Anthropic cut the price of Opus 5.5 by 20% today. OpenAI responded almost immediately, cutting Luna and Sol by about 50%.
Most people will see cheaper AI. For enterprises, this shows how quickly an AI strategy can go stale.
When price and performance shift overnight, model choice belongs in the system architecture. Models differ in reasoning, speed, cost, and task fit, so enterprise AI systems need to understand the work, assess the tradeoffs, and select the right model and reasoning level for each task.
@Glean’s auto routing addresses that problem. We evaluate models on real enterprise work, then route each task based on the quality, speed, and cost it requires. As the market changes, routing can update without forcing customers to rebuild their workflows.
Model prices will keep changing. Enterprise systems still need to understand the work and choose the right model for it.
We've seen an acknowledgment from the AI labs that models can no longer be judged on performance alone, cost has to be part of the equation. Good to see both @OpenAI and @AnthropicAI ship lower-cost models today that change the AI economics.
OpenAI's GPT-6 Luna and Sol both dropped 50% in price. @Glean evals found Opus 5.5 to cut the cost per query in half: from $0.82 (Opus 5) to $0.41, while simultaneously improving performance.
Giving AI more work requires a) the quality bar to be met b) the cost to pencil out. Really happy to see token costs heading in the right direction today, down.
GPT-6 Luna and Opus 5.5 are available in Glean today.
8
7
71
16,604
Most AI assistants are starting to look alike. The important differences will come down to what they understand about your work, and how safely they can act on your behalf.
every major company is arriving at roughly the same offering:
persistent memory, email/calendar/messages, browser + computer use, background tasks, proactive notifications, voice, app/tool execution, ambient context, & some notion of a personal agent sitting above everything.
remarkable levels of convergence with very little differentiation whatsoever.
20
5
123
17,831
Enterprise AI should understand what a company knows, and how it makes decisions and gets work done.
Operating intelligence shows up in four places.
- Individual intelligence is the judgment people apply in context.
- Managerial intelligence is creating the conditions for honest challenge and better decisions.
- Team intelligence is how groups coordinate and execute across functions.
- Process intelligence is understanding how work moves, where it gets stuck, and who can unblock it.
Together, they are the judgment and relationships that turn knowledge into action.
Most enterprise AI still centers on the explicit layer, including documents, tickets, and CRM records. The harder problem is the context around them. Why a signal matters. Who needs to be involved. What decision has already been made. Where work is actually stuck.
How well AI understands that context will determine how useful it becomes inside a company.
16
20
170
11,879
This is spot on. Our Work AI Index found that 69% of AI users admit to shipping work they haven’t verified, don’t fully understand, or can’t confidently stand behind.
If you can’t explain the goal and judge the result, don’t delegate it to AI.
Problem: I see developers using AI to implement tickets they don't understand.
Claude finishes the task and the developer just assumes it's right. They don't understand the ticket well enough to evaluate if AI was successful.
Rule: If I don't understand the goal, I shouldn't ask AI to do it.
2
2
10
1,537
Coordination neglect happens when AI saves one person an hour but creates more review, rework, and fact-checking for everyone else.
Organizations need to set an accountability loop around AI-assisted work. It should operate before, at, and after the handoff and the sender owns the downstream problems. If the code breaks, they fix it. If the document creates confusion, they rewrite it.
At @glean, we’ve built this accountability into the platform – grounding AI in company context and carrying permissions, approvals, and human review through the workflow. But the platform is only part of the answer. Organizations also need clear norms around who stands behind the work and fixes if something breaks.
2
5
791
Becoming AI-first is not just a technology strategy. It’s an organizational design problem.
Last week at @Glean:GO, leaders from some of the world’s largest companies—like @GM, @Ericsson, @Cisco, @Deloitte, @Mastercard, @Dell, and @GeneralMills—showed that even organizations with long histories can become AI-first. But the transition requires redesigning how work gets done, not simply adding AI to the existing stack.
