Opus 5.5 can write you a beautiful market analysis. But without actual price data, it's still just a very confident f*cking guess. @ExaAILabs and @PythNetwork are testing a different approach: give AI agents web research and market data directly. An AI research agent needs two things: → somewhere to read (Exa's agentic search across the web and specialist sources) → something to measure (Pyth's MCP server: feed discovery, live prices, history, candlesticks)
brandNo recent capture
Exa on X
Praise is the largest theme on the page — 20 people, many of them running agents on it — and next comes excitement over the pairing with Pyth's market data (12 people). The pushback is narrow: what a search costs per task, and what the benchmark claims prove.
First written 27 Oct · what changed · from 66 posts by 63 people
Tracking posts that say “@exaailabs”
Track this for your marketPosts, last 24h
—
Last captured 36h ago
Posts, last 7 days
139
19 in the last 72h
Velocity
20/day
7-day average
Momentum
—
Needs a recent capture to measure
Posts per day, last 30 days
as of Nov 3, 11:01 UTCWhat changed
as of 27 Oct · checked once a monthFirst written 27 Oct. When what people say changes, it shows here.
Tone· 43 posts from 40 people· written 27 Oct
How do people feel about Exa?
Praise is the largest theme: 20 people thank the team, recommend Exa or credit it inside their own stack, and the next largest, 12 people, are excited about its pairing with Pyth's market data. Doubt is a small minority, aimed at the benchmark claims and at what the approach adds.
Praise is broad rather than tied to one use: people describe Exa as the search layer under working agents — finding established projects to copy instead of letting a model hand-roll slop while coding, grounding a research assistant, preparing a client meeting without leaking data. The compliments are specific: search APIs designed for LLM workloads rather than retrofitted, and "clean docs, good comms, team transparency". One user says it took two to three months to see the value and began paying once the $10 limit was hit; another says search quality had been the weak spot for so long.
The Pyth news is described the same way by nearly everyone who posts it: Exa finds the earnings news, Pyth's MCP server returns prices, history and candlesticks, and the agent can say how a stock traded in the sessions after. Several of those posts add that it is early testing.
4 people doubt the claims — "every benchmark we tested", a missing DeepSeek comparison, asking why use Exa at all. 3 people say the price pays for itself per task, and 2 praise Dynamic Highlights for returning only snippets.
- Praises it20
- Excited about the Pyth market-data pairing13
- Doubts the claims4
- Worth the cost4
- Dynamic Highlights saves tokens2
See 5 of 43 postsHide the posts
Posted about @ExaAILabs testing @PythNetwork's data last week, but I went back to the blog and something else caught my attention 👀 The earnings example is pretty simple: Exa finds the news, Pyth's MCP server brings the prices and candlesticks from the sessions that followed. But the blog also mentions where this could go next: automated market monitoring, cross-asset research and market reports. Imagine an AI agent following financial news, checking how different assets reacted and putting together a report without you having to search for all that data yourself. Still early testing, but I'm curious to see how far they take this. Read more 👇 link
At the moment, most AI tools are good at reading articles, filings, and web pages. But when it comes to stuffs involving markets, they often have to guess or come up with approximate numbers, which is unreliable. This is where @PythNetwork🤝 @ExaAILabs comes into play. - Exa handles the why and the narrative news, events, company info and more. - Then Pyth handles the what actually happened in the market live prices, historical prices, performance over the relevant periods, candlesticks. An agent can then combine both cleanly in one answer producing a well put together Financial research.
@ExaAILabs pls continue to deliver Took me 2-3 months before I figured out how valuable it was I use it so much. I even started to pay this month cause the $10 limit got hit When coding, I ask Exa to find established projects who solved this before. List them & their approach. I pick best one, agent git clones & copies 1:1 preferably Instead of model handrolling slop
A new (tied) #1 on the Search Index! I tested both of these prior to @ExaAILabs 3.0 launch, the lite offering has an excellent Cost per Task/performance tradeoff. If your agent needs web search, check out our Search Index for performance/cost/latency tradeoff comparisons across 1700 tasks.
