 points while still using 48% fewer tokens. Taking Sonnet 4.5’s output as the 100% baseline, the token consumption looks like this:
Crucially, Anthropic’s effort control shapes every token the model produces, including tool calls — lower effort means fewer tool invocations. That is precisely the mechanism that matters for an agent like Computer, where tool calls are what credits pay for.
Google and xAI complete the picture
Google exposes a thinking_level control across current Gemini models, and xAI’s Grok offers reasoning effort from low up to xhigh on its newest releases. The pattern is unanimous: nobody ships maximum reasoning as the only option any more.
Perplexity Already Ships an Effort Selector in Its API
Here is the detail most coverage misses: the effort selector idea is not new to Perplexity at all. Its developer API already exposes a reasoning_effort parameter on the sonar-deep-research model, with low, medium and high settings that scale how much searching and reasoning a deep research request performs — and, with it, the cost.
From one model to a whole computer
What the leak describes is that same concept promoted from a single API model to the whole agentic product. That is a much bigger surface: an effort selector on Computer governs how many steps the agent plans, how many models it consults, how many tool calls it makes and how much it verifies before declaring a task done.
Why granular beats binary
A single “fast mode” toggle would have been the easy build. The word granular suggests Perplexity wants task-level tuning instead — the difference between one brake pedal and a full gearbox. For workloads that mix trivial fetches with heavyweight analysis, that distinction decides whether the setting is a gimmick or a genuine cost-control instrument.
The Bigger Pattern: Perplexity's Push to Cut Computer Costs
The effort selector is not an isolated experiment. Through July and August 2026, TestingCatalog has logged a string of Perplexity tests that all aim at the same target: making Computer cheaper to run.
Three cost levers in two months
In July 2026, Perplexity was spotted testing an OpenRouter integration for Computer, which would let users route tasks through cheaper or free third-party models instead of Perplexity credits. In August came signs of dynamic compute splitting between local models and cloud models — explicitly framed as attacking “one of the top blockers”, cost. And alongside the effort selector sighting, references to GPT-5.6 Sol Fast support suggest a faster, lighter orchestrator option is coming too.
What OpenRouter would change
The OpenRouter experiment deserves its own note, because it shows how far Perplexity is willing to go. TestingCatalog found the wiring present but not yet functional: a settings surface where a user could hand Computer tasks to third-party models billed through OpenRouter rather than through Perplexity credits. Routing through cheaper or even free models would, in TestingCatalog’s phrase, loosen “the constraint that has defined who can realistically lean on Computer day to day”. An effort selector and external routing together would give users two independent dials over the same bill.
What that says about the roadmap
Read together, the pattern is unmistakable: Perplexity knows the credit meter is the biggest obstacle between Computer and everyday use. The effort selector is the user-facing end of a cost-reduction programme that runs right through the stack — model routing, local execution and now per-task effort control.
What an Effort Selector Could Mean for Your Business
For businesses experimenting with agentic AI, the arrival of effort controls on a mainstream product is worth planning for, not just observing. It changes how you budget, how you delegate and how you evaluate.
Match the effort to the task
The core discipline is triage. Most business tasks an agent touches — status summaries, inbox drafts, simple lookups — do not need frontier-grade deliberation, while a due-diligence pack or contract review absolutely does. An effort selector lets you encode that triage instead of paying flagship rates for everything, exactly the thinking behind our guide to AI token costs: the cheapest request is the one that spends only what the task deserves.
Budget with real numbers
The published burn figures give you a planning baseline today: if a 500-credit workflow is your heaviest case, a 10,000-credit allowance is twenty such runs a month, and everything lighter stretches it further. A working low effort mode would stretch it further still — Anthropic’s data shows the same quality is often reachable at a quarter of the tokens. Build those scenarios into your AI strategy before the invoice does it for you.
