The $29 Subscription That Became a $750 Bill
For years, the pitch for AI coding assistants was refreshingly simple: pay a flat monthly fee, get effectively unlimited use. $10 for Copilot Pro, $39 for Pro+, $19 or $39 per seat for Business and Enterprise. It was the easiest line item in an engineering budget to plan around, because it didn't move. That changed on June 1, 2026, when GitHub moved every Copilot plan to usage-based billing, replacing the old Premium Request Unit system with 'AI Credits' tied directly to token consumption — 1 credit per penny of underlying model cost. The listed subscription prices stayed the same. What you got for that price did not.
The backlash was immediate and specific. One developer on the $39/month Pro+ plan reported burning through 8% of their 7,000-credit monthly allotment in two hours of normal work, projecting the whole quota would be gone in under two days. Others posted side-by-side monthly cost projections showing jumps from $29 to $750, and from $50 to $3,000, once actual token usage was metered instead of averaged into a flat fee. One user described spending more than $6 on a single change request and called the consumption 'impossible to predict.' TechCrunch quoted developer reaction as blunt as 'what a joke', and The Register reported users vowing to abandon the tool entirely in favor of routing requests through OpenRouter or other third-party gateways instead.
GitHub and Microsoft did not walk the change back. And that's the more important signal buried in the story: this wasn't a pricing misstep that got corrected under pressure. It was a deliberate, structural shift — because the old flat-fee model was economically unsustainable once agentic, multi-step coding workflows became the default way developers used these tools instead of simple autocomplete. The complaints were real, but the change was permanent, and every other vendor in the category was watching.
This Wasn't Just a Copilot Problem
If Copilot's billing shock were an isolated vendor misstep, it would be a footnote. It isn't. The FinOps Foundation's State of FinOps 2026 report found that 98% of FinOps practitioners now manage AI spend as part of their remit, up from just 31% in 2024 and 63% in 2025 — AI cost management went from a niche concern to a universal one in two budget cycles. More strikingly, 73% of enterprises exceeded their AI cost projections over the past year, with a 2.4x budget overrun as the single most common outcome. That's not a rounding error in a forecast. That's most companies planning for one number and paying nearly two-and-a-half times that.
The specific incidents read like a pattern, not a coincidence. Uber's CEO publicly acknowledged the company 'blew through the entire annual AI budget in a quarter' after Claude Code reached 84% adoption across its roughly 5,000-engineer organization — by April 2026, the entire year's AI tooling budget was gone in four months. Uber's response was to cap spend at $1,500 per employee, per tool, per month. Microsoft, the company that owns Copilot, separately canceled most internal Claude Code licenses in its own Experiences and Devices division after per-engineer costs climbed to $500-$2,000 a month — a company that sells AI coding tools for a living found its own engineers' usage of a competitor's agentic tool too expensive to keep funding. And in at least one documented case, an enterprise that forgot to configure a spending cap on its AI tooling was billed $500 million in a single month.
Put together, these aren't stories about careless engineers. They're evidence that the underlying pricing model — pay-per-token, uncapped by default, scaling with however aggressively your engineers choose to use agentic features — is structurally incompatible with how finance teams plan budgets. A CFO can model a $30-per-seat SaaS line item a year out with high confidence. Nobody can model token consumption a year out with any confidence at all, because it depends on how autonomously engineers let their tools run, and that behavior itself is shifting month to month as the tools get more capable.
Key Takeaways
- 98% of FinOps practitioners now manage AI spend, up from 31% in 2024 (FinOps Foundation, State of FinOps 2026)
- 73% of enterprises exceeded their AI cost projections in the past year; 2.4x budget overrun was the modal outcome
- Uber exhausted its entire annual AI budget in four months after Claude Code hit 84% adoption across ~5,000 engineers
- Microsoft canceled internal Claude Code licenses after per-engineer costs reached $500-$2,000/month
Why 'Just Set a Spending Cap' Doesn't Actually Solve It
The obvious response to runaway metered billing is to cap it — and most organizations are now doing exactly that. Uber's $1,500-per-employee-per-tool ceiling is the visible example, but it's becoming the default enterprise pattern: set a hard budget, and let usage throttle or shut off when it's hit. The problem is that a spending cap doesn't fix unpredictability, it just relocates it from the finance team to the engineering team. Instead of a CFO being surprised by an invoice, a developer is now surprised mid-sprint when their agentic coding tool stops responding because the monthly allotment ran dry on day nineteen.
