Every AI agent that searches the web before answering pays a tax most users never see: a round-trip to a search API, on top of the model tokens it takes to read and reason over the results. For a single chatbot reply, that tax is trivial. For an agent completing one task — comparing flights, researching a company, fact-checking a draft — it can mean dozens of search calls, and it's one of the quiet reasons agentic AI products cost more to run than a single chat reply suggests.
On September 24, 2026, Perplexity shipped a fix aimed squarely at that problem: a new 'Fast Search' option inside its Search API, built on a retrieval-and-ranking engine the company calls Photon — a from-scratch rewrite of Perplexity's search backend in Rust, built, Perplexity says, by a small team of engineers working alongside hundreds of AI agents of their own. Developers turn it on by setting a single parameter — search_type: "fast" — in a Search API call (Source: Perplexity on X).
Quick summary: Perplexity launched Fast Search on September 24, 2026 — a cheaper, lower-latency tier of its Search API built on a new Rust-based engine called Photon. It returns 95% of results in 230 milliseconds or less (160ms at the median) and costs $1 per 1,000 requests versus $5 per 1,000 for standard web search — 80% cheaper per call. Across six internal agent benchmarks covering 3,554 tasks, Fast Search scored 64.3% versus 64.0% for default search — essentially a tie — while cutting total estimated cost (search plus model tokens) by about 68%, according to Perplexity's own testing. The trade-off: in isolation, Fast Search's own relevance score dropped from 2.45 to 2.21 and answer availability fell from 59.6% to 56.7%, which is why Perplexity's documentation recommends standard search for rare, difficult, or ambiguous questions and Fast Search for high-frequency agent workflows.
What Photon actually changed under the hood
Photon replaces the retrieval-and-ranking layer that used to sit behind every Perplexity search call. Perplexity says the migration cut 99th-percentile retrieval-and-ranking latency from roughly 800 milliseconds down to about 65 milliseconds, and that the new system runs on approximately 20% fewer equivalent serving machines than the old one needed for the same traffic (Source: explainx.ai). None of that is visible to a developer calling the API — what they see is a response that, 95% of the time, comes back in 230 milliseconds or less, with a median of 160 milliseconds.
Fast Search vs. standard search: the real numbers
| Metric | Fast Search | Standard Web Search |
|---|---|---|
| Price per 1,000 requests | $1 | $5 |
| Median latency | 160 ms | Not disclosed (higher) |
| 95th-percentile latency | 230 ms | Not disclosed (higher) |
| Internal relevance score | 2.21 | 2.45 |
| Answer availability | 56.7% | 59.6% |
| Aggregate task score (6 benchmarks, 3,554 tasks) | 64.3% | 64.0% |
| Agent API invocation cost | $0.001/call (+ model tokens) | Not published at this granularity |
Read those rows carefully and there's an apparent contradiction: Fast Search scores worse than standard search in isolation — lower relevance, lower answer availability — yet ties it on real agent tasks. Perplexity's explanation is that raw retrieval quality isn't the same as finished-task quality: across 3,554 tasks in six agent benchmarks, the reasoning model compensated enough for Fast Search's slightly noisier results that final task performance came out essentially even, while the combined cost of search plus the model tokens needed to use it fell by about 68% (Source: explainx.ai).
Perplexity's own guidance splits the decision cleanly: use standard web search for rare, difficult, or ambiguous questions where getting the single best source matters, and reach for Fast Search for the high-frequency, repetitive lookups that dominate most agent workflows — the kind of calls where being slightly less precise 95% of the time costs less than being maximally precise on all of them.
Why a search-pricing change matters beyond Perplexity's own customers
This is a narrower release than a new foundation model, but it points at something bigger: as AI agents graduate from answering one question to completing multi-step tasks, the cost of the tools an agent calls — search, code execution, browsing — increasingly rivals the cost of the model itself. A claimed 68% cut in blended search-related cost, at a provider as widely embedded in agent stacks as Perplexity, changes the unit economics for every product built on top of it, not just Perplexity's own consumer app.
The same week, five smaller models also shipped
Fast Search landed in the middle of an unusually busy stretch for smaller, specialized models, most of which were announced within about a day of each other on September 23, 2026: Zhipu AI's Z.ai shipped GLM-5.3 Prime, a 1.5–2x-throughput version of GLM-5.3 with a 1-million-token context window at $2.80 input / $8.80 output per million tokens; Alibaba's Qwen team released Qwen3.8 Max Prime, a higher-throughput SKU of Qwen3.8 Max at roughly double the base model's price ($4 / $12 per million tokens) with identical weights and context; Fireworks AI introduced Ember-1, a reasoning model built on Moonshot's Kimi K3 that cut output tokens per task from about 49,300 to 29,900 in a production A/B test with performance essentially unchanged; Upstage released Solar Mini 4, a smaller, cost-efficient sibling to its Solar Pro 4 model; and Aion Labs shipped Aion-3.5, a roleplaying- and storytelling-focused model built on the GLM family that doubled its predecessor's context window to 262,144 tokens at an unchanged $3 / $6 per million token price (Sources: Fireworks AI — Introducing Ember-1, OrcaRouter — Qwen3.8 Max Prime, OrcaRouter — Aion 3.5, OrcaRouter — Solar Mini 4, OpenRouter — GLM 5.3 Prime).
Who should actually switch
- Teams running high-volume agent pipelines (research assistants, shopping agents, customer-support bots) where most queries are routine lookups, not edge cases.
- Developers who were already treating search-API cost as a margin bottleneck, since the $1-per-1,000 price and the claimed ~68% blended cost cut change that math directly.
- Anyone building on Perplexity's Agent API, where Fast Search now runs at $0.001 per call before model tokens.
- Not a good fit: products where a single wrong or incomplete source is costly — legal research, medical queries, financial due diligence — where the roughly 3-point drop in answer availability is a real risk, not a rounding error.
01What is Perplexity's Fast Search, in one sentence?
A cheaper, faster tier of Perplexity's Search API — $1 per 1,000 requests instead of $5, with 95% of results returned in 230 milliseconds or less — built on a new Rust-based engine called Photon, launched September 24, 2026.
02Is Fast Search less accurate than standard search?
On its own, yes — Perplexity's internal relevance score drops from 2.45 to 2.21 and answer availability falls about 3 percentage points. But across real agent tasks in six benchmarks, aggregate task performance was essentially tied (64.3% vs. 64.0%), because the reasoning model compensated for the slightly noisier retrieval.
03How much cheaper is it really?
The raw per-request price is 80% lower ($1 vs. $5 per 1,000 requests). Perplexity's broader claim — a 68% reduction in combined search-plus-model cost per completed task — comes from its own six-benchmark, 3,554-task internal evaluation, not an independent audit.
04What is Photon?
Photon is Perplexity's new retrieval-and-ranking engine, rewritten in Rust, that now powers Fast Search. Perplexity says it cut 99th-percentile retrieval latency from about 800ms to about 65ms and runs on roughly 20% fewer servers than the system it replaced.
05Should I switch my product to Fast Search?
If your workload is high-volume and tolerant of occasional imprecision — agent tool calls, routine lookups, bulk research — yes. If a single incomplete answer is costly (legal, medical, financial use cases), standard search remains the safer default, which is also Perplexity's own recommendation.
If you're deciding which search-and-reasoning combination actually works best for your own questions — rather than taking any vendor's benchmark numbers on faith — LumiChats gives you a side-by-side way to test that. With more than 40 AI models available in one subscription, you can run the same real query through several models at once and judge speed, cost, and answer quality for yourself instead of guessing from a spec sheet.
