Dust and Glean both put AI agents on top of the SaaS your team already runs, both let an admin pick the model per agent, and both speak MCP in each direction. The old routing rule — Glean for search, Dust for agents — stopped working in 2026: Glean Agents take actions across the same connector graph the Assistant reads, and Dust’s retrieval returns the source clause rather than a paraphrase. What separates them now is what you are buying. Glean sells a maintained index: a permission-aware graph across 100+ systems, extended into PDFs and contracts by its March 2026 acquisition of Aryn. Dust sells a builder: you choose the models, the tools, and the handful of sources you scope each agent to. So the routing question is whether your bottleneck is coverage — the answer sits in one of forty systems and nobody knows which — or construction — you know where the data lives and need agents that act on it.
Where Glean wins
Coverage is the product. Glean indexes 100+ connectors and enforces each source system’s ACLs at query time, so one question about renewal risk returns an answer stitched from Salesforce, Gong calls, the Slack deal channel, and the QBR deck. The Aryn acquisition in March 2026 pushed that reach past SaaS records into unstructured documents, turning PDFs, contracts, and research reports into schema-driven structured data inside the same index. For legal ops and customer success, where the answer is as likely to sit in a signed PDF as in a CRM field, that is the difference between an assistant that works and one that shrugs.
Governed MCP in both directions. Glean runs a remote MCP server that exposes company context to outside AI chats and IDEs, and an MCP gateway that brings third-party servers in — 17 preloaded, including Asana, Atlassian Rovo, Box, GitHub, HubSpot, Intercom, Linear, Notion, and ThoughtSpot, plus any compatible remote server — with central admin control over which tools reach which users. The March 2026 release added Tool Search, which narrows the candidate tool set to cut context consumption, Enterprise Memory, which replays action sequences that already worked, and Skills defined as SKILL.md specifications. Dust supports MCP servers natively and remotely too, but its published plan stops at 5 remote servers.
Scale proof and the compliance surface. Glean reached $300M ARR by May 2026, up 89% year over year, at a $7.2B valuation following a $150M Series F. It carries SOC 2 Type II, ISO 27001, ISO 42001, HIPAA, GDPR, and TX-RAMP Level 2. Dust holds SOC 2 Type II with US and EU data residency, which clears most commercial diligence but not an ISO 42001 or TX-RAMP line item.
Where Dust wins
You can start today, at a number you can read. Dust publishes the whole ladder: a Free seat at 0€ with 500 credits for the seat’s lifetime, Pro at 30€ per seat per month (24€ billed yearly) with 8,000 credits monthly, and Max at 150€ per seat per month (120€ billed yearly) with 40,000 credits monthly. Seat types mix inside one workspace, so occasional users cost nothing while power users get the Max allowance. Glean publishes no price at all — glean.com/pricing now redirects to the homepage — and buyers report a 100-seat minimum. Below that threshold, Glean does not quote you.
The meter is documented end to end. A Dust credit is token credits plus action credits, and the action tiers are published: 0 for internal tools, memory, and file management; 1 for web search, agent execution, and Frames; 3 for data retrieval, image and file generation, and third-party integrations. Programmatic use bills at $0.01 per credit. Credits reset monthly without rollover, and overage is capped and admin-enabled rather than silent. Glean publishes FlexCredit consumption rates — a Fast Mode assistant query runs about 3 credits at the median and about 15 at the 90th percentile — but no dollar price per credit, so you can model relative cost and not the invoice.
Model choice without a plan gate, over an open core. Dust ships 20+ models from OpenAI, Anthropic, Google, Mistral, and DeepSeek, selected per agent, with no model locked behind a higher tier — higher-capability models simply burn more credits. The core is MIT-licensed, so a team worried about lock-in can read it or self-host. Glean’s Model Hub spans OpenAI, Gemini, Anthropic, Amazon Nova, and its own Waldo model with bring-your-own-key support, but it is an admin-governed hub inside a closed platform where premium models consume FlexCredits.
Pricing reality — and the ceiling most buyers miss
Dust’s published price is real, but read what it covers. The Business plan runs to 100 people and caps you at 3 connectors out of the 20+ available, 5 Spaces, and 5 remote MCP servers. Glean’s entire premise is breadth across 100+ systems. If your reason for buying is “AI across our whole SaaS estate,” Dust’s published price does not describe your deployment — you land in Dust Enterprise, which is quote-only, exactly like Glean. Transparent pricing is a genuine Dust advantage for a scoped rollout of one team over three systems, and it evaporates for a company-wide one.