That tradeoff has real delivery consequences. Engineering teams that lean on AI agents for a meaningful share of their throughput are now exposed to a new failure mode that didn't exist under flat-fee pricing: a sprint that was tracking on schedule can stall not because of a technical blocker, but because the tooling budget ran out. Under the old Copilot billing, fallback to a lower-cost model kicked in automatically when premium quota was exhausted. Under the new system, that fallback is gone — usage is governed directly by available credits and admin-configured budget controls, full stop.
This is the part boards and CFOs tend to miss when they hear 'we capped it.' A cap controls the downside of overspend. It does nothing to make the underlying capacity predictable. You've traded a budget risk for a delivery risk, and delivery risk is much harder to quantify in a board deck than a dollar figure.
The Root Cause: Cheaper Tokens, Bigger Bills
The counterintuitive part of this story is that per-token prices have been falling fast — 60-80% year over year across frontier model providers. By the normal logic of a commodity input getting cheaper, AI tooling costs should be going down. Instead, total enterprise AI budgets are ballooning, with estimates showing typical organizational spend climbing from roughly $1.2 million to $7 million in a single year. Cheaper tokens didn't lower the bill. They made previously-uneconomical use cases viable, and usage expanded to fill the new headroom — a pattern anyone who has watched cloud compute costs over the last decade will recognize instantly.
Coding is the sharpest version of this dynamic. Autocomplete-style suggestions consume a handful of tokens per interaction. Agentic coding workflows — the kind where a tool reads a whole repository, plans a multi-file change, writes code, runs tests, reads the failures, and iterates — consume tokens on an entirely different order of magnitude, often across dozens of internal reasoning and tool-call steps for what looks to the developer like a single request. The billing model changed because usage patterns changed first: engineers stopped asking for a function and started asking for a feature, and the token cost of 'write me a feature, end to end' is not a small multiple of 'suggest the next line' — it's a different category of spend entirely.
That's the uncomfortable truth underneath the Copilot backlash: this isn't a pricing model that vendors will quietly reverse once the complaints die down. It reflects the actual, growing cost of running increasingly autonomous agents against increasingly large codebases. Vendors that tried to keep eating that cost under a flat subscription would eventually have gone bankrupt on their heaviest users. The economics were always going to catch up with the pricing.
The Board Is Now Asking Questions Your Engineering Leaders Can't Answer
Every engineering leader who has been asked 'what will our AI tooling cost next year?' in the last two months has felt the shape of this problem directly. The honest answer, right now, is: it depends on how many engineers adopt agentic workflows, how autonomously they let the agents run, how large the codebases they're pointed at are, and how aggressively a vendor decides to re-price mid-contract. None of those variables are things a CFO can plug into a spreadsheet with confidence, and boards that were comfortable approving AI tooling budgets as a rounding error in 2024 are now asking pointed questions in 2026 because the rounding error became a material line item.
This lands at exactly the moment multi-year planning matters most. Organizations locking in 2027 headcount and tooling strategy are trying to answer a question that has no stable answer: how much delivery capacity does a dollar of AI tooling spend actually buy, and will that ratio hold for the next four quarters? A dollar of Copilot spend bought a predictable amount of autocomplete assistance in 2025. It's buying a variable, workload-dependent amount of agentic capability in 2026, priced by a vendor that can and did change the terms with roughly one billing cycle of notice.
The engineering leaders handling this best aren't the ones with the biggest AI tooling budgets — they're the ones who've stopped treating AI tool spend as a fixed cost and started treating it as a variable one that needs the same forecasting discipline as cloud infrastructure spend got a decade ago. That means modeling at the 90th percentile of usage, not the average, and building contingency into planning rather than discovering the gap when the invoice arrives.
Key Takeaways
- Token prices fell 60-80% year over year, yet typical enterprise AI budgets grew from roughly $1.2M to $7M in the same period
- Cheaper tokens expanded viable use cases faster than they lowered per-use cost — the same pattern seen with cloud compute a decade earlier
- Agentic, multi-step coding workflows consume tokens at an order of magnitude above simple autocomplete
- Boards that treated AI tooling as a rounding-error budget line in 2024 are now demanding forecasting discipline in 2026
The Predictability Argument for Dedicated Nearshore Teams
None of this is an argument against using AI coding tools — the productivity gains are real, and no engineering leader serious about 2026 delivery is walking away from them. But it is an argument for rethinking what predictable engineering capacity actually looks like when one of your core tools now has a cost structure that can 10x or 50x between billing cycles depending on how your engineers use it that month. A pure AI-tooling-first strategy has quietly inherited the same forecasting problem that made cloud cost overruns notorious — except here it compounds with an active vendor renegotiating the terms underneath you.