A few themes stood out:
- 𝗠𝗼𝘃𝗲 𝗳𝗿𝗼𝗺 𝗿𝗲𝗮𝗰𝘁𝗶𝘃𝗲 𝘁𝗼𝗼𝗹𝘀 𝘁𝗼 𝗽𝗿𝗼𝗮𝗰𝘁𝗶𝘃𝗲 𝗰𝗼𝘄𝗼𝗿𝗸𝗲𝗿𝘀. On stage with @OpenAI’s @embirico and @CNBC’s @Kr00ney, we discussed why most companies barely tap AI’s potential: workers don't know what to ask. When AI has deep context—your role, OKRs, and daily tasks—it can proactively propose work by default, eliminating the friction of prompt engineering.
- 𝗥𝗲𝗱𝗲𝘀𝗶𝗴𝗻 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝘁𝗮𝘀𝗸𝘀. The biggest constraint is often the coordination around a task: handoffs, context switching, and silos. If the workflow stays the same, much of AI’s productivity gain is lost. This is why we built Glean Transform, to map how work actually happens and redesign it around AI.
- 𝗕𝗮𝗸𝗲 𝗶𝗻 𝗗𝗮𝘆 𝟬 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆. In my conversation with @Cisco’s @jpatel41 and @Deloitte’s Ashish Verma, a shared reality emerged. Enterprise AI cannot scale as a collection of ad-hoc initiatives. To give organizations the trust needed to hand over real, mission-critical work, Day 0 security and fine-grained data governance must be built directly into your core systems of record.
- 𝗠𝗲𝗮𝘀𝘂𝗿𝗲 𝗼𝘂𝘁𝗰𝗼𝗺𝗲𝘀, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝗮𝗰𝘁𝗶𝘃𝗶𝘁𝘆. @DaVita has saved 75,000 days of work using Glean. But as Madhu Narasimhan, CIO of DaVita, emphasized, the ultimate measure is human impact: giving a caregiver five more minutes with a patient.
- 𝗖𝗿𝗲𝗮𝘁𝗲 𝗰𝗹𝗲𝗮𝗿 𝗼𝘄𝗻𝗲𝗿𝘀𝗵𝗶𝗽, 𝗼𝗳𝘁𝗲𝗻 𝘁𝗵𝗿𝗼𝘂𝗴𝗵 𝗮 𝗖𝗔𝗜𝗢 𝗿𝗼𝗹𝗲. In a panel with CAIOs from Capgemini, Mastercard, Ericsson, and Zapier, leaders emphasized that without clear ownership, AI remains a collection of disconnected initiatives instead of becoming a new operating model.
Thank you to our customers, partners, speakers, and Glean team for showing what it looks like to build an AI-first organization in practice. We’re honored to be a partner in that journey.
11
4
34
8,460
The enterprise AI stack is being rebuilt.
Models are becoming more capable, and increasingly interchangeable. As AI moves from answering questions to doing mission-critical work, the constraint is no longer raw intelligence. It is organizational context: understanding how a company operates, what information matters, and what should happen next.
Without that context, enterprises get shallow answers, rising token costs from sending irrelevant information to expensive models, and AI sprawl as teams adopt disconnected tools. The answer is not simply more models. It is an intelligence layer that understands the organization and routes each task to the right model, agent, and workflow.
That is the thesis behind the 50+ new features we’re announcing today at @Glean:GO, including:
𝗚𝗹𝗲𝗮𝗻 𝗧𝗮𝘂: A new governed desktop AI workspace that combines an open agent harness with your enterprise context, unifying local files, applications, browser workflows, and code repositories in one workspace. That context is what lets agents delegate complex, multi-step work and deliver better performance, while respecting permissions.
𝗚𝗹𝗲𝗮𝗻 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲: Uses organizational context to route each task to the right model, while giving enterprises control over quality, usage, and cost. Our benchmark found that Glean reduced token costs by 81% as compared to Claude Cowork, and was preferred 78% of the time.