5 of the 43 posts behind this answer, one per person. The other 38, and every new one as it’s posted, come with tracking. Get early access
Pain points· 2 posts· written 27 Oct
What do people complain about with Exa?
Price is the only complaint here: 2 people say Exa costs too much, one of them because the per-search fee is most of an agent's bill for a task. No other problem is raised.
Both posts in the section are about what a search costs. One works through a published task bill and finds that 7.6 searches at $7 per 1,000 queries comes to about $0.053 of the $0.066 a task costs, so the search fee is most of the bill and a cheaper local model would save only around a fifth. The other is a blunt dismissal rather than an argument: a paid engine that people are told to stop recommending.
Nothing in the section covers reliability, missing features or trust in the data. With 2 posts, price is the only complaint with any weight here, and it is said about the per-search fee rather than about the subscription.
- Expensive per search2
Alternatives· 11 posts from 10 people· written 27 Oct
What are the alternatives to Exa?
DeepSeek is the comparison people name most: 4 people weigh Exa against its $0.75-per-1,000-query search API, and they split on whether a cheaper search is a cheaper answer. Every other alternative gets a single mention.
DeepSeek's search and fetch APIs at $0.75 per 1,000 requests set off the comparison. One tester ran 246 research questions through Astra in Codex and found Exa answered 60.4% correctly against 47.6% for DeepSeek, 47% faster at the median and 33% cheaper overall; DeepSeek saved about 13¢ on search per attempt, but the model spent 67¢ more in tokens and made 38% more search calls, so a correct answer came to $1.81 with Exa and $3.43 with DeepSeek. Others read it differently: one says DeepSeek offers the same quality for less and that Exa left it out of its comparison, and another says the quality gap leads to more searches per task.
The rest are single mentions. One person calls Serper cheaper; one keeps SerpAPI as the fallback behind Exa; one asks how Exa compares with Firecrawl. A benchmark puts Exa's fast tier first at 99.3% and $2.10 against getcontextdev's 98.0% at $1.32, within the same confidence interval; one points to evals comparing Exa with OpenAI, and one adds treg_ai as another good option.
- DeepSeek5
- Serper1
- SerpAPI1
- Firecrawl1
- getcontextdev1
- OpenAI1
See 5 of 11 postsHide the posts
@Deepseek_ai just released search + fetch APIs at $0.75 per 1,000 successful requests each. i wanted to know whether that makes @OpenAI's Codex cheaper to run, or whether @ExaAILabs still earns its higher price. i tested 246 research questions with Astra in Codex. with Exa, Codex answered 60.4% correctly versus 47.6% with DeepSeek on the reviewed answer key. it also took 47% less time at the median and cost 33% less overall. paying more for retrieval ended up saving money across the whole attempt. link
@Deepseek_ai saved about 13¢ on search + fetch per attempt, but Astra's model-token cost went up by 67¢. the full attempt ended up costing $1.63 versus $1.09 with @ExaAILabs. Astra made 38% more search calls with DeepSeek. when you're choosing a search provider, that behavior matters: your model can keep searching, reading results and making more model calls, even when each individual search is cheap. here, the total cost of getting a correct answer in your own agent, including model tokens and unsuccessful attempts, came to $1.81 with Exa versus $3.43 with DeepSeek. link
Exa conveniently forgot to include DeepSeek's search API. DeepSeek's API is $0.75 per 1k queries and the same quality. link
not close in terms of quality and as a result leads to more searches per task
Serper is cheaper
5 of the 11 posts behind this answer, one per person. The other 6, and every new one as it’s posted, come with tracking. Get early access
Questions· 11 posts· written 27 Oct
What do people ask about Exa?
The evidence questions are the largest group — 4 people ask how the evals were run, what they measure and what latency looks like — with 3 people asking which search tool to pick at all, and 2 and 2 asking whether a feature is live and whether they can try it.
The evidence questions are specific: whether the evals measure whether results actually let an agent finish the task, whether the gains hold on short agent queries as well as longer research ones, and whether the results split by vertical like coding and GTM. One wants p50/p95 latency for retrieval, highlights and generation against full text, to judge the compression tradeoff for interactive assistants; another wants the Revolut benchmark written up, with its providers and how many hops the hardest queries needed.
