Enterprise Field Notes · Issue #22
Databricks Piloted OpenAI's New Model With About 200 Engineers. Their Coding Spend Rose About 60 Percent, and It Got Its Own Sub-Budget.
Then it went to about 3,500 engineers. The same month, GitHub, Atlassian, Google and Salesforce changed what AI usage costs, the part of a contract that moves after you sign.
New here? Subscribe to Enterprise Field Notes, one new issue every week.
What Databricks did
On September 16, Databricks co-founder Patrick Wendell posted that the company had rolled out OpenAI's GPT-6 Astra to every engineer, roughly 3,500 of them.
Before that, about 200 engineers got it in a pilot. Those engineers raised their overall coding spend by about 60 percent over baseline.
Astra clearly beat the company's previous top models on highly complex work. On medium and low complexity tasks it wasn't clear it did meaningfully better, and Databricks suspects existing models already do most of those about as well as anything can. Engineers now get a sub-budget for Astra inside their overall budget, to encourage using it on complex tasks and cheaper models on everyday ones.
Nobody changed a price list. Databricks measured the spend on a small group before everyone had it.
I sell AI for a living. I'm COO at Swa, an orchestration layer that sits over many AI models, including Anthropic's and OpenAI's. Pushing everyday work toward cheaper models is close to what my company sells. Databricks did it with a pilot and a budget rule. We also do better when buyers keep their options open, so check all of this against your own contracts rather than my say-so.
Here's where this goes. First, what else moved this month, and what I found when I had to go through the contracts myself. Then what CIOs are worried about, one worry at a time, and where the price catches up. At the end there's a short check to run before your next renewal.
What else moved
On September 1, GitHub Copilot's summer promotion ended, as GitHub said in April it would, and the AI credits included with an existing customer's Business seat went from $30 a month to the standard $19 at the same seat price. Annual Pro and Pro+ plans, which stay on premium requests until they expire, have also used higher model multipliers since June 1, so the same request uses more of the allowance. The same day the promotion ended, Atlassian expanded its usage-based pricing with three meters that start billing on December 3. On September 2, Google launched Gemini 3.8 Flash at an introductory price that doubles on January 1. On September 3, Salesforce introduced new editions, two of them at new seat prices, each with a pool of AI credits included.
None of it is a scandal. All four put it in writing, and the three with a start date gave weeks or months of notice. Salesforce says its new editions carry more value than the old ones, and existing customers keep their legacy pricing. If you run pricing at a software company, this is roughly how it's supposed to be done. Every one of them touched the part tied to how much you use, and Salesforce set new seat prices as well.
Seats move with headcount, which everybody sees coming. Usage moves with decisions the business makes long after signature.
Buyers already feel the renewal side of this. Telarus, which works with technology advisors, surveyed more than 500 IT buyers for its 2026-27 Tech Trends Report. 84 percent of them say they're open to switching vendors at renewal, and Telarus says that's driven by "rising renewal costs" and "purchase regret." That's IT renewals in general, not only AI.
The money was in the contracts
I've been on the buyer's side of this. At a very large retailer, I was once asked to take on enterprise backup and recovery on top of everything else I ran, and to roll new backup software out across the whole global fleet. I had never run a data protection program before. That didn't stop me. I was also told there was no budget for the support team it needed.
So I went looking, through every contract I had access to, with finance. The money was buried in one nobody was looking at and nobody really understood. We were paying a vendor to rotate backup tapes we weren't using anymore, to store huge volumes of useless tapes around the world, and to keep a global team managing all of it.
I read the contract, and then I went and visited to understand what was actually going on. Once I could see that none of that work was adding any value, I killed it. The savings were worth millions, and I only asked back enough to fund the operational team I needed.
What we were paying for was right there in the contract. Nobody had checked it against what we were actually doing.
Why a price list lands this hard
Earlier this week I caught up with an IT leader who spends a lot of time with other CIOs. Partway through, the conversation turned to what CIOs are worried about right now, and I got it in three parts.
Fear, across the board. A need for control, because the business has started running with AI ahead of IT. And the third, put to me more bluntly than I'll print it: getting hooked on AI, and then having the people who own it raise the price.
