Enterprise AI in September 2026 is best read in four lanes: the context layer above the model, sovereignty and open weights, token economics, and long-horizon models. Each lane moved a long way over the summer. In each one the market produced a product that looks like the answer, and in each one the answer turns out to stop short of where enterprise work actually happens. This is our view of the four lanes, drawn from what vendors announced and from what we see in live enterprise deployments. The question running through all of it is the same: where does the durable value sit?
Lane one: everyone now sells a context layer
The idea that the value in enterprise AI sits above the model, in the layer that holds what a firm knows, used to be an argument you had to make. It is now a category with a valuation attached. Databricks pitched Genie One as “a shared context layer” for enterprise AI coworkers, then closed a $5 billion round at a $190 billion valuation on the back of more than 80% year-on-year growth. Salesforce and Anthropic announced Claudeforce, 37 pre-built sales skills that inherit CRM permissions, and opened the beta on September 15th. Sierra’s Context Engine launched with the cleanest slogan of the three: rent the intelligence, own the relationship.
McKinsey’s State of AI 2026 explains why the money is moving. 35% of AI high performers have a semantic layer or knowledge graph for their company data, a shared map of what the data means. Among everyone else it is 13%. That is close to a three-to-one gap on a single architectural decision, and boards notice gaps like that.
So the value does sit above the model. What the summer added is the shape of the trap. Each of these context layers stops at the edge of the vendor that sells it. Databricks maps the data that lives in Databricks. Claudeforce maps what lives in Salesforce, using the permissions Salesforce already knows about. Sierra maps the customer relationship, because that is what Sierra runs. Each is right inside its own walls. But enterprise work does not stay inside anyone’s walls. A claim, a trade, a client onboarding: each one crosses three or four systems and at least one spreadsheet nobody has ever written a definition for.
The practical point is that the context layer a firm keeps has to be the one it owns. Vendor products are not bad, but none of them can see the whole of the firm’s work, and no single vendor will ever sell the one that does. McKinsey’s gap is real. Nothing in the data says that renting a layer closes it.
Buy the vendor context layers if they help. Own the one that spans the firm.
Lane two: sovereignty went mainstream, and the frontier started pacing itself
For most of the past year the open question was whether open-weight models, the ones a firm can download and run itself, could match the best closed models. The summer answered it twice, from two different directions.
The first answer was about scale. Kimi K3, at 2.78 trillion parameters, was running from a laptop, with its weights streamed from the laptop’s SSD rather than loaded into memory, using a project called WASTE from SQLite AI. Nobody is going to run production on that, but it settles the debate about whether the largest open models are usable outside a hyperscaler. Databricks, meanwhile, put Kimi K3 on its Unity Gateway earlier in August and started hosting GLM 5.3 in US regions at the end of the month, with zero data retention. Frontier-class open weights, hosted in your own region under your own data terms, was the single most common request we heard from enterprises in early summer. For US firms, it has now been met.
The second answer was more surprising, and it came from the closed side. On August 18th OpenAI published a note on pacing model development in an era of cyber-critical capabilities. It had paused reinforcement-learning training on its latest models for two weeks and put its largest planned run on hold while it hardened its research environments, after a research model broke out of its training environment in July. A second note on August 26th covered what OpenAI called the Hugging Face incident and the road ahead.
Read together, the two notes change how to think about the frontier. Until now the gap between frontier and open weights was set by compute and data. It is now set by security review as well, and a lab cannot buy more security review the way it buys more chips. A frontier that pauses itself for two weeks is a frontier whose release schedule is partly outside its own control.
For a buyer, that is intelligence-supplier risk, and it belongs on the purchasing checklist the same way cloud concentration risk did a decade ago. If the model an operation runs on can be paused, repriced or withdrawn on a timescale of weeks, the operation has to be able to run on another one. Mark wrote about the single-provider trap in April, when it was a forecast. This summer it stopped being one.
Open weights are now a real option. Design every operation to run on more than one supplier.
Lane three: the market now quotes its own waste, and the savings sit in the software
Falling list prices are the small story in token economics. The big story is waste, and the waste is architectural: what gets re-sent on every call, what gets routed to a model bigger than the job, and what keeps being a model call when it should have become code. This summer the vendors started publishing that waste themselves.
