On this page · 12 sections
- The headline most people are getting wrong
- What actually happened at Microsoft
- What actually happened at Uber
- The pricing change that triggered both stories
- Why this is fundamentally an ROI problem, not a pricing problem
- The Microsoft–Anthropic paradox, explained
- Anthropic's side of the story
- What CFOs and engineering VPs should actually do
- What this means for Indian and global enterprises
- Frequently asked questions
- A short closing note
- References
Two stories got tangled together in May 2026. Microsoft's Experiences & Devices division announced it would stop using Anthropic's Claude Code internally by 30 June 2026, moving developers to GitHub Copilot CLI after token spend blew through the team's annual AI budget. The same week, Uber's COO told Fortune the company had burned through its planned 2026 AI coding budget in four months on Claude Code usage. Neither company "canceled Claude." Microsoft simultaneously joined Nvidia in a $5 billion investment in Anthropic and made Claude available across Microsoft Foundry on Azure. The story is not a retreat. It is the bill arriving for a new pricing model that most enterprise procurement teams were not ready for.
By Manu Shukla, Founder, eCorpIT. Last updated 28 May 2026.
The headline most people are getting wrong
"Microsoft and Uber canceled Claude" makes a clean tweet. It is also wrong, and the wrongness matters because the actual story is more useful.
What actually happened is two things at once. One internal-tooling cost decision at a specific Microsoft division. One budget blowout at one Uber engineering organisation. Both involve Claude Code, Anthropic's developer command-line tool, not Claude the consumer or API product. Both were triggered by the same root cause: token-based pricing meeting unexpectedly high engineer usage.
Underneath those two stories sits a corporate-level Microsoft-Anthropic relationship that got materially deeper in the same window, not shallower. Microsoft made Claude available across Microsoft Foundry on Azure. Microsoft 365 Copilot now uses Claude for its Researcher agent. Copilot Studio lets customers build on Claude. Microsoft put real money on the table at Anthropic. And Microsoft and Anthropic are in talks for Claude inference workloads to run on Microsoft's Maia 200 AI accelerator chips. That is not a brand pulling back.
The story is what every CFO and engineering VP should be planning against right now: pay-per-token pricing has changed the unit economics of AI tooling faster than most procurement processes can keep up.
What actually happened at Microsoft
Microsoft launched a Claude Code pilot in December 2025 inside its Experiences & Devices division, the org that builds the products most non-technical people associate with Microsoft (Windows, Surface, Microsoft 365 apps). The pilot was meant to compare Claude Code against GitHub Copilot CLI on real engineering work.
A few months in, the team hit the wall every flat-licence buyer eventually hits when they move to consumption pricing. Cybernews reported that Microsoft watched token-based billing consume the team's entire annual AI budget within months. The cancellation date is 30 June 2026. Affected developers transition to GitHub Copilot CLI in the weeks before that.
A few specifics worth flagging because they are usually missed.
This is not all of Microsoft. It is one division of Microsoft. Other divisions continue to use Claude through Foundry, in Microsoft 365 Copilot, in Copilot Studio and in Excel's Agent Mode.
GitHub Copilot CLI is also a Microsoft product. The internal switch is partly an "eat your own dog food" decision. The financial pressure made the political decision easier to make.
This is not Anthropic losing Microsoft as a customer. Anthropic still has a multibillion-dollar relationship with Microsoft as an investor, a hosting partner and a co-development partner. The Claude Code cancellation is an internal tooling line item, not a vendor relationship.
If you are a CFO at a company smaller than Microsoft, you should be reading the cancellation memo, not the headline. The memo says: a company with the budget to absorb any cost decided this specific cost was not worth the marginal productivity. That is a useful prior for everyone else's decision.
What actually happened at Uber
The Uber story is bigger and more revealing. Fortune's 26 May 2026 piece is the cleanest single source.
The shape:
Uber rolled Claude Code to roughly 5,000 engineers starting in late 2025. Adoption ran far ahead of plan. Usage went from 32% of engineers in February 2026 to 84% classified as "agentic users" by March, and around 95% of engineers using AI tools monthly by April. Roughly 70% of committed code at Uber was originating from AI tools by spring. The per-engineer token spend reportedly ran between $500 and $2,000 a month, as multiple outlets covering the same engineering disclosures have detailed.
The internal incentive that made the bill spike is the part nobody wants to admit out loud. Uber ran a leaderboard ranking engineering teams by total AI tool usage. Token volume was the score. The more tokens consumed, the higher the rank. As soon as that loop closed, engineers had every reason to use Claude Code aggressively and no reason to hold back. The leaderboard worked exactly as a leaderboard works. The budget did not survive it.
Then comes the more important quote. Uber's president and COO Andrew Macdonald told Fortune, in plain language, that it is hard to draw a connection between rising Claude Code use and consumer-product improvements. His phrase, widely quoted across the coverage, was: "That link is not there yet, right?"
