A leaked Accenture meeting, a tool called “caveman,” and monthly AI invoices that have tripled past $15 million: the story of enterprise AI in the summer of 2026 is no longer about who has the best model. It is about who can afford to keep the lights on. Internal Slack chats, dashboard screenshots, and emails obtained by 404 Media’s Joseph Cox and Emanuel Maiberg show large companies quietly walking back the “AI for everyone” stance they adopted twelve months ago, cutting off access to specific frontier models and pushing workers toward cheaper ones as bills spiral.
The Numbers Behind the Pullback
In one case that 404 Media reviewed, a company’s internal AI bill tripled to more than $15 million a month - and that is the floor of a documented range, not the ceiling. Adobe told staff it was ending unlimited access to Claude. Atlassian, Amazon, and Citi are among the half-dozen firms whose leaked materials show them cutting access to specific frontier models, urging workers to switch to “less powerful” alternatives, or both.
The same pattern is older than the past week. In a separate investigation published June 24, 404 Media reported that Uber had blown through its entire annual AI budget in four months, forcing the rides-and-delivery company to cap employee use of Claude Code and Cursor. GitHub responded by moving Copilot to per-token billing rather than the older flat enterprise plan. Accenture, which had required senior staff to use AI “or risk missing promotions,” discovered a less flattering detail: in a leaked internal meeting, the firm’s agentic AI strategy lead Justice Kwak said that non-engineers, not engineers, were driving the bulk of token consumption - the same kind of trivial “PDF to slides” tasks the AI vendors pitched as productivity wins.
How Workplaces Are Actually Cutting Costs
The throttling is not just procurement-side. Engineering teams inside AI labs have built and adopted a tool meant to shrink the bills from the seat level. According to 404 Media’s June 30 piece, developers at OpenAI, Nvidia, and GitHub are using an open-source plugin called “caveman” that rewrites model output so Claude Code, Codex, and Gemini respond in compressed, telegraphic prose. A senior OpenAI employee even contributed code adding Codex support to the project.
The plugin’s own repository tells the same story. JuliusBrussee/caveman, released April 4, 2026 and MIT-licensed, claims to cut roughly 75% of output tokens across Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, OpenCode, OpenClaw, and Pi. With more than 81,500 GitHub stars, it is one of the more visible pieces of evidence that cost - not capability - is now the binding constraint on enterprise AI rollouts.
That shift is visible across the stack. The same 404 Media reporting links to GitHub’s Copilot moving off flat-rate enterprise pricing and Uber’s CTO telegraphing budget exhaustion to product teams. Combined with the new 404 Media podcast episode on the “Tokenpocalypse” (hosted by Cox and Maiberg), it paints a market that has stopped subsidizing experimentation and started metering it.
What This Means
For everyday knowledge workers, the practical conclusion is uncomfortable: the seat at the best frontier model is no longer a default entitlement. Adobe and Citi’s internal rollbacks are an early read on a wider pattern. If your team relies on Claude or another top-tier model for daily work, assume budgets will be reviewed inside the quarter - and that the model you use today may not be the model you use in October.
For developers and operators, the caveman-style middleware plus per-token vendor pricing point to a near-term world where spend controls live above the model rather than inside it. That favors operators who can swap base URLs (as the caveman gateway design implies) and disfavors teams that hard-coded one provider. For local-AI readers, the same math that pushes a $15 million monthly bill down has a corollary: the off-peak cost of running an open-weight model on your own hardware looks cheaper every quarter, which is why 404 Media’s Tokenpocalypse reporting pairs naturally with the rise of local inference stacks like Ollama (now agentic, with integrations like the Ollama 0.17 release we covered in February) and llama.cpp, plus smaller fine-tunes (for example, the project page lists a “cavegemma” - a Gemma 4 31B fine-tune trained on caveman-style pairs). For a practical starting point, our self-host ChatGPT guide with Ollama and Open WebUI walks through a private, local chatbot in 15 minutes - no subscription, no data collection.
For policymakers and labor advocates, the reporting reframes AI access as an occupational safety question, not just a productivity question. When a tool becomes mandatory for career progression (the Accenture policy) and is then quietly throttled for cost, workers are caught in the middle of a budget fight that was not on the slide deck.
The Bottom Line
The “AI for everyone” promise of 2025 is colliding with the inference bill of 2026. Real leaked numbers from Uber, Accenture, Adobe, Atlassian, Amazon, and Citi show companies throttling access, switching to per-token pricing, and adopting compression middleware like caveman to keep monthly spend - in some cases north of $15 million - from breaking the budget. The next twelve months will be defined less by which model is smartest and more by which one a given enterprise is still willing to pay for at the end of the month.