AI cost management for engineering teams in 2026

Uber went viral earlier this year when it burned through its entire 2026 AI coding tools budget in just four months. Uber president and COO Andrew Macdonald put it plainly on the Rapid Response podcast: "AI is not free. Somebody's paying the bill."

Smaller versions of these surprise bills are landing on engineering leaders everywhere.

AI spend has become one of the largest line items in engineering budgets, right alongside observability and cloud infrastructure. But unlike those categories, most teams haven't figured out how to manage it yet.

And cheaper models won't fix it. Gartner projects inference costs will fall 90% by 2030, but agentic workflows consume so many more tokens per task that enterprise AI bills keep climbing anyway.

It’s a complicated and never-ending problem, but there are steps you can take right away.

Meta's $221 million month shows token usage doesn’t equal ROI

Meta employees burned 73.7 trillion tokens in a single month. That's roughly $221 million at list price – $2.65 billion a year, or the salary bill for about 9,000 engineers at $300K each. CTO Andrew Bosworth's warning to staff sums up the whole problem: "All motion is not progress and token usage alone is not a measure of impact of any kind."

This triggered Meta to move to a centralized AI gateway with real-time spend tracking and formal token budgets, and started pushing engineers toward its internal MetaCode assistant instead of Claude Code.

These bills surprise individual engineers too. One developer burned 1.1 billion input tokens in a single month and published what he learned about where the waste hides.

A lot of this overspending comes from tokenmaxxing – an initial trend to use AI as much as possible and make token usage the measure of success. It’s not surprising that metric backfired.

Here are some of the ways an influx of AI-generated code makes it harder to track ROI. 

The hard problem: ROI, invisible costs, and AI-generated code

Say you do everything right on the tool spend side – routing, budgets, dashboards. Those invoices still only show part of the cost. A budget cap without a quality gate just moves the bill from the AI line item to the maintenance backlog.

Chamath Palihapitiya shared hard numbers from his own company, 8090: "token costs are doubling roughly every 45 days, and the incremental productivity we get from that doubling is maybe 5-10% at most."

AI-generated code carries bills that land months later, in other budgets:

• AI code review – LinearB's 2026 Software Engineering Benchmarks Report, covering 8.1 million pull requests across 4,800 teams, found high-AI-adoption teams merging 98% more PRs – but review time up 91%, and AI-authored PRs waiting 4.6x longer to get picked up.
• Code maintenance – GitClear's analysis of 623 million code changes found the share of changed lines cleaned up during a commit fell from 21% in 2022 to 3.8% in 2026, and code revisited more than a year after it was written dropped 74%. AI workflows optimize for a closed ticket, not what the codebase costs to own in year three
• Comprehension debt – Google's Addy Osmani's term for the gap between how much code exists and how much of it anyone understands. In an Anthropic study he cites, engineers who used AI to write code scored 17% lower on a follow-up comprehension quiz. His warning: "a junior engineer can now generate code faster than a senior engineer can critically audit it."

Back on Rapid Response, Macdonald admitted what most engineering leaders won't: "Maybe implicitly there's more that is getting shipped, but it's very hard to draw a line between one of those stats and 'Okay now we're actually producing like 25% more useful consumer features.'"

Here are some tangible steps engineering leaders can take to manage AI costs so far. 

How engineering teams control AI costs

Gergely Orosz at The Pragmatic Engineer surveyed engineering leaders across mid-size and large companies, and a few clear approaches are separating the thoughtful orgs from the ones reactively throttling their developers.

• Separate the tool budget from the ROI conversation. AI coding tools are infrastructure. You wouldn't defund your CI/CD pipeline because you can't directly attribute revenue to it. Treat the tool budget as a cost-of-engineering-operations decision, and have the ROI conversation separately, with cleaner metrics.
• Audit before you control. Before setting limits, understand where the spend is actually going. API calls vs. subscriptions. Which tools, which teams, which use cases. Most orgs don't have this visibility. Get it first.
• Route by task, not by team. Not every prompt needs a frontier model. The CTOs Gergely Orosz surveyed at The Pragmatic Engineer are implementing smart routing – frontier models for complex reasoning and architecture decisions, cheaper models for autocomplete and boilerplate. Blanket cheaper-model policies punish your best users and push them to work around the rules; task-based routing, ideally at the platform level, applies the right model to the right work automatically.
• Per-engineer budgets that reward efficiency, not consumption. DoorDash gives every developer a high monthly token limit. Exceed it, and you justify why and commit to an efficiency plan for next month. Orosz predicts efficiency shows up in performance reviews next cycle anyway, the way cloud cost optimization did two years ago. But learn from Meta and Uber – both gamified raw usage with leaderboards and got runaway bills.
• Degrade instead of cutting off. One of the largest retirement-savings companies in the US downgrades engineers to cheaper models for the rest of the billing cycle once they hit a monthly cap. Keep expensive models for the work that justifies them.
• Subscription arbitrage for smaller teams. Startups with tight margins subsidize Claude Code Max or Codex Max subscriptions for devs instead of running API calls. Flat monthly cost. No surprise invoices.
• Knowledge-sharing over mandates. The orgs seeing results run regular internal sessions on efficient AI use – engineers sharing what's working. DoorDash has made this part of the culture. It's a faster path to savings than governance.

