OctenAI

Search infrastructure for AI. LLM-native real-time web search, extract & embeddings for agents. The complete retrieval stack.

[OctenAI @OctenAI]

Try it now: https://t.co/gZ4QecsccJ

[Kuan Zou @KZouAPT]

Introducing Octen Model Gateway.

One API key. Models and search together, with the fastest responses in the world.

Thousands of developers have joined in the last 2 weeks. To say thanks, everyone gets a 15% rebate on all consumption.

Try it now in the comments:

[OctenAI @OctenAI]

API: https://t.co/4VmpUlxXmh
Pricing: $0.1 / 1k calls (Lite)
Tech report: https://t.co/XP1rf3EJIF
Open source: https://t.co/TJSum0wXSI

[OctenAI @OctenAI]

Octen Pulse, our new search API, returns results in 10ms, fast enough for models to search while they reason.

It's open source, with a tech report on how we built it. Try it and tell us what you build. Link below.

[Kuan Zou @KZouAPT]

Today @OctenAI is releasing Octen Pulse, web search at the speed and cost of inference. The model is the index, so one search takes ~10 ms, about one LLM token. We're open-sourcing it with a tech report so you can test it yourself. Link below.

[OctenAI @OctenAI]

Give Jev a wider information space with Broad Search. More searching in parallel, less waiting on retrieval.

[Kuan Zou @KZouAPT]

I think a lot of the demos coming out featuring Jev @typesafeai are cool, but I think we built something that's far better suited for actual workflows by pairing it with Broad Search.

Jev is far too powerful to be searching the web one query at a time. Giving it a much broader information space by pairing it with parallelized search allows it to be much more useful in understanding, connecting, and reasoning across information.

And that's exactly what @OctenAI built Octen Broad Search for!

Here's an example of how it could work:

[OctenAI @OctenAI]

#1 in speed. #2 in cost efficiency. Top 3 in quality. The only search API in the top tier of all three, measured independently by Artificial Analysis.

[Kuan Zou @KZouAPT]

10 months ago, we started Octen with one belief: search infrastructure built for humans won't be enough for AI.

Most of the industry saw the same shift, but kept building with a human-search mindset. AI doesn't search like humans. It fires hundreds of queries in parallel, reads everything it retrieves, and needs clean, grounded content back in milliseconds, at a cost that still works at scale.

So we didn't adapt human search for AI. We rebuilt it from first principles.

Today, Artificial Analysis published its independent Search API benchmark. Octen is the only major provider ranked Top 3 in all three dimensions that matter:

#1 in speed
#2 in cost efficiency
Top 3 in search quality

We're also the only provider in Artificial Analysis' "Most Attractive Quadrant" for search quality vs. cost.

What makes me proudest is how we got here.

Many companies on this leaderboard have been building for 3–5 years and have raised hundreds of millions of dollars. Octen is 10 months old and has raised just $10M.

To me, product built per dollar raised is one of the most honest measures of technical strength in this industry. We've kept the team small but exceptional, moved fast, and put nearly everything into the core technology.

People used to assume search had to trade off speed, cost and quality. We've shown it can be faster, cheaper and better at the same time.

The product engine is working. Next, we're bringing Octen to every model, agent and physical AI product that needs to search.

Humans search with Google. AI searches with Octen.

We're just getting started. 🚀

[Kuan Zou @KZouAPT]

A new Harvard and Stanford paper evaluated 10 frontier LLMs against 26 embedding models across a 37-task benchmark.

Octen-8B ranked #1 among all embedding models at 77.2, a statistical tie with the top LLM: Gemini 3.1 Pro, at 77.6.

The cost gap is striking: $154 vs. $0.11 per benchmark pass (1,431x).

The paper's conclusion is a division of labour, not a replacement: embedding models for classification, similarity, and clustering; LLMs for reasoning-intensive retrieval. We think that's the right read.

You can read the whole analysis here: https://t.co/eVnAydnMYU

[Kuan Zou @KZouAPT]

Our Web Search API is the cheapest.

On top of that, we are now giving you the full text of the top 10 URLs by default, completely for free, on every search. 

We believe having the full context is a fundamental need for AI agents, so it should just be the default.

If you build agents, your typical calls ask for 10 results or fewer, so with this feature you may just stop paying for page content entirely.

Beyond the first 10, the price is only $0.50 per 1,000 results.

[Kuan Zou @KZouAPT]

We just benchmarked Octen Extract against 11 other fetch APIs across 100 pages.

The result: #1 overall with an 88% success rate.

We are not the single best on every individual aspect, and the report shows that openly.

But we took the top spot by holding up everywhere rather than just winning one category.

Why is that important?

Your AI reads every kind of website, so your provider's blind spots instantly become your own.

That is why consistent performance across all page types is better than perfection in just one.

And the most unique features of Extract:
Every result includes page structure and category metadata, giving your agent useful context about the page without requiring it to process the full content first. 

And you only pay for successful fetches.

The full evaluation report and reproduction scripts are live here: https://t.co/kdeOPNBCgN

[OctenAI @OctenAI]

With Extract, you no longer get just the page's content, you get a full understanding of the page. Extract is already available, and the best part: only successful URLs are billed.

Get started here: https://t.co/bDDdzL3Qlm

[OctenAI @OctenAI]

1. Page structure: Every page comes back labeled with what it is (article? login wall? CAPTCHA?) and what domain it belongs to. That way, a login wall or error pages never reach your context.

2. Topical Classification: Every response also carries a topical classification, meaning each batch gets sorted before it ever reaches the model: official and regulatory sources are treated as primary facts while restricted or sensitive domains stay out of the run.

3. Highlights: Octen Extract also returns relevance-ranked passages instead of the full page, so you get a highlight that gives you the exact answer the full page would, at a fraction of tokens.

[Kuan Zou @KZouAPT]

@OctenAI's Model Gateway is live on Product Hunt today! One API key for every frontier model.

Add one line and your favorite frontier model searches the live web itself with no tool definition, no handler & no loop.

I will be in the comments answering everything. Get 15% off your rebates and tell us how we're doing.

[Kuan Zou @KZouAPT]

Introducing Octen Model Gateway.

One API key. Models and search together, with the fastest responses in the world.

Thousands of developers have joined in the last 2 weeks. To say thanks, everyone gets a 15% rebate on all consumption.

Try it now in the comments:

[OctenAI @OctenAI]

@KeiSakaiX That is strange. Sending you a DM to look into it.

[OctenAI @OctenAI]

The fastest search for AI agents, validated by thousands of developers this past week. 62ms P50, built for concurrency. Try it yourself.

[Kuan Zou @KZouAPT]

https://t.co/74U3uhCebz

[OctenAI @OctenAI]

@PakaJoe Welcome in. Go push it hard and tell us where it bends.

[OctenAI @OctenAI]

@etoNEtrofim $1 per 1,000 calls. Field average runs about $6.20, so roughly 6x cheaper. Free credit to try at https://t.co/Kx1swA13q9.

[OctenAI @OctenAI]

@shengkun_ye @monid_ai Here we go. Fastest search on the planet for agents. Come point yours at it and see.

[OctenAI @OctenAI]

@thedevchandra @Zendesk @sarahfim @composio Loving the search spotlight!

[Dev Chandra @thedevchandra]

Last night, I showed off @OctenAI at the OpenAI hackathon with @Zendesk, then helped judge 5 impressive finalists. It was fun to share the stage with @Sarahfim of @Composio. And the best part was that devs kept saying the same thing: our search is crazy fast!