Why it matters
If you use Perplexity for search and sources, there is a new model in the picker, and the company says it does wide research better than anything it has tested. Two posts from Perplexity said two different things about who can use it. This page puts what each source says next to each other, with the date we read it.
Cheaper than Astra and twice the price of Opus per task. Cheap depends on what you compare it to.
Who gets Argon? Two answers
Who gets Argon? Two answers, two minutes apart
On 16 October 2026 at 15:03 UTC Perplexity posted on X: Gemini 4 Argon sets a new record on WANDR, and it is now available in Perplexity for Pro and Max subscribers. The post carried a chart of WANDR scores against cost per task.
Two minutes later Aravind Srinivas, Perplexity's CEO, quoted that post: Gemini 4 Argon is a frontier model, has set a new state of the art on WANDR, and is available now to all Perplexity users on fast search and in Computer, Perplexity's agent mode.
Gemini 4 Argon is Google's model, id gemini-4-argon on Google's API, with a context window of 1,048,576 tokens.
Sources: Perplexity on X (16 October 2026), Aravind Srinivas on X (16 October 2026), and Google, Gemini 4 Argon model page (updated 14 October 2026)
What Perplexity's own pages say
What Perplexity's own pages say
Perplexity's changelog of 19 October puts it this way: Gemini 4 Argon is available in Computer to Pro and Max subscribers, and Max subscribers can also use it in Search by choosing it in the model selector. The help center's table of Search models, as of 27 October, lists Gemini 4 Argon as No for Pro and Yes for Max. On the Free plan you cannot choose an advanced Search model by hand.
So the written pages side with the company post, and are narrower for Search. No page we found says free users get Argon. No page says either way whether Perplexity sends some searches to Argon on its own when you leave the model on Best. We have not tested it on our accounts.
| Where | Free | Pro | Max |
|---|---|---|---|
| Computer (agent) | not stated | Yes | Yes |
| Search, picked by hand | No model choice | No | Yes |
From Perplexity's changelog (19 October 2026) and help center (as of 27 October 2026), both read 28 October 2026. Plans can change. Enterprise: Enterprise Max Yes, Enterprise Pro No in Search.
Sources: Perplexity changelog (19 October 2026) and Perplexity help center, models by plan (as of 27 October 2026)
The record
The record: Perplexity's test, Perplexity's runs
WANDR stands for Wide ANd Deep Research. Perplexity published it on 14 July 2026: 500 data-collection jobs of the kind people hand to a researcher, such as mapping competitors, due diligence, literature reviews, market analysis, product comparisons and sourcing candidates. A job asks for a long list, a median of 50 entries with four facts each, and every fact needs evidence. The test and its grader are public on GitHub under the Apache-2.0 licence.
The chart above is Perplexity's, from its post of 16 October. Perplexity built the test, ran the models and drew the chart. It does not say which score the axis shows, which effort setting each model ran at, or which prices the cost per task uses. The July article reported soft and hard F1 scores on a different scale, so the two cannot be compared. On 13 October the repository switched its grader to GPT-6 Luna and regenerated the tasks; the chart does not say which grader it used.
We searched on 28 October 2026 and found no score for Argon on WANDR from anyone other than Perplexity. Anyone can rerun the test from the repository.
On Perplexity's chart a WANDR task with Argon costs USD 7.88. That is less than GPT-6 Astra (USD 13.34) and Claude Fable 5.1 (USD 12.76), and about twice Claude Opus 5.5 (USD 3.69) and GPT-6.1 Sol (USD 3.70). The ratio is our arithmetic on Perplexity's numbers.
| Model | Score on Perplexity's chart | Cost per task (USD) |
|---|---|---|
| Gemini 4 Argon | 0.802 | 7.88 |
| GPT-6 Astra | 0.735 | 13.34 |
| GPT-6.1 Sol | 0.654 | 3.70 |
| Claude Opus 5.5 | 0.647 | 3.69 |
| Claude Fable 5.1 | 0.601 | 12.76 |
| GPT-6 Sol | 0.566 | 2.23 |
| Claude Sonnet 5.5 | 0.550 | 2.73 |
| Grok 4.7 | 0.528 | 9.30 |
| DeepSeek V4 Pro 0813 | 0.359 | 0.75 |
Perplexity's runs on Perplexity's test, as published on 16 October 2026. Not measured by us or by a third party.
Sources: Perplexity on X, the WANDR chart (16 October 2026), Perplexity, WANDR benchmark article (14 July 2026), and perplexityai/wandr on GitHub (read 28 October 2026)
Read the sources
Read the sources
Every claim above comes from these pages, read on 28 October 2026. Plans and prices can change.
- Read the postPerplexity on X
The post of 16 October: the WANDR record and "Pro and Max subscribers".
- Read the postAravind Srinivas on X
Two minutes later, quoting it: "all Perplexity users", in fast search and in Computer.
- Read the changelogPerplexity's changelog, 19 October
Argon in Computer for Pro and Max, in Search for Max by picking it.
- Open the tablePerplexity's help center: models by plan
The table of Search models: Argon No for Pro, Yes for Max (as of 27 October).
- Read the articleWANDR, the test
Perplexity's article of 14 July: 500 wide research jobs, how they are scored.
- Open the repositoryWANDR on GitHub
The test and its grader, public under Apache-2.0. Anyone can rerun it.
- See the pricesGoogle's price for Argon on its API
Launch price through 30 November 2026; input and output double on 1 December.
Our screenshot of Google's Gemini API pricing page, 28 October 2026 - Read the pageGoogle's model page for Argon
The model id gemini-4-argon and its context window of 1,048,576 tokens.
Yes, that's my AI avatar in the video. Same opinions, better lighting.