The Model Routing Playbook: What We Learned Running an AEO Agency on Perplexity Computer
How WebbROI routes AI models for cost control across AEO audits, migrations, and content ops. Relative cost tiers, real user demographics, and a practical playbook for technical operators.
TL;DR
- Perplexity Computer doesn’t publish a per-model price list. Credits burn by task complexity, tools, subagents, and media — model choice is a relative lever, not a menu.
- We route models by job: premium reasoning for architecture and decisions, cheap operators for volume execution.
- We built our AEO audit and migration workflows around this split — and cut per-audit compute cost by roughly half without dropping quality.
- Skill you can steal: architect once on Claude Opus or Fable, execute on GLM 5.2 or Grok, verify in the credits ledger after every run.
Perplexity Computer is one of the AI tools we use at WebbROI. It’s not the only one — Claude Code, Claude Desktop, Cursor, and direct API calls each show up in different parts of our stack. We reach for Computer when the job needs fast setup, tight iteration loops, or built-in connectors to GitHub, Google Drive, Postiz, GSC, and the other services we already have wired up. Every AEO audit and Squarespace-to-Astro migration we run touches Computer at some stage.
Credits burn fast when you’re not paying attention. Most operators pick a model once and forget about it. That works until the bill catches up.
This post is what we’ve learned about controlling credit spend on Computer without dropping quality. It applies to anyone running multi-step technical work there — agencies, consultants, in-house teams, or solo operators.
What Actually Burns Credits
Credits power Computer’s heavier work — subagents, browser automation, wide research, deployments, media generation. Normal Ask/Search doesn’t burn them. The published rate is 100 credits ≈ $1.
Task ranges vary wildly. From Perplexity’s own guidance:
| Task type | Typical credit range |
|---|---|
| Light (single search, quick task) | ~100–350 |
| Complex (multi-step research, one deliverable) | ~350–950 |
| Heavy (deep research, coding, asset creation) | ~875–2,275 |
| Mega (multi-agent orchestration, video, large builds) | ~2,400–9,800 |
What actually drives the number:
- Steps. Each tool call, subagent, browser task, file read — additive.
- Model choice. Premium models cost more per call, but often finish in fewer calls.
- Media. Image and video generation are the biggest single spenders.
- Retries. A cheap model that fails and retries can cost more than a premium model that gets it right first pass.
Model choice is a lever, not the thing itself. You can burn 500 credits on Claude Opus doing one clean task, or burn 500 credits on Sonnet doing that same task twice because the first attempt was wrong.
Ground truth: after every run, check ⋮ → credits used in the thread, or account usage for the rolling picture.
The Model Landscape (as of mid-2026)
Computer orchestrates multiple providers behind the scenes and lets you nudge model selection when it matters. Availability changes by plan and account, so treat any specific model as best-effort at any moment.
Here’s how we categorize them for actual work. Cost tiers are relative — $ means cheaper per equivalent task, $$$$ means most expensive. Do not read these as dollar amounts.
Text and orchestration models
| Model | Relative cost | Best used for |
|---|---|---|
| Claude Sonnet 5.0 | $$ | Default workhorse — research, ops, drafts, most WebbROI work |
| Gemini 3.1 Pro | $$ | Large-context research, multimodal, budget-oriented deep dives |
| GPT-5.4 | $$ | Math, logic, structured reasoning, general fallback |
| Grok | $–$$ | Fast light tasks, real-time web, casual iteration |
| GLM 5.2 | $–$$ | High-volume operator loops, architect-then-execute workflows |
| Claude Opus 4.8 | $$$ | Hard planning, decision documents, work you don’t want to redo |
| GPT-5.5 | $$$ | Frontier agentic coding, terminal automation |
| Claude Fable 5 | $$$$ | Top-of-catalog quality when cost is secondary |
Image models
| Model | Role |
|---|---|
| Seedream 5 | Budget volume runs |
| Nano Banana 2 | Fast drafts, iterative concept work |
| GPT Image 2 | Default quality, best for images with visible text |
| GPT Image 1.5 | Transparent backgrounds only |
| Nano Banana Pro | Premium stills, client-facing hero images |
Video models
| Model | Cost tier | Role |
|---|---|---|
| Sora 2 | $ | Cheap draft iteration |
| Seedance 2.0 | $ | Budget variants with native audio |
| Veo 3.1 Fast | $$ | Most Veo quality at reduced cost |
| Sora 2 Pro | $$$ | Cinematic realism, physics-heavy scenes |
| Veo 3.1 | $$$ | Dialogue, creative control, premium video |
Who Uses What (and Why It Matters)
Understanding the user base of each model matters because it tells you what kind of work each one has been optimized against. Data below is drawn from the Stack Overflow 2026 Developer Survey, ICONIQ’s State of AI 2026 report, and vendor documentation.