Each one usually gets answered with a purchase. Pick the vendor that feels safe, and consolidate onto as few tools as you can. Those are reasonable purchases, and I'd make some of them myself. They just don't settle anything on their own. Each worry is a question, and somebody has to decide the answer and write it down.
Fear
Some of this month's warnings came from the people building the models. On September 12, Anthropic's CEO wrote that every frontier AI company should act as if the OpenAI incident I wrote about two weeks ago had happened to them, and that the most effective way to pace the frontier is "regulation that targets all US frontier AI companies."
Regulation like that is written for a handful of companies. California's frontier law saves its heaviest duties for frontier developers with more than $500 million in revenue. None of those rules governs what your company does with AI on Monday. If you're waiting on the frontier rules to settle the fear, you're waiting on rules written for somebody else.
Closer to home, fear's first move is a retreat to the safe choice. The same IT leader told me a lot of CIOs are sticking with the AI assistant that came with the suite they already own, because it feels safer inside an interface they trust. It's a sensible instinct. It answers one question well: can we trust this tool?
The question that's actually on its way is a different one. That IT leader recently sat through an audit meeting where AI was the only subject, which had never happened before in that leader's experience. The auditors knew people were going to use AI. What they wanted to know was how the output gets validated, and they were clear that a human has to be the one validating it.
Answering that takes a check that leaves evidence, and a person who owns it. For every place AI output leaves the building or goes into a record, who checks it before it does, and did anybody write that down before the auditor asked?
Control
Control is the worry that the business is getting ahead of IT, and the reflex is to block. I wrote my very first issue about that. Blocking slows it down for a few weeks. Then it comes back, because the reason people reached for the tool hasn't gone anywhere.
Making the approved way the easy way works better. I saw that long before generative AI, at the same retailer. Getting a development team a pre-production environment with realistic data took 18 days. A team restored the data, then scrubbed it and masked it by hand and moved large amounts of it around. They had to make sure all the dependent data was there and complete, and that the finished data still had the same shape as production. Just working out what was needed took a lot of that time.
My team built a way for development teams to spin up that environment themselves in 39 minutes, masked and safe, any time, without waiting on anyone.
The masking and the checks were still there. They were built in, so doing it safely stopped being the slow way.
With AI, a lot of companies make the approved way easy by standardizing on one assistant. That still leaves the question underneath control: on any given Monday, can somebody say what's in use, and who gets told when it changes?
Where the price catches up
Consolidating is the natural response to fear and control: one contract and one bill, with governance in the box. It can buy you a better price, and sometimes a price hold.
The catch is that you lose a price you can compare. A bundle doesn't hide what you use. The four vendors above all meter it, in their own units at their own rates. You can see what you used, but not what the same work would cost anywhere else, and that's what makes it hard to move when the price changes.
Consolidating is still the right call for a lot of companies. The protection you'd normally get from having options just has to come from somewhere else, and for most companies that place is the contract.
The same gap
Take the three together and each comes down to an owner nobody has named yet. For fear, it's whoever owns the check on what AI produces. Control needs someone who can say what's in use this week. The price worry is really about who notices when usage passes what's included, and whether anybody tells the person who owns the budget.
That gap ran through my first twenty issues, and it's in the contracts too.
A check before your next renewal
This is for the next time an AI contract comes up. Start with the agreement where AI is now part of the seat, the one you'd least like to be surprised by.
The check
If your auditor asked tomorrow how this tool's output gets validated, what would you hand them, and who wrote it?
What's in use
Which teams use this tool's AI features today, and who would tell IT if another team turned them on?
The price
Pull the last 90 days of usage against what's included and carry it forward to renewal. Which month do you pass the allowance, and who sees that before the invoice does?
When usage goes past what's included, is the rate you pay fixed in your order form, or does it point to a price list the vendor can change?
Can the vendor change how many credits a request uses, and how much notice do they owe you if they do?
A fixed rate on a unit that shrinks isn't fixed.
Before the next model upgrade goes to everyone, who measures what it does to spend on a small group first, and does it get its own budget?
