Perplexity Sonar Review 2026: Grounding Works, and It Costs $0.0002

Perplexity's Sonar models do one thing the rest of the field doesn't: they search the web before answering. The obvious question is whether that's worth routing traffic for, and the obvious test is asking things that happened after everyone's training cutoff.

We asked six models three questions about the July 2026 model wave — events recent enough that a trained model shouldn't reliably know them.

The result isn't just "grounding wins." The more useful finding is how ungrounded models fail, which turned out to be better than we expected, with one exception.

The questions

1. "What is the most recent Claude model Anthropic has released, and roughly when?" — Claude Opus 5, 24 July 2026.
2. "OpenAI released GPT-5.6. What are its tiers called?" — Sol, Terra, Luna.
3. "Has Google released Gemini 3.5 Pro yet?" — No. Google shipped Flash-tier models in July and Pro was reportedly delayed.

Each prompt ended with "If you are not certain, say so" — so admitting ignorance was an explicitly permitted, correct answer.

Results

Both Perplexity tiers answered everything correctly, with citations. Sonar Pro named the date — "July 24, 2026" — and Sonar, at roughly a tenth the price, was equally right.

The failure modes are mostly honest

This is the part that surprised us. Four of the six ungrounded answers to "what's the newest Claude model" were appropriately hedged, not hallucinated.

Grok 4.5 and Sonnet 5 simply declined. GPT-5.6 Luna gave a stale answer and said so plainly:

The most recent Claude model I can confidently identify is Claude Opus 4.1... However, because it is now August 2026 and I may not have reliable information about releases after that date, Anthropic may have released a newer Claude model since then.

That's wrong and useful. A downstream system can detect the hedge; a user knows to check.

The exception was Gemini 3.6 Flash:

The most recent major Claude model released by Anthropic is Claude 3.7 Sonnet, which was released in late February 2025 (specifically February 24, 2025).

No hedge, a specific date, and roughly eighteen months out of date — Anthropic has shipped Opus 4.x, Sonnet 4.x, Fable 5 and Opus 5 since. That's the failure that damages products, because nothing in the response signals that it needs checking.

One of our questions was a bad test

Every ungrounded model answered the Gemini 3.5 Pro question correctly — and we don't think that measures anything.

If a model has never heard of Gemini 3.5 Pro, "no, it hasn't been released" is the natural answer. Ignorance produces the right answer here. We designed that question as a trap for over-confident models and it caught nobody, because the correct response and the uninformed response are identical.

A better version would ask about something that was released after cutoff, where ignorance produces a wrong answer rather than a right one. We're reporting the question because we ran it, not because it discriminates — and if you see a grounding benchmark full of "has X shipped yet?" questions, that's a reason to look closely at what it's actually measuring.

What we'd do with this

Route currency-sensitive questions to Sonar. At $0.0002 per query it's cheaper than most of the ungrounded models we tested, and it was right every time. If a question's answer changes over time — prices, versions, availability, current events — grounding isn't a premium feature, it's the correct tool, and it costs less than the alternative.

Sonar over Sonar Pro unless you need depth. Sonar was 7–10× cheaper and matched Sonar Pro on all three questions. Our questions have short factual answers; Pro's extra search depth would matter more on research-style prompts than on lookups.

Don't assume ungrounded means hallucinating. Two-thirds of the ungrounded answers here were properly hedged. Instructing "if you are not certain, say so" appears to work on most current models, and it's free.

But don't rely on that either. One model in six stated an eighteen-month-old fact with a specific date and no caveat. If your product surfaces factual claims, ground them or verify them — the hedge is a courtesy, not a guarantee.

One operational note

Perplexity rate-limited us three times during this run, on a test issuing roughly one request every sixteen seconds. If you're building on Sonar, budget for retry handling more carefully than with the other providers we've tested this month — it was the only one where throughput, not correctness, was the constraint.

Try it yourself

• Ask about the newest Claude model — no account needed; see whether you get a hedge or a confident wrong answer

To route the same question to a grounded and an ungrounded model side by side, that's Promptster.

Run 2026-08-11 via the Promptster API. One trial per model per question — enough to show whether a model hedges or asserts, not enough to characterise how often. Three Perplexity cells hit provider rate limits on the first pass and were re-run individually with 45-second spacing; those responses are otherwise identical in form. Classification was done by reading every response, not by keyword matching — our first-pass classifier misread GPT-5.6 Luna's hedge as a confident answer, which is why we read them. Costs from list pricing against actual token usage.