Some links in this article are affiliate links. Automated Wealth Lab may earn a commission at no extra cost to you. See our Affiliate Disclosure.
- The three billing shapes
- Perplexity Sonar Deep Research: five meters, one report
- OpenAI: premium tokens, simpler math
- Google Gemini Deep Research: the only published per-task estimate
- The full comparison: cost per finished report
- The self-host alternative: run the researcher yourself
- FAQ
- Methodology
- AWL verdict
- Sources
Disclosure: This article contains affiliate links to DigitalOcean and Hostinger. If you sign up through them, Automated Wealth Lab earns a commission at no extra cost to you. The API providers compared at the top of this article pay us nothing — the hosting links appear only in the self-host alternative section, where a VPS genuinely is the product being discussed.
A “deep research” API call is the most misleading line item in AI pricing. There is no per-report price. There is a pile of meters — input tokens, output tokens, reasoning tokens, citation tokens, individual search queries — and a report is whatever the pile adds up to after the agent finishes wandering the web. Two identical prompts can cost 40 cents or 4 dollars depending on how many searches the agent decides to run.
So this article does the only honest thing: it takes each provider’s published meters, applies stated token assumptions, and computes the cost per finished report for Perplexity Sonar Deep Research, OpenAI’s deep research models, and Google’s Gemini Deep Research agents. Every figure below is a modeled estimate built from list prices checked in October 2026 — labeled as such, with the assumptions printed next to it so you can re-run the math on your own workload.
The three billing shapes
| Provider / product | How it’s billed | Published rates (Oct 2026) |
|---|---|---|
| Perplexity Sonar Deep Research | Tokens + reasoning + citations + per-search fee | $2/1M input, $8/1M output, $3/1M reasoning tokens, $2/1M citation tokens, $5 per 1,000 searches |
| OpenAI o3-deep-research (last published rates) | Tokens only | $10/1M input, $40/1M output |
| OpenAI o4-mini-deep-research | Tokens only | $2/1M input, $8/1M output |
| Google Deep Research (standard agent) | Tokens at Gemini 3.1 Pro rates + search grounding | $2/1M input, $12/1M output; search grounding $14 per 1,000 queries |
| Google Deep Research Max | Same meters, bigger appetite | Same rates; roughly double the searches and input volume |
Note what’s not in the table: a per-report SKU. Google is the only provider that publishes its own per-task estimate — roughly $1–$3 per standard task and $3–$7 per Max task — and even Google frames it as an estimate, because the meters decide.
Perplexity Sonar Deep Research: five meters, one report
Sonar Deep Research has the most complex bill in the category. On top of input ($2/1M) and output ($8/1M) tokens, it charges separately for reasoning tokens ($3/1M), citation tokens ($2/1M) and each search query ($5 per 1,000). The reasoning and citation meters are the ones that bite: a deep research run generates far more internal reasoning text than final report text.
Perplexity’s own documented worked example, as reported by pricing trackers: one complex query that ran 21 searches, generated 193,947 reasoning tokens, 19,028 citation tokens and 11,395 output tokens cost approximately $0.82. Tracker estimates for more typical queries put a single report around $0.41 and up. Scale that honestly: a product generating 1,000 reports a month at $0.41–$0.82 is a $410–$820 monthly API bill for the research layer alone — before your app’s own models, hosting or margins.
Two operational caveats. First, the search fee means your prompt design is a cost control: vague prompts fan out into more searches. Second, the surface is moving — coverage in late 2026 reports Perplexity ended support for the older Sonar chat-completions endpoint (September 2026) and now routes deep research through its Agent API, where Sonar Deep Research maps to a high preset. Budget on the meters, not the endpoint name, because the endpoint names are churning.
OpenAI: premium tokens, simpler math
OpenAI’s dedicated deep research models bill on tokens alone — no per-search fee — but at premium token rates. The last published rate for o3-deep-research was $10/1M input and $40/1M output; its smaller sibling o4-mini-deep-research listed at $2/1M input and $8/1M output. (One pricing tracker reports o3-deep-research was retired in July 2026 with OpenAI steering users to newer flagship models — treat the model name as perishable and the rate structure as the durable information.)