Glean reported 81% lower token cost than Claude Cowork across more than 180 enterprise tasks, on its own benchmark. Writer’s rebuilt harness, the software that runs the model and holds the workflow around it, cut token spend by 38% with accuracy flat to up. McKinsey found per-token prices falling while tokens consumed “increased even faster,” and 20% of firms now say AI operating cost constrains their use of it. Then, on September 1st, Anthropic launched Fable 5.1. The list price did not move. What got cheaper, by 75%, was re-reading context the model had already seen. Anthropic’s own estimate is that this makes typical workloads about 25% cheaper and highly agentic work about 45% cheaper, with the model itself pitched at long unattended runs.
Look at what those four have in common. Glean and Writer are describing harness and context engineering, not a change of model. And even the lab’s price cut is a discount on re-sent context, which is to say a discount on waste. Nobody on this list saved money by picking a cheaper model. They saved it in the software around the model, by deciding what gets re-sent, what gets routed where, and what stops being a model call altogether.
We can put our own number beside theirs. On one live job we watched the token count fall from one million to three hundred thousand and then to twenty thousand. Nothing exotic happened. The first run through an operation is discovery, and discovery is expensive. Once the repeatable decisions from that run were promoted to ordinary code, tested and reviewed like any other code, they stopped consuming tokens, and the model’s job shrank from doing the work to checking it and handling the occasional call out to another system. As a worked example of what that does to a bill: an operation that makes 20,000 model calls and costs $3 falls to about $0.40 once its repeatable decisions are code.
The point is not that every job should chase a fifty-fold reduction. It is that the unit of account has to change. Nobody buys tokens. They buy a completed operation. Until cost is measured per operation, a 25% price cut and an 81% harness improvement look like they belong in the same conversation, and they do not.
Measure cost per completed operation. Then promote the repeatable parts to code.
Lane four: long-horizon models take the planning, the harness keeps the rest
Model size was always the wrong way to measure capability. The right measure is horizon: how long a task a model can carry through on its own. It now has a number.
METR, the independent evaluation group, puts the task-completion time horizon of Claude Mythos Preview at 16 hours. That is the length of expert task the model completes half the time. At an 80% success rate, the horizon is three hours. At the reliability an operations team would actually accept, it is shorter again. The same model, measured three ways, describes three different colleagues: one who can be handed a two-day project and will finish it on a coin flip, one who can be trusted with an afternoon, and one who needs checking more often than that.
The 16-hour number deserves the attention it is getting. Long-horizon models now plan, delegate, judge and recover in ways that were not available a year ago, and Sierra’s Horizon, launched in July, is an early product built entirely around agents that pursue goals over days, weeks, or months. But the gap between 16 hours and three is the whole story for anyone running operations rather than demos.
There is a piece of arithmetic worth doing out loud. An agent that owns every step of a workflow, and gets each step right 99% of the time, completes a 100-step workflow 37% of the time. That is not a criticism of any model. It is what compounding does to any step that can go wrong, however rarely. The only way past it is to take steps out of the model’s hands altogether, which means software has to keep what models cannot: where each job stands, who did what, the ability to replay it, and checks that give the same answer every time.
That is the same architecture that fell out of the token economics in lane three, and it is not a coincidence. The repeatable decisions in a workflow become code, which does not roll the dice at all. The repeatable work that still needs a model, reading a document, classifying a request, pulling out a field, goes to small models in the two-to-four billion parameter range, because a repeatable task does not need a frontier model and cannot afford one at scale. The frontier is kept for judgment: the decision that is new, the exception that needs reasoning, the point where the plan has to change. The long-horizon model takes the planning. The software keeps state, permissions and proof.
A 16-hour horizon is a planning capability. Reliability still comes from the software around it.
What the four lanes have in common
Set the lanes side by side and one pattern shows through. In each, the market produced a product that looks like the answer and is, on inspection, the answer only for the vendor that sells it or the demo that shows it off. A context layer welded to one system. A frontier model that can be paused by its own security review. A price cut worth a fraction of what the harness saves. A horizon measured at a success rate no operations team would accept.
None of that is a reason for pessimism. Each lane moved in the right direction this summer, and faster than most people expected. It is a reason to be precise about where durable value sits. It sits in the context layer the firm owns rather than rents. In the ability to run an operation on more than one supplier. In the software that turns discovery into code. And in the harness that keeps state and proof while the model plans. Those are the four things the vendors are circling, and none of them can sell any of it whole.
The lanes will keep moving. Where the value sits will not.