AI Models Right Now: What It Does, What It Costs, and When It's Worth It

The $37B Question: Everyone's Spending, Almost Nobody Can Prove ROI

Here's the uncomfortable math: enterprises will spend $37 billion on AI in 2025 — a 3.2x jump from $11.5B in 2024 — yet when researchers examined 250 public enterprise AI case studies, only 10 of them (4%) paired an observed operational result with explicit monetized ROI evidence, according to an analysis of enterprise AI case studies.

That gap is the single most important thing to understand before you pick a model or sign a contract. The operational side of the story is actually well documented: 90% of those case studies published some quantified outcome. Clinicians using AI scribes cut documentation time by more than half. Coding teams report 15%+ velocity gains. But "our team saves two hours a week" and "this project returned cash" are very different claims — and the second one is almost never backed by audited numbers.

The industry itself is split on what this means. Menlo Ventures' 2025 enterprise AI report frames the market as a genuine boom, pointing to broad adoption, real revenue, and productivity gains at scale. Yet the same report cites an MIT study claiming 95% of generative AI initiatives fail. Both can't be the whole truth — and honest readers should treat this as an unresolved debate, not a settled fact.

The reconciliation matters more than the argument. The 250-case study is explicit about it: the low 4% figure "does not mean only 4% of enterprise AI creates value." It means the public evidence base is strong on time saved, throughput improved, and quality raised — and weak on realized financial return. As that study puts it: pipeline is not revenue, estimated savings are not audited cash savings, and faster content creation does not automatically mean incremental profit.

• Operational gains (hours saved, output quality) are widely reported and relatively easy to measure.
• Monetized ROI (payback, breakeven, audited cash savings) appears in only a small fraction of public cases.
• The strongest ROI pattern wasn't tied to any model or vendor — it was tied to the structure of the work itself.

That last point is the thread this article pulls on. The spending boom is real, the hype is real, and the evidence gap is real. What's missing for most buyers isn't another vendor pitch — it's a practical way to answer three questions: what does a model actually do, what does it truly cost once you're in production, and under what conditions does it pay for itself.

One note before we dig in: this evidence problem extends beyond models to the ecosystem around them. Agent skills — the instruction packages people bolt onto AI agents — are proliferating fast, and published research found roughly 26% of them carry at least one vulnerability. That's why we're building SkillzBench, a free, independent directory that will verify and benchmark agent skills before you deploy them. Proof, not promises, is the standard worth demanding everywhere in AI.

What Models Actually Do — and the Pricing Tiers Behind the Hype

The gap between AI's hype and its price tag is enormous — and understanding both is the difference between a smart deployment and an expensive experiment. Independent benchmarking by Artificial Analysis tracks 134 models, and the spread is striking: top proprietary models score around 58 on the intelligence index, while the best open-weights models land at 44–46. Speed varies just as wildly, from a blistering 1,407 tokens per second down to far slower reasoning models that trade latency for depth.

What does that mean in dollars? API pricing has stratified into three recognizable tiers, according to a pricing comparison of major providers:

• Budget tier (~$0.10–$0.30 per 1M input tokens) — fast, cheap models suited to high-volume, routine work like classification, extraction, and simple chat.
• Mid-range ($1–$3 input) — the workhorse tier where most enterprise workloads probably belong.
• Flagship ($5–$10 input) — reserved for genuinely complex, high-stakes reasoning where the quality gap justifies the premium.

Two pricing mechanics quietly dominate your actual bill. First, output tokens consistently cost 3–5x more than input tokens — so a model that writes long responses can cost far more than its headline rate suggests. Second, cached input can cut costs by 50–90% on repeated system prompts, with major providers charging as little as 5–10% of the input price on cache hits.

Here's the part most pricing comparisons miss: token price isn't your real cost. Artificial Analysis explicitly cautions that its figures show pay-per-token API costs only — not infrastructure, engineering, or supervision costs. As one enterprise leader told an a16z survey, prompts tuned for one provider mean "changing models is now a task that can take a lot of engineering time." And Microsoft's architecture guidance notes that real cost depends on how your workload's usage patterns interact with model characteristics, not on simple category comparisons.

The practical takeaway: match the tier to the task, not the brand. Most routine work doesn't need flagship pricing, and audit your token mix — input, output, cache, reasoning — before concluding any model is "too expensive." If you're extending models into agent workflows, the same discipline applies to the skills you bolt on: SkillzBench is building a free directory and MCP connector that will let you verify and benchmark agent skills before they touch your bill.

The ROI Rule: It's the Structure of the Work, Not the Model

Here's a counterintuitive finding from an analysis of 250 enterprise AI deployments: the wins didn't cluster around a particular model or vendor. According to the study, "the strongest pattern was not tied to one model or vendor — it was tied to the structure of the work."

That structure has four markers, and they're worth memorizing before you pick a model at all:

• A clear unit of work — a discrete task you can count, like a support chat or a code review
• A measurable baseline — you know how long or how costly the task is today
• Enough repetition to generate telemetry — volume that lets you detect real improvement, not noise
• An output that can be checked against an objective standard — correctness isn't a matter of taste

The best-evidenced wins fit this profile almost perfectly. Coding accounts for $4B of the $7.3B in departmental AI spend, with teams reporting 15%+ velocity gains — a task with clear units (commits, tickets), strong baselines, and objective tests. Healthcare ambient scribes cut documentation time by more than 50% in a domain where clinicians spend roughly an hour documenting per five hours of care. And Nubank's AI assistant handles over 2 million chats a month, resolving about 55% of Tier 1 inquiries — high-volume, objectively verifiable work.

