Enterprise AI bills are climbing faster than most companies anticipated, and a 20-year-old developer from Amsterdam thinks he's found a way to cut them by two-thirds.
OpenAI disclosed in late 2024 that AI agents—autonomous systems that write code, manage workflows, and handle customer inquiries—now generate 55% of all output tokens among its European Union enterprise customers. Those operations aren't cheap. According to industry research, the average organization spends over $117,000 monthly on inference and orchestration alone, and nearly half say the costs exceeded their projections. Julius Brussee, founder of a Y Combinator-backed startup called Caveman, believes the industry has been approaching the problem backward.
"It started as a joke," Brussee wrote recently. "Make Claude talk like a caveman so my AI bill stopped beating my grocery bill." The joke accumulated 111,000 stars on GitHub before it became a company.
The cost squeeze stems from how agents actually work. Unlike chatbots that respond once and move on, agents generate massive volumes of text—API calls, code blocks, multi-step reasoning chains. They re-read context repeatedly as workflows unfold across minutes or hours. A poorly configured caching strategy can multiply costs tenfold. Output tokens, the text agents write back, dominate enterprise bills, yet most companies lack real-time visibility into where those tokens come from or how to reduce them without breaking functionality.
Ninety-five percent of enterprises now run at least one agent-enabled workflow in production, according to IDC's mid-2024 survey, though only about a quarter report meaningful cost visibility, according to recent surveys from IDC and KPMG. Agent adoption surged from 69% to 86% year-over-year, and nearly a third of organizations say employees spend more than six hours weekly just overseeing these systems. Sixty percent run more than ten agents simultaneously. The infrastructure bills, as one might expect, are piling up faster than the dashboards tracking them.
The technical fixes exist but remain fragmented. Anthropic charges premium rates for cache writes—1.25 to 2 times base prices—but cache hits cost only a tenth of standard input rates. Miss the cache, and you're paying ten to twenty times more than necessary. OpenAI noted in September that one customer trimmed costs 20% simply by nudging cache hit rates from 85% to above 90%. Gartner noted in August that substantial token costs could be eliminated without quality loss through a combination of model tiering, open-weight alternatives, smarter caching, intelligent routing, and compression. The problem, analysts noted, is that enterprises lack tools to observe real workloads and deploy these fixes safely.
Adobe Research demonstrated in a June 2024 study that output-side compression—teaching models to respond more tersely—can reduce realized costs by 1.4 to 2.4 times per model, occasionally reaching threefold savings, without sacrificing quality. The same paper warned that input-side "caveman" prompts—stripping grammar and shortening user instructions—backfire, actually increasing net costs. Separate research on multi-agent routing systems showed 72% cost reductions with 97% quality retention by directing workflow sub-steps to cheaper models. Another study found that "keepalive" strategies, which maintain cache residency during idle periods, can slash post-pause request costs by up to 12.5 times.
Yet these levers underdeliver when applied piecemeal. Community benchmarks published over several months reported mixed or negative net savings in long-running coding agents when compression reduced output tokens but failed to address context bloat or cache misses. The economics require end-to-end visibility—tracking input tokens, cache writes, hits, output tokens, and reasoning tokens per agent, per workflow—plus the ability to test changes and roll back failures without disrupting production systems.
Brussee, who previously worked at Stacklink, built Caveman as a full-stack optimization platform that observes agent workloads, identifies inefficiencies, tests improvements, and deploys only the changes that pass automated quality checks. The platform combines output compression (marketed as "say less"), input compression ("read less"), cost-aware routing, and what the company calls "eval-gated deployment"—essentially, nothing ships without passing real-world tests.
The output compression achieved a mean 65% token reduction across a ten-prompt benchmark, Caveman reported, preserving code and identifiers verbatim while compressing prose. Input compression reduced tokens by 33% across full Claude coding sessions on an internal benchmark, passing all 18 quality checks. The "Auto" routing feature, launched in Caveman 4.0, delivered 54% cost savings on standard benchmarks at the same quality score while completing tasks 50% faster. A fourth component maintains prompt cache residency across idle gaps.