A dedicated nearshore engineering team offers something the metered-billing model structurally cannot: a fixed, contracted monthly cost for a known quantity of senior engineering capacity, regardless of how token-hungry any given sprint turns out to be. That doesn't mean choosing between AI tools and human engineers — the strongest setups combine both, with senior engineers who use AI coding assistants as a force multiplier under governed, capped usage, rather than as the entire delivery mechanism. The predictability comes from the team structure, not from abstaining from the tooling.
This is precisely the kind of hybrid model Stepto builds with clients: senior, direct-hire engineers in Serbia working as an extension of your team, at a fixed monthly rate you can put in a budget a year out with actual confidence — who also use frontier AI coding tools daily, under sane usage governance, to move faster on the work that tooling is genuinely good at. The result is a total cost of engineering capacity that a CFO can forecast, even in a year where the AI tooling market itself is still re-pricing every few months. When the variable in your budget is 'how many senior engineers we've contracted' rather than 'how many tokens our agents burned this sprint,' 2027 planning gets a lot less speculative.
Key Takeaways
- Metered AI tooling costs can swing 10-50x month to month based on usage patterns alone — a volatility profile most finance teams have never had to model for a delivery tool
- A dedicated nearshore team provides fixed, contracted monthly capacity cost, independent of token consumption
- The strongest model is hybrid: senior engineers using AI tools as a force multiplier under capped, governed usage — not a replacement for predictable staffing
- Total cost of engineering capacity becomes forecastable when the variable is headcount, not token burn
A Practical Framework for Budgeting AI Coding Tools in 2026
Whatever mix of tooling and staffing you land on, a handful of concrete practices separate the organizations getting blindsided from the ones that aren't. Start by forecasting AI tooling spend at the 90th percentile of last quarter's actual usage, not the average — the Copilot backlash and Uber's budget blowout both happened to organizations that planned around typical usage and got hit by tail usage instead. Negotiate enterprise commit pricing with hard, per-seat usage caps and real-time alerting rather than a single account-level kill switch, so a budget ceiling doesn't take down an entire team's tooling mid-sprint the way blanket caps have.
Separate your budget line for autocomplete-style assistance from your budget line for agentic, multi-step workflows — they have entirely different cost profiles and lumping them together is exactly how organizations end up surprised. Review AI tooling spend monthly, not quarterly, given how quickly vendors have shown they'll change pricing terms with limited notice. And build your delivery model so that AI tooling access, if throttled or re-priced again, degrades your team's velocity rather than stopping it outright — which in practice means not building your core delivery capacity entirely around tools whose pricing you don't control.
None of this requires exotic tooling or a dedicated FinOps hire on day one. It requires treating AI coding tool spend with the same seriousness engineering leaders learned to apply to cloud infrastructure costs after the first round of unpleasant AWS invoices a decade ago — because the pattern, and the lesson, is the same one repeating itself.
Key Takeaways
- Forecast AI tooling spend at the 90th percentile of recent usage, not the average
- Use per-seat caps with real-time alerting instead of a single account-level kill switch
- Budget autocomplete assistance and agentic multi-step workflows as separate line items — their cost profiles don't compare
- Review AI tooling spend monthly; vendors have shown they'll change pricing terms with limited notice
The Bottom Line
The Copilot billing shock wasn't a one-off vendor misstep — it was the moment the industry's pay-per-token economics collided publicly with how engineering teams actually budget, and every signal since then (Uber's four-month budget burn, Microsoft's internal license cancellations, a 2.4x average cost overrun across enterprises) confirms it's the new normal, not an anomaly. AI coding tools aren't going away, and they shouldn't — but treating them as the entire foundation of your delivery capacity means building your budget on a number that can move 10x between billing cycles at a vendor's discretion. Engineering leaders who pair disciplined, capped AI tool usage with a fixed-cost core of senior engineering capacity are the ones who'll walk into 2027 budget planning with an actual number to defend, instead of a token-consumption forecast nobody can make stick.
Building a team in Eastern Europe?
StepTo helps European and US companies build senior-led nearshore engineering teams in Serbia. Let's talk about what your next engagement could look like.
Start a conversation