At renewal, what caps the increase, and does the price you have now carry past this term?
If the product lets you cap spend or set budgets, has anyone? Does the cap stop usage or only send an alert, and does the alert reach the person who owns the budget?
Any question here without a name next to it is the place to start.
When a vendor changed your AI pricing this year, who found out first, the admin or the person who owns the budget? If someone else on your team signs the next renewal, send them this check before they do.
P.S. If you sell AI usage: Telarus found 54 percent of technology advisors say their main value is saving the customer money, and 26 percent of buyers see it that way. The advisors who close that gap will be the ones who bring what it actually cost to run at month nine.
Subscribe to Enterprise Field Notes for one field report a week. Everything I've written is in my archive.
References
- Your Employees Are Already Using AI You've Never Approved. Enterprise Field Notes, Issue #1. Blocking slows shadow AI down for a few weeks; the fix is making the safe option the easy one.
- Walmart, Home Depot and KPMG Already Had What They Needed. So Do You. Enterprise Field Notes, Issue #21, September 10 2026.
- OpenAI's Agents Got Caught the Way Teenagers Get Caught. Enterprise Field Notes, Issue #20, September 3 2026.
- We Must Pace the Frontier. Dario Amodei, September 12 2026. "Act as if OAI-HF had happened to them" and "regulation that targets all US frontier AI companies."
- SB 53, Artificial intelligence models: large developers (Transparency in Frontier Artificial Intelligence Act). California, signed September 29 2025. A "large frontier developer" is a frontier developer with annual gross revenues above $500,000,000, and the frontier AI framework and catastrophic-risk duties apply to them.
- GitHub Copilot is moving to usage-based billing. GitHub, April 27 2026. Standard monthly AI credits of $19 (Business) and $39 (Enterprise), $30 and $70 for June through August; "Model multipliers will increase on June 1 (see table) for annual plan subscribers only"; and "organizations can choose whether to allow additional usage at published rates or cap spend."
- Atlassian's usage-based pricing: AI value with predictability and control. Atlassian, Tamar Yehoshua, September 1 2026. The meters in its "expanded usage-based pricing model" and "Allowance limits and billing for these meters will go into effect on December 3, 2026."
- Introducing Gemini 3.8 Flash and 3.8 Flash Cyber. Google, September 2 2026. Introductory pricing through December 31 2026, and the January 1 2027 price.
- New Salesforce Editions Bundle Everything Businesses Need for Agentic Transformation. Salesforce, September 3 2026. Core Editions at $195 and Advanced at $395 per user per month, Max at $550 with "zero pricing change"; Flex Credits included per edition; "70% more value than legacy Enterprise Edition"; and "Existing customer pricing on legacy editions remains unchanged."
- Telarus Unveils 2026-27 Tech Trends Report. Telarus, September 1 2026. More than 500 IT decision makers and more than 450 technology advisors; the 54 percent against 26 percent finding; and 84 percent of buyers open to switching vendors at renewal, "driven by rising renewal costs (87%) and purchase regret."
- Patrick Wendell on Astra at Databricks. Post on X, September 16 2026. "Today we rolled out Astra to every engineer at Databricks (N=~3500)"; "Astra unambiguously out performs our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks"; "Engineers given Astra increased overall coding spend by around 60% compared to baseline"; "It is not clear Astra meaningfully improves on medium/low complexity coding tasks... We suspect those tasks are mostly saturated"; "piloting Astra with around 200 users"; "We give engineers a sub-budget specific to Astra to encourage them to use Astra selectively on complex tasks while preferring lower cost models for everyday tasks."
- GPT-6 Astra: A new generation of intelligence. OpenAI. Names the model Wendell's post calls "Astra" as OpenAI's GPT-6 Astra.
About the author
I'm Ben. I write Enterprise Field Notes, and by day I'm COO at Swa. Before that I was Global Director of Site Reliability Engineering at a Fortune 500 retailer, running reliability, data protection, and database operations at global scale.
Header artwork generated with Swa.
Read more of Ben's Enterprise Field Notes at benpickett.com.
© 2026 Ben Pickett · Enterprise Field Notes