So how do you judge whether your deployment is actually paying off? Use the three-result framework from the ROI methodology: capacity released, cash impact, and service or risk improvement. The critical discipline is that "an unchanged salary bill does not fall because a task takes fewer minutes" — released hours only count as savings when there's a defensible financial use for them.

That leads to the study's sharpest warning: pipeline is not revenue. Salesforce's 30,000+ internal leads and $20M pipeline lift sound impressive, but estimated savings are not audited cash savings, and faster content creation doesn't automatically mean incremental profit. This matters because only 4% of the 250 public cases combined an operational result with explicit monetized evidence — the rest reported time savings and throughput without the cash proof.

Finally, stress-test your assumptions before committing. In the ROI framework's worked example — 3-year ROI of 50%, NPV of €54,478, payback at 24 months — the downside scenario flipped NPV to negative €35,214. If your business case only survives under optimistic assumptions, it isn't a business case.

One practical note: if you're extending these workflows into agents, the same "verify before you count" logic applies to the skills you plug in. SkillzBench is building a free directory that security-scans and benchmarks agent skills before deployment, because published research found roughly 26% of skills carry at least one vulnerability. Structure your work well, verify your components, and the model choice becomes the easy part.

How to Choose: Match the Tier to the Task, Then Squeeze the Token Mix

The wrong model in the wrong seat quietly destroys your AI budget — while the right routing strategy cuts it before you even negotiate. Start with task fit, not brand loyalty.

Microsoft's architecture guidance lays out a clear filter sequence: match the model to the workload first, then weigh cost, context window, and compliance requirements (Microsoft's model-selection framework). Agentic workloads place extra weight on reasoning quality, latency predictability, and orchestration support — so a cheap model that stutters mid-chain can cost more than a pricier one that finishes cleanly.

Budget tiers run roughly $0.10–$0.30 per million input tokens; flagships command $5–$10, and output tokens cost 3–5x input across the board (current API pricing breakdown). That spread is why 37% of enterprises now run 5+ models side by side, with 76% buying rather than building — the multi-model stack is the new default, not an edge case (a16z enterprise AI survey).

Before you blame a provider for a bloated invoice, audit the hidden levers that move real cost more than headline price:

• Caching — cache hits on repeated system prompts cut costs 50–90%, and hit rates vary by provider routing, which silently changes your effective per-token price.
• Token mix — real spend depends on the blend of input, cached, reasoning, and output tokens. A model with a low headline rate can lose on output-heavy workloads.
• Switching costs — one enterprise leader put it plainly: "all the prompts have been tuned for OpenAI," and re-tuning them for a new provider is real engineering time.

Reserve flagship models for complex, high-risk work where reasoning depth pays for itself. Route high-volume routine tasks — classification, extraction, ticket triage, code review passes — to budget tiers, where a hundred daily reviews can cost a few dollars a day against meaningful developer hours saved.

Price-to-performance has become the actual differentiator as model quality converges, and API prices tend to drop every 3–6 months. Re-evaluate your routing quarterly rather than annually; the efficiency frontier moves fast enough that last quarter's optimal mix is already stale.

If your routing includes agent skills to extend what each model can do, verification matters as much as selection — published research finds roughly 26% of skills carry at least one vulnerability. SkillzBench is building a free, independent directory that will security-scan, license-check, and benchmark skills in a sandbox before you deploy them, so tier selection and capability selection both stay evidence-driven.

ctaText: Browse verified agent skills on SkillzBench socialProofText: Every skill security-scanned, license-checked, and benchmarked with a published Lift score — free to explore on SkillzBench.

Going Further: Extending Your Model Investment with Verified Agent Skills

Agent platforms capture only $750M of the $8.4B horizontal AI market, while copilots command $7.2B — a reminder that agent workflows remain the frontier, not the mainstream. That gap is exactly why extending your model investment into agent skills deserves scrutiny before enthusiasm.

Agent skills are lightweight instruction packages that teach tools like Claude Code, Codex, Cursor, and Gemini CLI how to perform specific tasks. They are cheap to add and easy to distribute, which is precisely what makes them risky: published research cited by SkillzBench found that roughly 26% of skills carry at least one vulnerability. Extending your stack with unvetted skills can quietly widen your attack surface.

The diligence burden falls on you, and it is heavier than most teams expect:

• Provenance: Which commit produced this skill, and has it changed since?
• Security: Does it execute code, call external services, or request credentials?
• Licensing: Can you legally use it in a commercial workflow?
• Actual value: Does it improve output, or just add tokens and latency?

That last question matters most. As models converge in quality, enterprise leaders surveyed by a16z report that pricing and measurable performance — not brand — now drive model choice. The same discipline should apply to the skills you bolt onto those models.

SkillzBench (skillzbench.com) is building a free, independent answer. Its directory continuously crawls and dedupes skills from GitHub, pins every version to a commit SHA, runs security scans and license checks, and assigns a trust score from security, quality, license, maintenance, and community signals. Every skill is then benchmarked in a sandbox with the same model, with and without the skill, publishing the Lift in percentage points — always shown separately from ratings. A free MCP connector will let any agent load only the skills you choose, and everything stays free: revenue comes from blog advertising, and money never touches scores or listings.

Skills, proven — join the waitlist to be first in when SkillzBench launches.