Octen Search Menempati Peringkat Ketiga di Artificial Analysis Search Index: Tercepat per Tugas, Termurah di Antara Para Pemimpin

Octen Search scored 77 on its first appearance in the Artificial Analysis Search Index, good for third place on the board — and it got there in 17.2 seconds per task, quicker than anything else Artificial Analysis tested. The product itself is not new: Octen Search has been commercially available since April 22, 2026, when its parent company APITECH AI announced a $10 million seed round led by Square Peg. What changed is that as of Artificial Analysis's September 8 data there is finally a third-party number attached to it, and the number is better than the earliest independent hands-on testing of Octen Search would have predicted.

What Artificial Analysis is actually measuring

The Search Index went live on August 18, 2026, and its design is unusually clean for a leaderboard. Every provider runs inside the same agent harness — Stirrup, Artificial Analysis's own open-source agent — driven by the same model, GPT-5.6 Luna at medium reasoning, and scored by the same grader. The agent gets two tools, web_search and web_fetch, up to 25 turns per task and at most ten results per search, with known contamination sources filtered out. A run that burns all 25 turns without calling finish scores zero, so leaving the agent unable to close out a task is punished as hard as retrieving nothing.

The only variable is the search provider. That is the property most search comparisons lack, and it is what makes the placement of Octen Search readable as a statement about retrieval rather than about which vendor bundled the strongest answer model.

Scores are the equal-weighted mean of three evaluations on a 0–100 scale. DeepSearchQA F1 runs over a 900-task split of broad research questions and grades the answer as a list of items. BrowseComp accuracy uses a 200-sample hard subset of multi-hop facts where the answer is a single verifiable value. AA-Omniscience accuracy uses 600 held-out general-knowledge questions and exists mainly to isolate how much search adds beyond what the model already knew. Because the index is a straight mean, a single soft component drags the headline number down, and the component profile is often more useful than the ranking. Running the same agent with no search tools at all scores 33, so nearly everything a provider earns is lift over the model's own parameters.

Artificial Analysis says it tested 20 search products across 10 providers; the default charts display 12 of them.

Where Octen Search landed

The top of the September 8 table, with per-task costs calculated by adding the search and model-token columns Artificial Analysis publishes per 1,000 tasks — the leaderboard itself does not print a single combined figure:

• Perplexity Search (medium) — Index 80, 28.8s per task, about $0.091 per task

• Perplexity Search (high) — Index 79, 29.7s, about $0.091

• Octen Search (highlights) — Index 77, 17.2s, about $0.058

• Perplexity Search (low) — Index 77, 37.4s, about $0.105

• Parallel Search (advanced) — Index 75, 42.9s, about $0.084

• Brave Search (LLM context) — Index 75, 24.4s, about $0.130

• You.com Search (highlights) — Index 74, 20.7s, about $0.117

• Exa Search (auto) — Index 74, 33.3s, about $0.127

The component scores are where the placement gets more interesting than the ranking. Octen Search scores 80 on DeepSearchQA, 86 on BrowseComp and 66 on AA-Omniscience. Parallel Search (advanced), the provider that led the index at launch, scores 81, 77 and 67. So Octen Search beats Parallel Search by nine points on the multi-hop fact-finding benchmark and matches it on broad research, while trailing the Perplexity Search settings on the general-knowledge lift — 66 against 72 for Perplexity Search (medium). In plain terms, Octen Search is very good at hunting down things that are genuinely hard to find, and offers less extra on questions the model already half-knows.

One small wrinkle worth noting for anyone quoting this: Artificial Analysis's own announcement of the result put Octen Search at 16.9 seconds per task, while the live table shows 17.2 on September 8 data. That is re-run drift, not a contradiction, but it is the reason to cite the table rather than the social post.

The speed is the headline; the cost split is the story

17.2 seconds per task makes Octen Search the fastest entry on the index. Perplexity Search (medium) needs 28.8 seconds, Parallel Search (advanced) 42.9, and Firecrawl Search 63.2. Artificial Analysis also reports Octen Search as fastest per individual query, averaging 0.21 seconds per search — the fastest call time measured, independent of how many calls a task takes.

The number worth staring at, though, is the cost breakdown. Per 1,000 tasks, Artificial Analysis puts Octen Search at $9.07 of search spend and $49.15 of model-token spend. The tokens cost roughly 5.4 times the searches. That is not a quirk of Octen Search — it is the structural point this index keeps making. A search provider does not only bill you for retrieval; it decides how many turns the agent takes, and every extra turn is model tokens at the model's rate.

Artificial Analysis made the same argument with Parallel at launch: advanced mode cost more per search than basic ($0.048 against $0.045) but less per task ($0.084 against $0.11), because better results cut token use from roughly 339,000 to 169,000 per task. Cheaper calls and cheaper tasks are different things, and only one of them shows up on your invoice.

Read against the rest of the board, Octen Search's roughly $0.058 per task is the lowest of anything scoring 75 or above. TinyFish Search is cheaper at about $0.035, but it scores 71 and takes 60.4 seconds to finish a task.