𝗚𝗹𝗲𝗮𝗻 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺: Built on your Enterprise Graph, it understands how work happens across your organization, identifies the best opportunities for AI, recommends the right agents and skills, and measures the impact on your business.
To help us announce these new capabilities, I’m looking forward to being joined by leaders from some of the world’s most innovative organizations, including @GM, @nvidia, @Snowflake, and @OpenAI, who will share how they are using Glean in mission-critical workflows.
See you today at Glean:GO in San Francisco and online.
56
16
315
52,272
Glean:GO making an appearance in NYC 🗽
See you tomorrow at 9 am PT for our morning keynote. glean-it.com/4gxHM5c
6
5
21
1,333
Enjoyed this conversation with @ricmac.
Organizations are looking for the flexibility to choose the right model for each task while maintaining context, security, control, and governance needed to deploy AI broadly. Capabilities that go beyond what a single model provider can offer.
The industry is moving toward a multi-model and open future. Proud that @glean will play an important role in making that future practical.
Model routing is hot news (see Stripe / OpenRouter). It's also increasingly important in enterprises. In my latest @latentspacepod article, I talk to @glean CEO @jainarvind about how model routing helps control AI costs for organizations. latent.space/p/glean-model-r…
5
2
14
2,148
Owning intelligence means owning what the system learns from.
That includes your context, your definition of quality, your workflows, and every workaround users make in production. The model is an ingredient, but your accumulated judgment is the proprietary asset.
The race for the AI application layer is not only about UI, workflows, or GTM... it is a fight for the intelligence layer itself.
1
2
25
4,181
The economics of AI matter just as much as their capabilities. They are shaped by the architecture around the model, not only the model itself: how work is routed, how context is retrieved, and how each task is orchestrated.
In our benchmark, @Glean was 4x more cost-effective, averaging $0.45 per task versus $1.84 for Claude Cowork, a 75% reduction in cost. That gap came from making better decisions at every layer of the stack.
The best systems will not send every task to the most powerful model. With the right context, they can understand what each task requires and route it to the least expensive model capable of meeting that bar.
We've been investing heavily in our harness and routing capabilities, and we put them to the test benchmarking Glean's token costs against Claude Cowork. The results were striking: @Glean is 4x more cost-effective, averaging $0.45 per task versus $1.84 for Claude Cowork.
That 4x advantage comes from two things compounding: 2.9x lower token volume and a 1.4x cheaper blended rate per million tokens. Here's how:
- Model family routing: Glean made use of Luna which is 10x cheaper than Claude Sonnet and widely capable. We’re able to strike the balance by routing between open and closed models.
- Model tier routing: In Glean, Opus was used 10x more (29% vs 2.8%) but surgically for the right things and balanced by other models.
- Better context: Glean’s harness and indexing capabilities result in fewer tokens consumed; Claude Cowork used 3x the tokens per query on average using 88.8M versus 29.8M in Glean.
More results coming out at Glean:GO! Hit me up if you're still looking for an invite.
2
1
32
4,229
AI slop is everywhere. 69% of AI users admit to shipping work they haven’t verified, don’t fully understand, or couldn’t confidently defend, according to @glean's Work AI Index.
And even when AI output is factually correct, it can still be generic, too long, or clearly not written by a human.
The deeper problem is what researchers call “mental proof”: visible evidence that someone engaged with the work, cares about the result, and understands the person on the other end. When work feels like slop, it erodes trust and relationships.
This is hard to solve. At Glean, we’re still figuring out how to keep AI from making our internal communication longer, blander, and less human. A few principles guide us:
- 𝗜𝗻𝘃𝗲𝘀𝘁 𝗶𝗻 𝗮 𝗰𝗼𝗻𝘁𝗲𝘅𝘁-𝗿𝗶𝗰𝗵 𝗔𝗜 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺. AI can help protect against slop when it understands the context around the work. It should match the length of a communication to its complexity. When communication involves a sensitive relationship, ambiguity, or higher stakes, it should recognize that more human judgment may be needed and prompt the user to pause or add humanity.