Read that quote carefully. He is not saying Claude Code is bad. He is not saying engineers should stop using it. He is saying the company cannot yet prove that the dollars going out are producing proportional consumer-facing value coming back. That is the unanswered question at every large enterprise running AI tooling at scale right now.
The pricing change that triggered both stories
To understand both blowouts, you have to understand how Anthropic's billing model has actually worked in 2026, because most published commentary gets the timing wrong.
Anthropic's published API pricing as of mid-2026 sets Claude Opus 4.7 at $5 input and $25 output per million tokens, Claude Sonnet 4.6 at $3 input and $15 output, and Claude Haiku 4.5 at $1 input and $5 output. Prompt caching reduces cached input cost by 90%. The Message Batches API gives a flat 50% discount on token prices in exchange for async processing inside 24 hours.
Those rates are not high by historical AI standards. The trap is that agentic tools like Claude Code use far more tokens than a chat session does. A single autonomous coding task can chain dozens of API calls, each carrying tens of thousands of tokens of context. The per-call cost looks small. The per-task cost is much larger. The per-engineer-per-month cost is what shows up on the bill.
A separate Anthropic billing change is landing in June 2026 that will make the picture even sharper. On 15 June 2026 Anthropic splits Claude Pro and Max subscriptions into two buckets. Interactive use (Claude.ai chat, Claude Code in the terminal, Claude Cowork) keeps drawing from your existing subscription limits. But the Agent SDK, the headless claude -p mode, Claude Code GitHub Actions and third-party agents move to a separate monthly credit (roughly $20 on Pro, $100 on Max 5x, $200 on Max 20x) billed at full API rates, with no rollover. Overflow falls back to API pricing or gets rejected depending on a setting.
What this does in practice is unmask the agentic spend that was previously hidden inside a flat seat fee. Companies that did not budget for true API rates on agent workloads will see those costs starting in June. That is the change Microsoft and Uber are looking at and accelerating their decisions around.
A separate detail worth knowing: the Opus 4.7 tokenizer change can produce up to 35% more tokens for the same fixed text. Every Opus 4.7 call is mechanically more expensive than the same Opus 4.5 call, even if the model output is identical.
Why this is fundamentally an ROI problem, not a pricing problem
Token costs are the visible symptom. The deeper issue is the one Macdonald put into the open.
Enterprise software has always been hard to ROI. People shrug at $40 per seat per month for a CRM that may or may not be moving deals. Token billing breaks that comfort. When you see $500 to $2,000 per engineer per month flowing to one vendor for one tool, the question "is this worth it?" is no longer rhetorical. The finance team will ask. The board will ask. And the engineering leadership will not yet have a defensible answer, because measuring AI-tooling output against business outcomes is a problem nobody has fully solved.
We see three early patterns in our client work, where we have helped CFOs and engineering VPs frame this conversation.
Pattern one. The leaderboard backfire. If you incentivise engineers on token consumption, you will get token consumption. You will not necessarily get features shipped or bugs closed. Uber's leaderboard is the cleanest public example of this. The fix is to measure on outputs (PRs merged, bugs fixed, customer-impacting features shipped) rather than inputs (tokens used).
Pattern two. The agentic surprise. Engineering teams pilot AI tooling on chat workloads, where token usage is predictable, then move to agentic workloads, where it is not. A single autonomous "fix this bug" task can issue 40 to 100 calls. The pilot cost model does not survive the rollout.
Pattern three. The "no baseline" problem. Most teams cannot tell you what they shipped in the six months before AI tooling, in metrics that the CFO trusts. Without that baseline, every conversation about AI ROI is anecdotal. The teams that have the most credible AI investment cases are the teams that started measuring developer output before they bought the tool.
None of these patterns are about Claude specifically. The same dynamics apply to GitHub Copilot, Cursor, Cody, Replit, Codeium and any agentic SaaS sold on usage. Claude Code happens to be the most visible right now because two large customers have publicly disclosed their numbers.
The Microsoft–Anthropic paradox, explained
Here is the part that confuses casual readers. How does Microsoft cancel Claude Code in one division in the same month it commits more money to Anthropic at the corporate level?
The answer is that those are different decisions made by different teams on different time horizons.
The strategic Microsoft-Anthropic relationship is about owning the AI infrastructure layer for the next decade. In November 2025, Microsoft and Nvidia put $5 billion into Anthropic as part of a broader deal where Anthropic committed $30 billion in Azure spending. In 2026 Microsoft made Claude available across Microsoft Foundry on Azure, integrated Claude into Microsoft 365 Copilot's Researcher agent, enabled Claude in Copilot Studio for custom agent development and added a Claude option to Excel's Agent Mode. As of May 2026, Anthropic and Microsoft are in talks for Claude inference workloads to run on Microsoft's Maia 200 AI accelerator chips, which would lower Anthropic's hosting costs and lock in more long-term Azure consumption.
That is a corporate bet on Anthropic being one of the two or three companies that matter in frontier AI for the foreseeable future.