- Claude Sonnet 5.0 has the highest admiration rating in the developer market — 67.5%. Professional developers use it at ~45%; learners only 30%. It’s the model senior engineers migrate toward as they gain experience. (Stack Overflow data via PPC Land)
- Claude Opus 4.8 is positioned for production coding with minimal oversight. Anthropic reports it’s roughly 4× less likely than its predecessor to pass code flaws without flagging them. Popular with senior engineers and enterprise teams. (Anthropic)
- Claude Fable 5 is Anthropic’s newest — a “Mythos-class” model built for long-horizon autonomous work with a 1M token context. Aimed at the ~31% of developers now deploying AI agents. (Global Tech Council)
- GPT-5.4 and 5.5 remain the volume leaders. OpenAI’s models hold ~81% developer usage. ChatGPT drives 55% of enterprise AI conversations. GPT-5.5 in particular scores 82.7% on Terminal-Bench 2.0, positioning it as the agentic-coding leader in OpenAI’s stack. (DevOps.com, OpenAI)
- Gemini 3.1 Pro leads on multimodal understanding — text, images, video, audio, code. Strong for creative and prototyping work. Gemini Flash sees ~35% developer usage; Pro sees less but is deeper. (Google DeepMind)
- Grok has ~117M monthly active users, heavily driven by X integration. Audience skews male (67.5%) and young (25–34 largest segment). Real-time web access is its differentiator. Mostly consumer/social — thin in professional developer surveys. (Daily AI Mail)
- GLM 5.2 has minimal Western enterprise adoption. Chinese models overall have closed most of the performance gap with US models per the Stanford AI Index, but adoption in Western technical stacks remains under 1%. Best used as a cost-efficient operator in a routed workflow. (Stanford HAI)
The takeaway: there’s a “ubiquitous default” (OpenAI) and an “expert preference” (Anthropic). Grok is a consumer product with a technical use case. GLM is a cost lever. Gemini is the multimodal specialist. Choose accordingly.
How We Route Models on Computer
Perplexity Computer is where we run the workflows that benefit from fast setup and connectors we already have wired up — GitHub, Google Drive, GSC, Postiz, our Google Business Profile. For deep repo work we tend to reach for Claude Code; for one-off API calls we go direct. Computer’s strength is the middle: multi-step, multi-tool jobs where the connector coverage saves us more time than a custom script would.
Here’s how we route within Computer specifically:
Architect on premium. Execute on cheap.
For AEO audits: Claude Opus 4.8 designs the audit framework, plans the deliverable structure, and produces the executive summary. GLM 5.2 or Grok runs the repetitive execution — pulling GSC data, checking robots.txt across a batch of client sites, generating per-URL findings.
Result: the premium reasoning happens once, at the start. The volume work happens on a cheaper model. Total cost per audit dropped by roughly half versus running everything on the same premium model.
Blog research: Sonnet default, escalate on quality signal
Every blog post starts on Sonnet 5.0. It’s fast, cheap enough, and produces first-draft quality that’s usually good enough to publish with light edits. When Sonnet’s output feels thin — vague framing, missed nuance, weak structure — we re-run on Opus. About 1 in 4 posts.
Images: Nano Banana 2 for drafts, GPT Image 2 for final
Every social image concept starts on Nano Banana 2. Iterate fast, throw away the misses. When we’ve locked the composition, we regenerate the final on GPT Image 2 — same prompt, higher fidelity, better text rendering. Total cost is lower than starting on GPT Image 2 and iterating there.
Wide browsing: Grok or GLM for volume, never premium
For batch tasks — checking 40 client sites for schema markup, scraping competitor pricing pages — we route to whichever budget model is available. Premium models on batch work is money on fire.
The exception: client-facing deliverables
Anything a client will read — proposal, audit report, migration plan — gets Opus or Fable at the final polish step. The economics work: one saved back-and-forth with a client is worth 10 audit runs.
The Playbook
Steal this if it’s useful:
- Ask/Search for single facts. Computer only for multi-step work. This alone saves the most.
- Force cost mode in the prompt. “Use GLM 5.2 or Grok for execution. Optimize for cost.” The orchestrator respects explicit routing.
- Architect on premium once, execute on cheap. For any workflow you’ll run more than three times, split the cognitive design work from the volume work.
- Skills + threads beat re-explaining. Load a skill or continue a thread instead of re-briefing every session. The context is already there.
- Prototype small before wide runs. Before firing
wide_browseon 100 URLs, run 3 and check the output. - Check
⋮ credits usedafter every run. It’s the only ground truth. Everything else is estimation. - When quality fails, escalate — don’t restart. If Sonnet’s output is weak, hand the same thread to Opus. Don’t burn the setup context.
FAQ
Does Perplexity publish exact credit costs per model?
No. Computer bills by task complexity — steps, tools, subagents, media, retries — not by a per-model tariff. Model choice affects cost as a relative lever. Check the credits documentation for the current rate and task-range guidance.
Can I pick which model runs my task?
Sometimes. The UI model picker exposes a subset of available models depending on plan and account status. Computer’s internal routing may still use other models for subtasks. Availability changes — treat it as best-effort.
Which model should I default to for most work?
Claude Sonnet 5.0 is the best value default for research, ops, and drafts. Escalate to Opus for hard planning or client-facing deliverables. Drop to Grok or GLM for volume execution.
How much can routing actually save?
In our AEO audit workflow, splitting architect (Opus) from execute (GLM) cut per-audit compute cost by roughly 50% versus running everything on premium — no measurable quality drop in the deliverable. Your mileage will vary by workflow.
What’s the biggest single cost sink in Computer?
Media generation. A single video run can burn more credits than 20 text-only sessions. Prototype the concept on Nano Banana 2 or Sora 2 before committing to Veo 3.1 or Sora 2 Pro.
Model routing is one of the highest-leverage skills for any technical operator on Perplexity Computer. Pick by job, verify by ledger, and never pay premium for volume work.
WebbROI is an AEO + web development agency. We build fast, answer-ready sites on Astro + Cloudflare Pages and run AEO audits that show you exactly where your business is cited in ChatGPT, Perplexity, Gemini, and Google AI Overviews. Get a fixed-price AEO audit.