Model the two by shape. Sixty ops people on Dust Pro pay 1,800€ a month, or 1,440€ billed yearly, every number published in advance, with free seats absorbing the occasional users. Those same sixty people cannot buy Glean at all. At 300 seats across a wide estate both are quote-only: Glean at buyer-reported $40-50 per user per month for the base license plus roughly $15 per user per month when the AI tier is broken out, which puts most mid-market and enterprise contracts in the $200K-480K per year band all-in; Dust Enterprise at volume pricing over workspace-pooled credits, where the number tracks how hard agents retrieve rather than headcount. On neither platform is the headline the bill — on Dust it is the credit tail, on Glean it is FlexCredits plus the crawl and storage load each new connector adds at renewal.
Implementation effort
Both fail the same way, and it is not the AI. Glean’s ramp is index hygiene: a Drive folder shared with anyone who has the link surfaces to anyone who can ask, and ACL inheritance does not repair upstream over-sharing. Freeze the connector list at rollout and route additions through a quarterly review rather than a self-serve request. Dust’s ramp is scoping and ownership: three connectors on the published plan forces you to decide which three systems matter before you build anything, and someone has to own the agent library or you finish the quarter with fifty half-working bots and no audit story. On both, audit permissions on the top five connected systems before rollout and keep write-capable agents behind an approval step until their output earns it.
Verdict
Pick Glean when coverage is the problem — the answer is somewhere across dozens of systems including signed PDFs and contracts, and nobody knows where; when you have 100+ seats and a permissions model already in shape; when you want MCP governed centrally in both directions; and when procurement wants ISO 42001, HIPAA, or TX-RAMP on the certificate list. It is the maintained-index pick.
Pick Dust when you already know where the data lives and need agents built on top of it; when you want a price before a sales call and a rollout that can begin with one person on a free seat; and when per-agent model choice and an MIT-licensed core carry weight. It is the builder pick, and the only one of the two that a team under 100 people can actually buy.
Pick neither when your estate is essentially Microsoft 365 and Copilot inside that suite covers most of the surface at a published $30 per user per month; when you would rather own the orchestration yourself with n8n and your own model keys; or when the job is one autonomous process running end to end rather than a knowledge layer, which is Relevance AI’s ground.
Still undecided — count your seats. Under 100, the decision is already made: Dust, because Glean will not quote you. Over 100 with a wide estate, pilot Glean and keep Dust as the scoped builder for the teams that want their own agents. The two coexist more often than either vendor admits, and Glean’s remote MCP server is one of the five remote servers a Dust workspace can connect, so the index can feed the builder instead of competing with it.
Dust and Glean both put AI agents on top of the SaaS your team already runs, both let an admin pick the model per agent, and both speak MCP in each direction. The old routing rule — Glean for search, Dust for agents — stopped working in 2026: Glean Agents take actions across the same connector graph the Assistant reads, and Dust’s retrieval returns the source clause rather than a paraphrase. What separates them now is what you are buying. Glean sells a maintained index: a permission-aware graph across 100+ systems, extended into PDFs and contracts by its March 2026 acquisition of Aryn. Dust sells a builder: you choose the models, the tools, and the handful of sources you scope each agent to. So the routing question is whether your bottleneck is coverage — the answer sits in one of forty systems and nobody knows which — or construction — you know where the data lives and need agents that act on it.
Where Glean wins
Coverage is the product. Glean indexes 100+ connectors and enforces each source system’s ACLs at query time, so one question about renewal risk returns an answer stitched from Salesforce, Gong calls, the Slack deal channel, and the QBR deck. The Aryn acquisition in March 2026 pushed that reach past SaaS records into unstructured documents, turning PDFs, contracts, and research reports into schema-driven structured data inside the same index. For legal ops and customer success, where the answer is as likely to sit in a signed PDF as in a CRM field, that is the difference between an assistant that works and one that shrugs.
Governed MCP in both directions. Glean runs a remote MCP server that exposes company context to outside AI chats and IDEs, and an MCP gateway that brings third-party servers in — 17 preloaded, including Asana, Atlassian Rovo, Box, GitHub, HubSpot, Intercom, Linear, Notion, and ThoughtSpot, plus any compatible remote server — with central admin control over which tools reach which users. The March 2026 release added Tool Search, which narrows the candidate tool set to cut context consumption, Enterprise Memory, which replays action sequences that already worked, and Skills defined as SKILL.md specifications. Dust supports MCP servers natively and remotely too, but its published plan stops at 5 remote servers.
Scale proof and the compliance surface. Glean reached $300M ARR by May 2026, up 89% year over year, at a $7.2B valuation following a $150M Series F. It carries SOC 2 Type II, ISO 27001, ISO 42001, HIPAA, GDPR, and TX-RAMP Level 2. Dust holds SOC 2 Type II with US and EU data residency, which clears most commercial diligence but not an ISO 42001 or TX-RAMP line item.