How to pilot effort tiers today
You do not need the feature to adopt the discipline. Sort the agent tasks you run, or plan to run, into three buckets: routine (summaries, drafts, lookups), standard (multi-source research, report assembly) and heavyweight (due diligence, audits, anything a decision rests on). Log the credits each bucket actually consumes over a fortnight of normal use. When the effort selector arrives, that log becomes your mapping — routine tasks go straight to low effort, heavyweight tasks keep full depth, and you will know within days whether the middle tier is safe to step down.
Questions to ask before you rely on it
Three things remain unknown and belong on any evaluation checklist. Does low effort degrade accuracy on your tasks, and by how much? Are credits priced differently per level, or does lower effort simply consume fewer of them? And will the effort selector be available on Pro-tier credits, or gated to Max? Until Perplexity answers officially, pilot with capped auto-refill and measure.
The Stakes: Computer Is Carrying Perplexity's Growth
Why would Perplexity pour this much engineering into cost controls? Because Computer is now the engine of the company’s commercial story, and anything that limits its daily use limits the business.
The growth numbers behind the push
According to Reuters, citing The Information, Nvidia has discussed joining a Perplexity funding round at a valuation above $30 billion — up from $20 billion in September 2025. The same reporting puts Perplexity’s annualised revenue above $750 million as of August 2026, from under $250 million at the start of the year, with the growth attributed largely to Perplexity Computer. The company has also signed a $750 million Microsoft Azure deal and has floated a 2028 IPO ambition.
A large audience waiting on the price
The user base gives the cost work its urgency. Perplexity’s chief executive said the service handled 780 million queries in May 2025, around 30 million a day and growing more than 20% month on month — and that was before the Comet browser went free worldwide and put an agentic front door on every one of those users’ desktops. Only a fraction of that audience will ever pay $200 a month; a low effort mode is how the remainder becomes addressable.
Cheaper to run means bigger to sell
A product whose heaviest users hit a $200 auto-refill in a fortnight has a ceiling. Every lever that lowers the cost per task — routing, local compute, a low effort mode — raises that ceiling, widens the funnel from the free Comet browser into paid Computer usage, and strengthens the revenue line those valuation talks rest on. The effort selector is small UI with large strategy behind it.
Effort Selector FAQ
Is the effort selector available now?
No. As of 24 August 2026 it has only been spotted in testing by TestingCatalog. Perplexity has not announced it, and there is no release date. Features seen at this stage can change substantially or never ship.
What effort levels will Perplexity Computer offer?
Unknown. The leak describes the selector as granular and confirms a low effort mode; the full list of levels is not stated in the post text. Rival products range from three steps to seven, so there is no safe assumption to make.
Will low effort make Perplexity Computer cheaper to run?
That is the stated intent — the leak links low effort directly to credit consumption. Industry precedent is encouraging: effort and thinking controls elsewhere cut token usage sharply, and Anthropic’s published figures show quality holding at a fraction of the tokens. Whether Computer’s implementation delivers similar ratios is untested.
How would this compare with ChatGPT’s thinking-time picker?
ChatGPT’s picker adjusts how long a single model deliberates over one answer. An effort selector on Perplexity Computer would govern something broader: a whole agentic run — planning steps, tool calls, model consultations and verification passes across many minutes of autonomous work. The consumer-facing shape is similar, which is surely what the leak’s “this looks familiar” nodded at, but the cost surface it controls is an order of magnitude larger.
Should businesses wait for the effort selector before trying Computer?
There is no need to wait. Credits can be capped, auto-refill is configurable, and low-stakes pilots reveal your real burn profile now. If the effort selector ships, an organisation that already knows which of its tasks are heavy and which are light will benefit from day one.
References
TestingCatalog on Threads: Perplexity working on a granular effort selector for Perplexity Computer
TestingCatalog: Perplexity tests OpenRouter integration for Computer
TestingCatalog: Perplexity released Personal Computer to all Max subscribers
Builder.io: I tried Perplexity Computer
DataCamp: Perplexity Computer review and tutorial
Reuters via Investing.com: Nvidia discusses Perplexity investment at $30 billion-plus valuation
OpenAI API documentation: Reasoning models and reasoning_effort
Anthropic: Introducing Claude Opus 4.5
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.