The Experiences & Devices Claude Code cancellation is about line-item cost control in one division for one developer-tooling product. It does not contradict the corporate bet. Microsoft's E&D team using GitHub Copilot CLI does not change the fact that Microsoft Foundry sells Claude to other customers, that Microsoft 365 Copilot uses Claude under the hood, or that Microsoft owns equity in Anthropic.
This is normal big-company behaviour and people unfamiliar with it routinely misread it. AWS has competed with Snowflake while hosting Snowflake. Google has competed with OpenAI while paying OpenAI for partnership terms. Microsoft has, for years, simultaneously partnered with and competed with the same companies depending on the layer. Treating any one of those decisions as definitional is a mistake.
Anthropic's side of the story
It is easy to read Microsoft and Uber and conclude Anthropic is hurting. The numbers say the opposite.
Anthropic reportedly hit a $30 billion revenue run-rate in 2026, after what the company's leadership described as "crazy" 80x growth. The June 15 subscription split is widely read as a margin-protection move at exactly the moment Anthropic has the leverage to make it. If your customers are using your product more than your seat fee covers, you raise prices on the part that uses the most. That is what splitting agentic spend off subscriptions does.
Anthropic also has plenty of room to optimise unit costs over the next 18 to 24 months. The Maia 200 chip talks with Microsoft are about exactly that. So is the prompt-cache discount. So is the batch-API discount. The model price card is one lever among several.
The underrated risk for Anthropic is enterprise sales-cycle damage. CFOs who saw their counterparts at Microsoft and Uber publicly walk into a budget wall are going to ask harder questions in 2026 H2 procurement than they did in 2025 H2. Pricing transparency, predictability and cost-cap mechanisms will become standard procurement requirements faster than they otherwise would have.
What CFOs and engineering VPs should actually do
A few concrete moves, in roughly the order we recommend to clients.
Cap the token bill before you launch the pilot. Set a per-engineer monthly budget at the API level (Anthropic's "usage credits" toggle, Cursor's spend cap, Copilot Enterprise's policy). Hard caps are the only reliable defence against token surprises. Soft caps with "review monthly" do not work.
Measure outputs, not inputs. Track PRs merged, features shipped, bugs closed, customer NPS deltas. Track them by team, both before and after AI tooling launches. If you cannot show before-and-after on outputs, you do not have an ROI story when the CFO asks.
Do not run a token leaderboard. This sounds obvious until you are sitting in a planning meeting where someone proposes it. Token volume is not a goal. Output is the goal. Reward the output.
Default to grounded, constrained usage in production agents. Agentic workflows that retrieve and answer from a known knowledge base are dramatically cheaper than agentic workflows that explore freely. Architecture choices made on day one show up on the bill on day 180.
Pre-negotiate the June 15 change. If you have an enterprise contract, talk to your Anthropic rep about how the subscription split affects your team. Many enterprise contracts have flexibility procurement does not realise.
Plan for the audit. When the bill is high enough to need explaining to the board, you will be asked to defend three numbers: total spend, output per dollar spent, and what the world looks like without the tool. Have those numbers ready before someone asks.
For the broader strategic question of whether AI coding tools are worth it, our view is straight. They are. The productivity lift on routine engineering work is real. The economics are not yet pencilling cleanly because the pricing models are mid-evolution and the measurement frameworks are immature. Both of those problems will get solved. The companies investing now with cost discipline will have a head start when they do.
Need a senior outside read on your AI tooling bill? eCorpIT runs AI-cost-and-ROI reviews for engineering organisations in India, the UK and the US. We do the token-economics math, the output-baseline build, and the procurement conversation. Talk to our team about a review.
What this means for Indian and global enterprises
The lessons translate directly to teams smaller than Microsoft and Uber, with a few adjustments.
If you are an Indian SME or a global SaaS company running a 20-engineer team, you cannot afford a $500-$2,000 per engineer monthly AI bill without a clear ROI story. The right move is to start with a hard per-engineer cap (say $100 or $150 per month), measure output impact for one quarter and then scale only if the data justifies it.
If you are a services or consulting company, the per-engineer-per-client billable hour question matters. AI tooling that lets a senior engineer ship 1.5x in the same hour is a margin-creating tool if the rate stays the same. If the tool eats the margin, it is a margin-destroying tool with extra steps. Measure both sides.
If you are an Indian outsourced engineering team selling to US or UK clients, AI tooling can be a competitive moat or a competitive liability depending on how it is priced into the client engagement. The teams that bake AI productivity into their pricing while keeping costs transparent are the ones that win the next 24 months.
Frequently asked questions
A short closing note
The Microsoft and Uber stories are not the end of enterprise Claude usage. They are the start of a much more disciplined conversation about token economics, output measurement and procurement controls that the industry needed to have anyway.
Anthropic is going to keep growing. Microsoft is going to keep selling Claude through Foundry. Uber is going to keep using AI tooling. What changes is how all of them, and the rest of us, measure whether the bill matches the value.
If you want a senior read on your own AI tooling bill, that is what we do.