- 𝗟𝗲𝗮𝗱𝗲𝗿𝘀 𝘀𝗵𝗼𝘂𝗹𝗱 𝗺𝗼𝗱𝗲𝗹 𝘁𝗵𝗼𝘂𝗴𝗵𝘁𝗳𝘂𝗹 𝗔𝗜 𝘂𝘀𝗲, 𝗮𝗻𝗱 𝗺𝗮𝗸𝗲 𝗶𝘁 𝘀𝗮𝗳𝗲 𝘁𝗼 𝗽𝘂𝘀𝗵 𝗯𝗮𝗰𝗸. I try to model this myself, and push my team to do the same by saying something like: “This is too long. Please rewrite it more concisely.”
- 𝗧𝗵𝗲 𝗽𝗲𝗿𝘀𝗼𝗻 𝘄𝗵𝗼 𝘀𝗵𝗶𝗽𝘀 𝘁𝗵𝗲 𝗼𝘂𝘁𝗰𝗼𝗺𝗲 𝗺𝘂𝘀𝘁 𝗿𝗲𝗺𝗮𝗶𝗻 𝗮𝗰𝗰𝗼𝘂𝗻𝘁𝗮𝗯𝗹𝗲 𝗳𝗼𝗿 𝗶𝘁. Otherwise, it’s too easy to deflect responsibility to the AI.
- 𝗠𝗲𝗮𝘀𝘂𝗿𝗲 𝘄𝗵𝗮𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀. If you measure vanity metrics like output volume instead of quality and outcomes, you’ll optimize for more output, even when that output creates more work for everyone else.
More AI output does not create more value if people spend their time sorting through slop.
21
8
106
12,976
Models can reason over context, but they shouldn’t be trusted to define their own identity, permissions, scope, or environment.
The control plane around the model must continuously verify those conditions–using deterministic access controls, permission-aware retrieval, auditability, and least-privilege execution. Enterprise AI needs guardrails enforced by infrastructure, not inferred from instructions.
Alignment without context integrity is not safety.
An AI agent can faithfully follow its instructions and still be dangerous if its understanding of reality is wrong.
Anthropic recently disclosed that Claude models gained unauthorized access to the real systems of three organizations during cybersecurity evaluations.
The agents had been told they were operating in a simulation with no internet access. But a configuration mistake gave them access to the live internet.
They treated real production systems as part of the exercise and kept pursuing the goal they had been given.
OpenAI separately disclosed that models found a previously unknown vulnerability, escaped an isolated evaluation environment and compromised Hugging Face.
These were not simply failures of intelligence.
The deeper problem was that the agents were acting inside a false understanding of reality.
We have spent years asking whether an AI system will follow our instructions.
We now also need to ask whether it correctly understands the environment in which those instructions are being executed.
This creates a new security requirement.
Context integrity.
Before an agent acts, the system must continuously verify where it is, which resources are in scope, whose authority it carries, what it is allowed to do, and when that authority expires. Just in time permission for every action. At just the right time. For just enough time. Assessed in real time.
Those facts cannot live only inside a prompt. They must be verified and enforced by the infrastructure around the model.
A prompt is not a security boundary.
Zero trust taught us to never trust identity and always verify access. And provide least privileged access.
Agentic AI adds another dimension.
Never blindly trust context. Continuously verify reality.
The next security perimeter is not just the agent’s identity.
It is the agent’s understanding of reality. The most dangerous agent may not be misaligned. It may simply be mistaken. And in an agentic world, a false belief can become a real breach.
4
1
38
8,282
In many discussions, I find that context is often being referred to as a synonym for more data, but more data doesn't necessarily mean more context.
Organizational theory is a good way to think about context, which breaks enterprise knowledge into four categories: know-what, know-how, know-why, and know-who.
Take a sales deal.
