Glean

Rohan Paul

@rohanpaul_ai

Token costs "doubling every 45 days" against a productivity lift of "5%." Glean founder says its becasue of architecture. Enterprise AI bills are rising because cheaper tokens now power longer, more complicated task chains. Today, one prompt can trigger retrieval, tool calls, and multi-step reasoning loops. Hence the number of tokens needed for a task is increasing faster than the rate of decrease in the token price. Routing and context are better than raw model swaps because they are engineering based. Cheap outputs cost a lot when mistakes mean more work, more tries, more supervision, or angry customers. “Permission-aware context means the models don’t have to learn something new about the company every time a request is made,” Glean says. Its benchmark showed that it had ~30% less tokens and 2.5x more preferred answers than other MCP tools. That gap grew as the tasks grew harder, and the reported context-layer win rate grew from 66% to 73%.

Aakash Gupta

@aakashgupta

Glean has been running an enterprise index since 2019, across 275+ applications. In 2026 the rest of the market is catching up to the same conclusion. The catch: most are arriving at the floor and treating it like the ceiling. "We index your data" can describe wildly different systems. Per-app embeddings improve retrieval inside one source. But a customer lives in the CRM, the support queue, call notes, and planning docs. If the index doesn't unify those into one permission-aware foundation, the model has to reconcile fragments on every query. And coverage gaps are silent. No error message. Just confident answers with pieces missing. Glean's argument: indexing is one of five parts of a system of context. Indexing finds the information, graphs connect it, memory carries it forward, connectors pick the right retrieval path for each source, and tools act on it. Each part compounds the others. The buyer question has shifted. Ask how the index works: does it unify across apps, preserve entities and permissions, and prepare relevance before the model spends a token? Get that right and tokens go toward reasoning instead of sorting weak context. Then ask what's built beyond it. Indexing is table stakes now. The system around it is where the next few years get decided.

George from 🕹prodmgmt.world

@nurijanian

Good essay on the context management market that is converging on the same product 1. Jevons Paradox makes context management inevitable. As agent adoption goes up, the volume of text-based data (code output, writing, agent traces) explodes beyond what Slack, Notion, and GitHub were designed to handle. 2. The landscape is fractally crowded but most players aren't real competitors - personal knowledge bases (GBrain, Obsidian), agent memory layers (Letta, Engram), observability tools (Braintrust), agent dev tools (Mintlify), integration companies, data moat builders (Applied Compute), established companies pivoting (Notion, Glean), AI employees (Viktor), and dedicated company brains (Stash, Sentra) - A 4-person design agency and a billion-dollar tech company buy the "same" product category - Some players are genuinely complementary at different parts of the stack - Frontier labs have their own internal FDE/product teams working on it too 3. Five patterns are emerging across all players: - multiplayer/version control - accumulating skills (token spend benefits the org as a reasoning cache) - hybrid retrieval (vector + KG + keyword) - dream-time compute (separate custodian agent that indexes/deduplicates while the worker agent works) - external connections with inherited permission scopes We all know this is not consolidated or solved yet, so there is more room for more solutions. It's actually quite a fascinating market to track because on the one hand you could say well this could be a winner-takes-all market and whoever figures this out is going to be the substrate for agentic work. But will it be possible to lock it down and to prevent churn? What moats will emerge in this market will be very interesting.

Sahil Rajput

@_sahilrajput

Tried to dig deeper. This is what I found. Rasen: No website. No founder. No LinkedIn. No funding. No news. Just a name on the PMO press release. Sypha AI: Incorporated May 2025. 9 months old. One known customer, a government visa outsourcing company. Father of the founder: Ashish Chauhan, MD & CEO of the National Stock Exchange of India. Origin Bio: About a year old. Zero customers. Still in research stage. Zero revenue. Mother of the founder: Ashwini Bhide, Additional Chief Secretary to the Chief Minister of Maharashtra and MD of Mumbai Metro Rail Corporation. In the same room: Glean ($4.6 billion). Rubrik (NYSE listed). Miko (140 countries, IIT founders). One company doesn’t exist on the internet. One is 9 months old with a single government client. One has a parent running Maharashtra’s biggest infrastructure project. 100,000+ AI startups in India. Somebody decided these three deserved a seat at the table. I am not saying anything. I am just telling you what I found.

Chidanand Tripathi

@thetripathi58

Read this article from Glean about why AI at work keeps failing. Companies let AI search through everything, then wonder why it's not helpful. The problem is AI can find files but doesn't understand what matters: > Which customer deals are actually happening right now > Who you need to talk to about something > What decision needs to happen next > How old the information is Glean goes beyond just indexing. They build understanding of how your teams work, what's connected, and what actually matters for getting things done. AI can find stuff. Knowing what it means is what helps you work.

Suryansh Tiwari

@Suryanshti777

everyone's blaming token prices for rising AI bills wrong problem - one request now triggers retrieval + tool calls + reasoning loops - tokens per task are climbing faster than price per token drops - most teams are optimizing the wrong number the fix isn't cheaper tokens it's smarter routing - feed the model your company's context before it reasons - send hard tasks to frontier models, everything else to cheaper ones Glean tested this in Claude Cowork same model, only context changed responses preferred 2.5x more win rate jumped from 66% to 73% on complex tasks the model was never the moat the architecture is full thread from @jainarvind below

Arvind Jain

Co-founder & CEO

Former Google Distinguished Engineer and Rubrik co-founder who started Glean in 2019.

T.R. Vishwanath

Co-founder

Engineering veteran from Google and Meta who co-founded Glean's search platform.

Tony Gentilcore

Co-founder

Former Google engineer and Glean co-founder focused on the company's core technology.