Elastic Labs validated the output-compression approach across eight live scenarios, reporting a 64% response-token reduction with "zero accuracy loss," preserving technical syntax and field names intact. Adobe Research's subsequent study confirmed the broader findings: output-side compression cuts realized costs substantially, but input-side shortcuts hurt both cost and quality. GreenPT, a language model API provider, integrated Caveman and related compression variants in August, offering them as simple model-ID swaps and reporting output-token reductions ranging from 22% to 87% depending on workload characteristics.
Recently Brussee released a 1.2-megabyte model called "trim-1" that decides which lines of context an agent actually needs, running at 694,000 tokens per second on one CPU core with no GPU or API call required. The company cautioned in August that end-to-end savings depend heavily on cost denominators: if input or cache writes dominate a particular bill, output compression alone won't move the needle much. Caveman Cloud, now in private preview, prices every model call per agent, builds quality checks from actual agent activity, and automatically tries cheaper variants.
The competitive landscape splits into specialized niches. Compresr, also backed by Y Combinator, offers a context-compression API for long prompts. Token Company builds token-compression models. Headroom and RTK focus on compressing terminal output, logs, and tool results before they reach the language model. LiteLLM, an open-source gateway with more than 59,000 GitHub stars, provides cost-aware auto-routing and savings dashboards, though it faced security vulnerabilities in October that allowed API-key theft. Portkey, an enterprise AI gateway acquired by Palo Alto Networks on April 30, offers semantic routing and unified billing. Cloudflare AI Gateway provides caching and cost analytics. None bundle compression, routing, caching management, and eval-gated rollout the way Caveman positions its offering.
The observability segment includes HoneyHive, which raised $7.4 million for agent evaluations, along with LangSmith and Langfuse. These platforms track token and cost breakdowns but don't auto-generate optimization proposals or test them against production-derived quality checks—a gap Caveman is trying to fill.
Three forces make this market particularly urgent right now. Agents generate far more tokens than human chat users, and OpenAI's EU data confirms they now represent the majority of enterprise output volume. Providers keep repricing: Anthropic reduced cache read costs on its latest model in September; OpenAI shipped improved caching instrumentation for GPT-6 the same month. Modal, Baseten, and Together AI all updated pricing within recent weeks. Meanwhile, enterprises have no standard playbook. McKinsey's September State of AI report noted budgets are being reallocated as agentic coding reduces traditional software purchases, pushing model and infrastructure optimization squarely into procurement decisions.
Regulatory timelines add another layer of pressure. The EU AI Act's Article 50 transparency obligations took effect in August, and the AI Omnibus entered force in July, extending risk-management and transparency requirements across agent systems. NIST launched an AI Agent Standards Initiative in February, with workstreams on agent identity, authorization, and interoperability entering public comment this year. IEEE P1968, a recommended practice for governance of autonomous AI systems, remains in progress.
Engineering leaders deploying agents should probably watch three trends, perhaps more closely than they'd like. Providers will keep repricing caching and reasoning tokens, making cost models unstable quarter to quarter. Community benchmarks show that single-lever optimizations—compression alone, or routing alone—deliver inconsistent results across real workloads; full-stack observability that attributes costs to input, cache, output, and reasoning tokens per agent is becoming necessary infrastructure. Eval-gated deployment is no longer optional. The risk of quality regression under aggressive compression or routing, visible in recent community tests, requires automated testing against production-derived checks before any optimization goes live.
Caveman's traction—110,000 GitHub stars, validation from Elastic and Adobe Research, and Y Combinator backing—suggests the market may be consolidating around platforms that observe, test, and deploy optimizations rather than forcing engineers to manually stitch together proxies, routers, and dashboards. Brussee declined to disclose funding beyond Y Combinator, and the company's YC profile lists team size as one. The platform's private-preview status and in-development components mean most features aren't yet generally available, which leaves some questions about production readiness at scale.
Still, the core claim—that agent costs can be cut 65% without performance loss—aligns with Gartner's recent forecast and with a growing body of research on compression, routing, and caching strategies. Whether enterprises will trust a one-person startup with their AI infrastructure remains an open question, but the economic incentive is hard to ignore when monthly bills average six figures and visibility remains poor.
"The models got smart enough a while ago," Brussee wrote. "Now the work is making them cheap enough to leave running."