Modeled cost per report, assuming a run that consumes 250,000 input tokens (the agent re-reads a lot of web) and produces 30,000 output tokens including its reasoning trace:
| Model | Input cost | Output cost | Modeled cost per report |
|---|---|---|---|
| o3-deep-research ($10/$40) | $2.50 | $1.20 | ~$3.70 |
| o4-mini-deep-research ($2/$8) | $0.50 | $0.24 | ~$0.74 |
The assumption is doing real work in that table — double the input volume and both figures roughly double — which is precisely the point: with token-only billing, report length and source volume are your price. The mini model at ~$0.74 per modeled report is the budget disruptor here, sitting in Sonar Deep Research’s price neighborhood with a far simpler bill to forecast.
Google Gemini Deep Research: the only published per-task estimate
Google’s Deep Research agents (standard and Max) run on Gemini 3.1 Pro, billed at that model’s standard rates — $2/1M input, $12/1M output at the standard tier — with Google Search grounding billed separately at $14 per 1,000 queries after a monthly free allowance of grounding prompts on paid tiers. Google publishes the workload shape: a standard task runs about 80 searches and processes roughly 250K input tokens (50–70% of them typically cached, which cuts the effective input cost) and ~60K output tokens; a Max task runs about 160 searches and ~900K input tokens.
Search grounding alone adds $1.12 per standard task (80 × $14/1,000) and $2.24 per Max task — a floor under every report before tokens are counted. Google’s own estimate for the all-in figure: $1–$3 per standard task, $3–$7 per Max task. Our read of the published meters is consistent with that range. Access note: these agents live behind Google’s Interactions API in preview, are asynchronous, and require a paid tier — you architect for minutes-long jobs, not chat latency.
The full comparison: cost per finished report
| Provider | Modeled / documented cost per report | Forecast difficulty | Watch out for |
|---|---|---|---|
| Perplexity Sonar Deep Research | ~$0.41–$0.82 (documented examples) | Hard — 5 meters | Reasoning + citation tokens dominate; search fan-out |
| OpenAI o4-mini-deep-research | ~$0.74 (modeled, 250K in / 30K out) | Medium — tokens only | Input volume grows with sources read |
| OpenAI o3-deep-research | ~$3.70 (modeled, same assumptions) | Medium — tokens only | Premium output rate; model lineup churn |
| Google Deep Research (standard) | $1–$3 (Google’s estimate) | Medium | $14/1K grounding fee is a hard floor |
| Google Deep Research Max | $3–$7 (Google’s estimate) | Medium | ~900K input tokens per task |
Set against the consumer alternative, the framing flips: a $20/month chat subscription with deep-research included looks cheap until you need the output inside a product — at which point the API bill above is simply the cost of goods. If you’re selling reports, your gross margin lives in the gap between these numbers and your price, which is why we modeled this alongside the broader pattern in what AI agents really cost per task. And whatever API you pick, the output still needs verification before it ships to a client — our citation trust test explains why, with numbers.
The self-host alternative: run the researcher yourself
There is a fourth option that isn’t an API at all: an open-source research agent (GPT Researcher is the reference implementation) running on your own server, calling a standard LLM API. We compared that path head-to-head in GPT Researcher vs Perplexity, and wired it into automations in the automated research pipeline guide.
The economics: a modest VPS — a 4 GB DigitalOcean Droplet at $24/month or a Hostinger KVM 2 VPS (8 GB RAM, listed at $8.99/month on its promotional term, renewing at $14.99) — replaces the per-report agent fee with a flat server cost plus whatever base-model tokens your researcher consumes. At 100+ reports a month, that trade usually wins on cost. What it loses on: your time. You own the crawling failures, the rate limits, the updates and the 3 a.m. outage — the managed APIs’ per-report premium is, in large part, a fee for not owning those.
FAQ
How much does one deep research API report actually cost in 2026?
Documented and modeled figures cluster between roughly $0.40 and $7 per report. Perplexity Sonar Deep Research examples land around $0.41–$0.82; Google estimates $1–$3 for its standard agent and $3–$7 for Max; OpenAI’s models model out between ~$0.74 (o4-mini-deep-research) and ~$3.70 (o3-deep-research) on a 250K-input/30K-output assumption. Your prompt’s breadth is the biggest variable.
Why can’t providers just charge a flat per-report price?