Where Dust wins
You can start today, at a number you can read. Dust publishes the whole ladder: a Free seat at 0€ with 500 credits for the seat’s lifetime, Pro at 30€ per seat per month (24€ billed yearly) with 8,000 credits monthly, and Max at 150€ per seat per month (120€ billed yearly) with 40,000 credits monthly. Seat types mix inside one workspace, so occasional users cost nothing while power users get the Max allowance. Glean publishes no price at all — glean.com/pricing now redirects to the homepage — and buyers report a 100-seat minimum. Below that threshold, Glean does not quote you.
The meter is documented end to end. A Dust credit is token credits plus action credits, and the action tiers are published: 0 for internal tools, memory, and file management; 1 for web search, agent execution, and Frames; 3 for data retrieval, image and file generation, and third-party integrations. Programmatic use bills at $0.01 per credit. Credits reset monthly without rollover, and overage is capped and admin-enabled rather than silent. Glean publishes FlexCredit consumption rates — a Fast Mode assistant query runs about 3 credits at the median and about 15 at the 90th percentile — but no dollar price per credit, so you can model relative cost and not the invoice.
Model choice without a plan gate, over an open core. Dust ships 20+ models from OpenAI, Anthropic, Google, Mistral, and DeepSeek, selected per agent, with no model locked behind a higher tier — higher-capability models simply burn more credits. The core is MIT-licensed, so a team worried about lock-in can read it or self-host. Glean’s Model Hub spans OpenAI, Gemini, Anthropic, Amazon Nova, and its own Waldo model with bring-your-own-key support, but it is an admin-governed hub inside a closed platform where premium models consume FlexCredits.
Pricing reality — and the ceiling most buyers miss
Dust’s published price is real, but read what it covers. The Business plan runs to 100 people and caps you at 3 connectors out of the 20+ available, 5 Spaces, and 5 remote MCP servers. Glean’s entire premise is breadth across 100+ systems. If your reason for buying is “AI across our whole SaaS estate,” Dust’s published price does not describe your deployment — you land in Dust Enterprise, which is quote-only, exactly like Glean. Transparent pricing is a genuine Dust advantage for a scoped rollout of one team over three systems, and it evaporates for a company-wide one.
Model the two by shape. Sixty ops people on Dust Pro pay 1,800€ a month, or 1,440€ billed yearly, every number published in advance, with free seats absorbing the occasional users. Those same sixty people cannot buy Glean at all. At 300 seats across a wide estate both are quote-only: Glean at buyer-reported $40-50 per user per month for the base license plus roughly $15 per user per month when the AI tier is broken out, which puts most mid-market and enterprise contracts in the $200K-480K per year band all-in; Dust Enterprise at volume pricing over workspace-pooled credits, where the number tracks how hard agents retrieve rather than headcount. On neither platform is the headline the bill — on Dust it is the credit tail, on Glean it is FlexCredits plus the crawl and storage load each new connector adds at renewal.
Implementation effort
Both fail the same way, and it is not the AI. Glean’s ramp is index hygiene: a Drive folder shared with anyone who has the link surfaces to anyone who can ask, and ACL inheritance does not repair upstream over-sharing. Freeze the connector list at rollout and route additions through a quarterly review rather than a self-serve request. Dust’s ramp is scoping and ownership: three connectors on the published plan forces you to decide which three systems matter before you build anything, and someone has to own the agent library or you finish the quarter with fifty half-working bots and no audit story. On both, audit permissions on the top five connected systems before rollout and keep write-capable agents behind an approval step until their output earns it.
Verdict
Pick Glean when coverage is the problem — the answer is somewhere across dozens of systems including signed PDFs and contracts, and nobody knows where; when you have 100+ seats and a permissions model already in shape; when you want MCP governed centrally in both directions; and when procurement wants ISO 42001, HIPAA, or TX-RAMP on the certificate list. It is the maintained-index pick.
Pick Dust when you already know where the data lives and need agents built on top of it; when you want a price before a sales call and a rollout that can begin with one person on a free seat; and when per-agent model choice and an MIT-licensed core carry weight. It is the builder pick, and the only one of the two that a team under 100 people can actually buy.
Pick neither when your estate is essentially Microsoft 365 and Copilot inside that suite covers most of the surface at a published $30 per user per month; when you would rather own the orchestration yourself with n8n and your own model keys; or when the job is one autonomous process running end to end rather than a knowledge layer, which is Relevance AI’s ground.
Still undecided — count your seats. Under 100, the decision is already made: Dust, because Glean will not quote you. Over 100 with a wide estate, pilot Glean and keep Dust as the scoped builder for the teams that want their own agents. The two coexist more often than either vendor admits, and Glean’s remote MCP server is one of the five remote servers a Dust workspace can connect, so the index can feed the builder instead of competing with it.