What the result settles, and what it does not

Octen's own numbers have always been aggressive, and they remain vendor-reported. The company claims 62 ms P50 and 68 ms P90 latency on the SealQA Hard benchmark, throughput above one million queries per second per account, new content indexed within five minutes, and a number three global placement on DeepResearch Bench at an overall 55.58. None of those has been independently reproduced, and several are not directly comparable to anything else published.

Two independent tests now exist, and they point in different directions. Artificial Analysis, holding the model and harness fixed, puts Octen Search third on quality and first on speed. A separate independent benchmark published in August 2026 measured Octen Search at 359 ms P50 and 941 ms P95 — several times the vendor's claimed latency — and an nDCG@10 of 0.495, just under the 0.5 line and behind Parallel, Brave, Kagi, Perplexity, Context, Tavily and Exa on that metric. Octen's team told that reviewer its own benchmark primarily tests latency rather than result quality, which is a fair clarification and also an admission that result quality was not the thing being claimed.

Both results can be correct, because they ask different questions. That earlier test graded whether the top ten results looked relevant by metadata heuristics on its own query set; Artificial Analysis grades whether an agent, with a real model driving it, finishes a task correctly. What the index placement firmly establishes is narrower than "best search API" and more useful than a vendor claim: on agentic task completion, with the model held constant, Octen Search is competitive with the strongest providers and faster than all of them.

What it costs to actually run Octen Search

Octen's published pricing, which the company last updated on September 7, 2026:

• Search API calls — $1 per 1,000 calls, an 80% discount on the $5 list price that has applied automatically since July 16, 2026

• Result limits — the first 10 results inside a search are free; extras bill at $0.50 per 1,000 results

• Other endpoints — Image Search and Video Search at $5 per 1,000 calls; Extract at $1 per 1,000 successfully processed URLs

• Embeddings — octen-embedding-8b at $0.07 and octen-embedding-0.6b at $0.01 per million tokens

• Throughput plans — free tier up to 10 QPS; Base up to 20 QPS at no charge once the account holds credit; Builder up to 50 QPS at $2,099/month; Pro up to 200 QPS at $13,999/month; Max up to 500 QPS at $33,999/month

• Registration — $5 in free balance

The throughput tiers are the part that quietly breaks the cost math. That $0.058 per task is a pure per-call calculation. If your workload needs 200 concurrent queries per second, you are paying $13,999 a month before the first search runs, and the per-task figure stops being the interesting number.

There are also the gaps that were flagged when Octen launched: zero-data-retention terms, compliance certifications and regional availability are not published in the same detail as they are by established providers. If you are shipping into a regulated environment, that matters considerably more than a three-point index gap.

If it is third, why would anyone still pick Perplexity Search?

The three-point gap to Perplexity Search (medium) is smaller than it looks, and so is the case for switching outright. Perplexity Search wins on AA-Omniscience by 72 to 66, and it returns a synthesised, cited answer in a single call, which some pipelines want and this index — built around a tool-calling agent — does not reward. Octen Search's advantages are speed and price: about 1.7 times faster per task and roughly 36% cheaper per task than Perplexity Search (medium) on these figures.

The practical summary: if your agent is latency-bound, or your bill is dominated by retrieval at volume, Octen Search is now a defensible default rather than a gamble. If your pipeline depends on a synthesised answer, or on the compliance posture of an incumbent, the index gives you no reason to move.

Where this leaves your agent stack

The Search Index measures a layer most teams treat as plumbing, and it makes a point that reaches past search. On Octen Search's own row, the model's tokens cost 5.4 times the retrieval bill. Whatever you choose for retrieval, the model side is where the money goes, and it is the side you least want to re-integrate every time you audition a new search vendor.

That is the case for keeping the model path behind a single endpoint. OrcaRouter serves 200+ models behind one API at 0% markup — provider list price passed through, so a vendor price cut reaches you the same day — with automatic failover across providers and a routing DSL for composing several models into one call. Keep the model route fixed, swap the retrieval layer behind it, and if the new provider turns out to be weaker on your traffic than it is on DeepSearchQA, you have changed one dependency instead of two. You can see the catalogue at OrcaRouter's model catalogue.

To be explicit about what we do and do not serve: Octen Search is not in OrcaRouter's catalogue and we do not proxy it — that call goes to Octen directly. What we route is the model on the other end of it.

Tontonan Berikutnya

Three things would move this story. Artificial Analysis re-runs the index as providers ship changes, so a second data point at or near 77 would turn a debut into a trajectory, while a fall would suggest the first placement was flattered by the query mix. Octen has not published its own per-benchmark DeepSearchQA, BrowseComp and AA-Omniscience numbers, which is the cheapest credibility available to it. And the $1-per-1,000-calls rate has been in place since July 16; if that promotion lapses back to the $5 list price, the search line on Octen Search's cost roughly quintuples and the price case weakens considerably.

For anyone building an agent this quarter the move is small. The index is a quality-per-cost signal, not a compatibility test. Run your own query distribution against Octen Search before you move a production retrieval path, and treat a 77 as a reason to benchmark rather than a reason to migrate.