The tool-choice questions are broader: why use Exa at all, what is best and cheap for news and article search, and which company plays Exa's role for Chinese content.
2 people ask whether a feature is live — whether the Codex plugin has been updated, and whether Dynamic Highlights is generally available or still a research preview, since the docs and the blog post disagree. 2 ask about access: whether they can use it, and whether there is a free trial.
- How does it measure up in tests?4
- Which search tool should I pick?3
- Is this feature live?2
- Can I try it for free?2
See 5 of 11 postsHide the posts
For interactive assistants, I'd compare time to a supported answer as well as tokens: retrieval + highlights + generation, including any full-page fallback. Do you have p50/p95 latency for that path versus full text? That would make the compression tradeoff much easier to judge.
Splitting the evals out by vertical like coding and GTM is more useful than one headline number. Did the gains hold up the same on short agent queries, or mostly on longer research ones?
For the agent search evals, do you measure whether the results actually let it finish the task?
Is Revolut's benchmark written up anywhere? Curious which providers were in the comparison and how many hops the hardest queries actually needed.
@ExaAILabs hires only the best marketing engineers /s my question: why should i use exa?
5 of the 11 posts behind this answer, one per person. The other 6, and every new one as it’s posted, come with tracking. Get early access
Wishes· 5 posts· written 27 Oct
What do people want from Exa?
5 people each ask for something different, so no wish is shared: a map of how to navigate websites, pages cleaned of menus and SEO filler, an open-source embedding model, a side-by-side comparison of the search endpoints, and no training on API data.
Each request comes from one person and concerns a different part of the product. One proposes an internet registry of how to navigate any website, so an agent can ask Exa for a site already mapped out instead of reading it again, saving tokens and time. One wants pages stripped of menus and SEO filler while keeping code examples, error messages and the explanation around them, so an agent stops burning context on clutter. One asks whether the embedding model could be open-sourced. One wants a comparison across the various search endpoints. One asks Exa to stop training on API data, saying that would let them integrate it into their platform.
Nobody repeats anyone else's request, and none of the wishes is about a feature the others also mention.
- A navigation registry for websites1
- Cleaner pages for agents1
- Open-source the embedding model1
- Compare the search endpoints1
- Stop training on API data1
See the 5 postsHide the posts
is @ExaAILabs going to build a internet registry of how to navigate any website (for agents) ? agents go to a website, query Exa who has already mapped it out for the agent -> agent saves token and time for its principal, and Exa profits
Getting rid of the menus and SEO filler would help a lot. Keep the code examples, error messages and the explanation around them, so the agent isn’t burning context on page clutter.
Any chance to open source this embedding model?
A comparison like your Jev once between various search endpoints would be super interesting!
would be super handy if you stopped training on api data guys - we could integrate you into our platform then!
In their words
How do people talk about Exa?
The phrases people use most, counted once per person.
- 1“market data”8
- 2“mcp server”5
- 3“early testing”4
- 4“history and candlesticks”3
- 5“per task”3
- 6“agent search”2
- 7“search endpoints”2
- 8“dynamic highlights”2
- 9“feed discovery”2
Posts are shown the way X shows them. One of them yours? Here’s why it’s shown, and how to take it off.
Also in Web data for AI
- Apify—Reception is warm and practical: as many people vouch for Apify outright as describe a real job it delivered.247 posts / 7d
- Firecrawl—The biggest thread is the output: people praise clean markdown and structured JSON an agent can read, while the free US government records index draws the questions and wishes.232 posts / 7d
- Browserbase−48%Reception is warm: plain praise leads at 11 people and real work follows at 8, with login handling the feature people name.201 posts / 7d
- Tavily—The largest group of people says Tavily quietly does the search job inside their agents and apps; the argument around it is price, and how it scores and bills next to Exa and DeepSeek's new, far cheaper Search API.136 posts / 7d
Get every post about Exa, filtered and labelled
This page is the summary. Tracking gives you the stream behind it for any brand, keyword or @mention: each new pain point, request and question as it’s posted, with the spam and promotions already taken out.