𝗞𝗻𝗼𝘄-𝘄𝗵𝗮𝘁 is the account plan, pricing, security questionnaires, and redlines.
𝗞𝗻𝗼𝘄-𝗵𝗼𝘄 is how to move the deal forward, like when to involve security and how to sequence the work.
𝗞𝗻𝗼𝘄-𝘄𝗵𝘆 is rationale, like which objections indicate real risk and which approvals are routine.
𝗞𝗻𝗼𝘄-𝘄𝗵𝗼 is the social knowledge, like who decides, who has handled a similar issue in the past, and who needs to be involved.
Most of what AI can retrieve today falls in the know-what category. The rest is trapped in scattered conversations, unwritten routines, and informal human networks.
We've spent the last 7 years at @Glean building toward this gap. Our context layer to connect know-what, know-how, and know-who, while inferring know-why from how work actually happens.
An AI tool with access only to know-what can find the account plan but will fall short in telling you who to involve, which objection deserves attention, or what to do next.
It’s the difference between knowing the account, and knowing how to move it forward.
4
2
27
1,956
The model layer is moving fast. Prices, speed, and capability are all changing. Enterprises shouldn’t have to rebuild their AI stack every time it does.
We built @glean for this reality, a model-agnostic context and intelligence layer that lets companies use the best model for each job, with the knowledge, permissions, and workflows to make it useful.
Model choice will keep changing. Enterprise architecture shouldn’t.
We are committed to pushing the model frontier across cost efficiency, capability, and speed.
Starting today, we are reducing prices for GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% , and offering a faster option for GPT-5.6 Sol in the API.
Luna and Terra’s lower prices are reflected in how usage is counted in Codex and ChatGPT Work, so your usage goes further.
7
7
39
4,960
AI is exposing a workflow design problem, not just a productivity problem. When the cost of building falls, the scarce resource is knowing what is worth building. The key advantage goes to teams that can create shorter loops between evidence, decisions, and action.
AI closed 90% of your support tickets. you released 10 new features this month but barely have any usage metrics yet to validate them. what now?
the bottleneck is now the organizational culture. most orgs are now sitting on an unexpected surplus of capacity, and don't know what to do with it yet.
a lot of the software org culture is still based on "the old days", when work was slower, planning took weeks and development took months. so culture was optimized to do things as fast as possible, as close as possible to "right" on the first try.
the immediate response from orgs to AI, in most cases, has been to keep the exact same approach and processes - just do the same thing, faster!
the problem is planning can't catch up. work planned for the whole quarter is done by the first month, and planning for the next quarter is still half baked.
a lot of people, myself included, have asked: "so... where is all the new software that AI is building?"
I think this question is mostly answered by the fact that most organizations, so far, aren't producing much more software. they're just producing the same thing, faster. so the remaining time they are either wondering what's up, doing internal politics, or just cutting half the engineering team because it's "superfluous".
the culture in software must adapt and switch mindsets. the edge is not anymore in producing the same with a smaller workforce. the edge is in leading the same workforce to produce MORE.
1
1
16
1,767
Over the next few months, I expect more AI vendors will announce indexing capabilities, reflecting a broader recognition that indexing, and a system of context built on it, is foundational to enterprise AI. The important question is how deep, permission-aware, and production-ready that index really is.
That’s good for the market. But indexing in the enterprise is not just connecting to a few apps and retrieving documents.
The hard part is building a system that can crawl data across hundreds of apps, preserve permissions, normalize identities and content, rank what matters, and keep the index fresh.
We’ve been doing that at @glean for the past seven years.
The advantage is having a solid foundation of context ready before the model starts reasoning. Instead of spending extra loops piecing together fragmented information, AI can get to a better result faster and with less token burn. Search is where this starts, but the impact is much broader than search. It improves the work AI can do across the enterprise.
So if more vendors are moving in this direction, it’s validating.
Everyone agrees indexing matters. What enterprises need to evaluate is how deep, permission-aware, and production-ready that index really is.
9
6
119
28,179