Because the work per report varies enormously — one query needs 8 searches, another needs 80 — and every provider’s cost structure is meters underneath. Google is the only one publishing a per-task estimate, and even that is a range, not a price.
What is the biggest hidden cost in Sonar Deep Research?
The meters beyond input/output: reasoning tokens at $3/1M and citation tokens at $2/1M, plus $5 per 1,000 searches. In Perplexity’s documented worked example, the reasoning tokens alone (~194K of them) cost more than the visible output. Budgets built only on input/output rates undercount badly.
Is it cheaper to self-host an open-source research agent?
At volume, usually yes on cash cost: a $24/month VPS plus base-model API tokens replaces per-report agent fees. At low volume, or when your time is worth anything, the managed APIs are cheaper in total. The break-even is mostly a question of how many reports you run and how much operational babysitting you’ll tolerate.
Which deep research API is easiest to budget for?
OpenAI’s token-only models and Google’s published per-task estimates are the most forecastable. Perplexity’s five-meter structure produces the lowest documented per-report costs on typical queries but the widest variance — cap it with prompt discipline and search limits where the API allows.
Methodology
What we did: Took each provider’s published rates (Perplexity’s five-meter Sonar Deep Research pricing; OpenAI’s last published deep-research model rates; Google’s Gemini 3.1 Pro rates plus its $14/1,000 grounding fee and published per-task estimates) and computed per-report figures using stated assumptions — 250K input / 30K output tokens for the OpenAI models; Perplexity’s own documented worked example for Sonar. All prices cross-checked across vendor-documented figures and independent pricing trackers, October 2026.
What we did NOT do: We did not call any of these APIs, generate reports, or measure report quality. The per-report numbers are arithmetic on published rates with the assumptions printed beside them — not invoices. One tracker reports o3-deep-research was retired in July 2026; we present its last published rates as the reference point for OpenAI’s premium tier and flag the churn rather than pretending the lineup is stable.
Conflict disclosure: DigitalOcean and Hostinger are affiliate partners. They appear only in the self-host section; the three API providers compared pay us nothing, and the self-host option is presented with its operational costs stated, not as the default recommendation.
AWL verdict
If you’re building a research product, Perplexity Sonar Deep Research offers the lowest documented per-report cost but demands the most billing vigilance; Google’s agents are the easiest to budget because Google does the estimating for you; OpenAI’s mini deep-research model is the simple-bill budget pick. Above ~100 reports a month, price the self-hosted route seriously — a flat VPS bill plus base-model tokens beats five meters, if you’re willing to be the on-call engineer your API provider used to be.
Sources
- Perplexity Sonar Deep Research meters ($2/$8 tokens, $3/1M reasoning, $2/1M citations, $5/1K searches) and the ~$0.82 worked example: https://suprmind.ai/hub/perplexity/pricing/ (verified October 2026)
- Perplexity API structure and Sonar request fees: https://www.finout.io/blog/perplexity-pricing-in-2026 (verified October 2026)
- Sonar Chat Completions end-of-support and Agent API migration (September 2026): https://www.profitec-ai.com/blog/openai-deep-research-api-connectors (verified October 2026)
- OpenAI deep research model rates — o3-deep-research $10/$40, o4-mini-deep-research $2/$8 per 1M tokens: https://www.scriptbyai.com/gpt-gemini-claude-pricing/ (verified October 2026)
- o3-deep-research last published rates and reported July 2026 retirement: https://www.economize.cloud/resources/open-ai/pricing/o3-deep-research/ (verified October 2026)
- Google Deep Research agent task estimates ($1–$3 standard, $3–$7 Max), workload shape (80/160 searches, ~250K/~900K input tokens) and Gemini 3.1 Pro rates: https://tokencost.app/blog/gemini-deep-research-agent-cost (verified October 2026)
- Google Search grounding fee ($14 per 1,000 queries; free monthly allowance on paid tiers): https://nerova.ai/costs-roi/deep-research-max-pricing-explained-2026 (verified October 2026)
- DigitalOcean Droplet pricing (Basic 4 GB / 2 vCPU, $24/month): https://www.digitalocean.com/pricing/droplets (verified October 2026)
- Hostinger VPS pricing (KVM 2: $8.99/mo promotional, renews $14.99/mo): https://gizmodo.com/best-web-hosting/hostinger (verified October 2026)