 Of course, each of these approaches changes depending on your toolset. 

Claude Code, OpenCode, Copilot: the cost picture across tools

Tool comparisons only make sense once you see the cost models underneath them. Every provider sells some mix of four:

• Flat per-seat subscriptions – fixed monthly fee with a usage allowance baked in (Claude Pro/Max, Claude Team, Codex Max). Predictable; hit the cap and you throttle or wait
• Seat plus metered usage – low seat fee, then pay per token (Claude Enterprise at $20/seat + usage, Copilot Business/Enterprise with pooled AI Credits). Governance dashboards included; the bill scales with usage
• Pure API – no seats, pay per token. Maximum flexibility, maximum exposure. You build your own guardrails
• At-cost and open-source inference – OpenCode Zen, open-weight models. Cheapest per token, with a capability ceiling

The trend line matters more than any single plan: the market is converging on metered. GitHub moved Copilot from flat premium requests to usage-based AI Credits on June 1. Anthropic is shifting agent-heavy usage toward metered pricing. Flat-rate subscriptions still exist, but treat them as an arbitrage window, not a permanent strategy.

Here's where each tool sits:

Open-weight models just changed the build vs buy math

Chamath Palihapitiya describes what the AI overspend looks like from the CFO seat: a Fortune 500 quietly compounds AI spend inside OpEx, misses a quarter by a few pennies of EPS, and traces the money "back to the $56-per-million tokens of intelligence, when a $0.50 version could have done the job."

His case for taking cheap models seriously: "The cheap models are now 80% to 95% as good as the frontier models on most tasks. If Pepsi is a fraction of the price of Coke and tastes almost the same in the dark, you at least have to manage that as a risk."

This all means it’s worth considering open-weight AI models that can do the job you need at a fraction of the cost of Anthropic or OpenAI models.

Open-weight models hit frontier performance this summer:

• Moonshot's Kimi K3 took first place on Frontend Code Arena, ahead of Claude Fable 5 and GPT-5.6 Sol – at $3/$15 per million tokens against Fable 5's $10/$50.
• Qwen 3.8, DeepSeek-V4, and Thinking Machines' Inkling are pushing the same direction
• The UK's AI Security Institute puts open-weight models 4–7 months behind the frontier, down from 6–10 months a year ago

Bridgewater's AIA Labs used Thinking Machines' Tinker to tune an open model for financial document triage: 84.7% accuracy against 78.2% for the best frontier model tested, at roughly one-fourteenth the cost per task. Prime Intellect post-trained a model for Ramp that beat Claude Opus at spreadsheet search and ran 27% faster than Haiku, at far lower cost.

Replit CEO Amjad Masad expects this to become the standard enterprise play: use frontier models to figure out a workflow, then distill the known use case into a smaller model you own.

So should you build on open weights? For most teams, buying is still the right call to start. Frontier models and the agent harnesses inside Claude, Codex, and Copilot deliver value across a company faster than any internal build.

Building starts to make sense when you have:

• High-volume, well-understood workflows where a tuned open model matches frontier output at a fraction of the cost
• Specific workflows you want to own end-to-end
• AI automations you already know you need to scale
• Irreplaceable market IP that you don’t want to leak

And the tooling gets better by the month. Palihapitiya's advice applies either way: "Measure the output you get, and treat intelligence like any other input cost."

Want help with AI cost management?

None of this is easy. You're managing a spend category and technology that didn't exist two years ago. Shifting capabilities, billing models that change quarterly, tools your engineers have strong opinions about, and ROI nobody has fully figured out how to measure – including the companies spending the most.

This is work we do at Codingscape every week: we help engineering leaders audit their AI spend and pick the right cost models, set up Claude and Copilot for executive teams, and build custom internal AI tools on open weights when building is the right call.

If you want a guide who's already working deep in the intricacies of managing AI costs on engineering teams, let's talk.