AI is not free labor. It is rented compute with a friendly interface. The industry has spent the last two years training buyers to look at model leaderboards, chat subscriptions, and per-token rate cards. The uncomfortable truth is that none of those numbers tell you what a workflow actually costs.
The hot example right now is the tokenizer. When Anthropic announced Claude Opus 4.7, the headline pricing looked stable: $5 per million input tokens and $25 per million output tokens, the same as Opus 4.6. But in the same release, Anthropic told developers that Opus 4.7 uses an updated tokenizer, and that the same input can map to roughly 1.0 to 1.35 times as many tokens depending on content type.
That is the whole story hiding in plain sight. If the model charges per token, and the tokenizer turns the same work into more tokens, the effective price of the work changes even when the rate card does not. This is not a conspiracy. It is math. But it is the kind of math most AI buyers are not yet set up to measure.
What a tokenizer actually does
A tokenizer is the piece of software that converts text into chunks the model can process. Those chunks are called tokens. A token may be a whole word, part of a word, a punctuation mark, a code fragment, or a strange-looking substring that only makes sense to the model’s vocabulary.
The model never sees your document as a human sees it. It sees a sequence of token IDs. Those IDs point to learned numerical representations, and the model performs its attention and neural network operations over those representations. More tokens means more work for the system to read, reason over, and generate.
This matters because AI vendors do not bill you for pages, emails, contracts, support tickets, or tasks completed. They bill by token. So a tokenizer is not just an implementation detail. It is part of your cost structure.
The Opus 4.7 lesson
Opus 4.7 is useful precisely because the change is visible. Anthropic did the responsible thing by warning developers that token usage could change. The mistake many buyers will make is stopping at “pricing remains the same” and never asking the next question: same price per token, or same price per workflow?
In production, the answer can be meaningfully different. OpenRouter analyzed real traffic from users who moved from Opus 4.6 to Opus 4.7 and found that actual costs rose 12 to 27 percent for prompts above 2,000 tokens after caching and completion-length changes were taken into account. That is the kind of increase that can hide inside a migration ticket and show up later as a finance problem.
OpenAI’s current frontier pricing tells the same broader story from a different angle. Its API pricing page lists GPT-5.5 at $5 per million input tokens and $30 per million output tokens, with GPT-5.4 at $2.50 and $15. Whether that higher tier is worth it depends on token efficiency and task quality, not on the model name. A smarter model that uses fewer retries may be cheaper in practice. A slightly smarter model that burns more tokens on a low-value workflow may be the most expensive way to do a simple job.
The model price is the sticker. The workflow cost is the invoice.
Why the job-replacement story is overdrawn
This is where the AI labor debate usually loses the plot. People compare an AI subscription to a salary and conclude that mass job replacement is inevitable. That is too simple. In the short-to-medium term, the constraint is not imagination. It is compute, reliability, and cost.
The agents that can replace meaningful white-collar work are not single-turn chatbots. They read long context, call tools, inspect files, make plans, revise outputs, recover from failures, and often ask other models to check their work. That is exactly the pattern that increases token use. Frontier autonomy is not a free feature. It is a compute strategy.
Deloitte’s 2026 compute outlook is a good sanity check here. It argues that inference is becoming the dominant AI workload and that demand for compute is still rising faster than chip efficiency. In plain English: using AI at scale is not making the infrastructure problem disappear. It is moving the bottleneck from training demos to production inference.
That is why the labor math is more stubborn than the hype suggests. A person with a laptop, judgment, and domain context is not being compared against a magical $20/month assistant forever. They are being compared against a stack of GPUs, memory, electricity, orchestration, security, evals, retries, monitoring, vendor margin, and human review. Some jobs will absolutely change. Some entry-level tasks will be compressed. But the claim that whole categories of professional labor become instantly uneconomic assumes compute abundance and permanent subsidy.
The labor data is already more nuanced than the slogan. Goldman Sachs Research recently estimated a modest net drag on U.S. payroll growth from AI, while also finding that augmentation is increasing employment in some roles. That fits the pattern I see in real businesses: the first-order effect is not “fire everyone.” It is “make the expensive people less trapped by low-value work.”
The subsidy will not last forever
Today’s AI pricing is still distorted by growth-stage behavior. Labs want adoption. Platforms want developers locked in. Consumer products want habit formation. Enterprise tools want logos. So some of the true cost of intelligence gets buried inside generous plans, bundled credits, temporary promotional pricing, or investor-funded infrastructure buildout.
That does not mean vendors are doing anything sinister. It means the market is still discovering the real clearing price of high-quality inference. When the subsidies narrow, the price signal will get louder. More usage caps. More metering. More priority pricing. More task budgets. More distinction between cheap models for routine work and expensive models for deep reasoning.
When that happens, a lot of the “AI replaces humans because it is cheaper” narrative will look naive. Humans are expensive, but they are also incredibly efficient at messy judgment under uncertain context. For many workflows, the winning system will not be AI instead of people. It will be a smaller number of people operating better AI systems with ruthless cost discipline.
Do not use a frontier model because it is impressive. Use it because the marginal quality is worth the marginal compute on that specific step.
Open-source AI becomes a cost-control skill
This is why open-source models matter. Not because every company should immediately run its own frontier lab. Most should not. The point is that businesses need options below the most expensive hosted model. They need local inference, private deployment, fine-tuned specialists, quantized models, and cheap hosted models that can handle routine work without turning every workflow into a premium-token burn.
A local open-source model running on a workstation, server, or modern AI-capable device will not beat the best hosted frontier model at hard reasoning. That is fine. It does not have to. It only has to be good enough for the right slice of work: classification, extraction, deduplication, first-pass summarization, redaction, template drafting, search expansion, routing, and low-risk internal assistance.
In other words, using a frontier model to read every inbound email is the AI version of using a cannon to kill flies. The expensive model belongs where ambiguity, risk, and upside justify it. Everywhere else, the future belongs to cheaper models, local models, and systems that know when to route up.
This is not only a technical choice. It is a job function. The next valuable AI operator will understand the difference between:
- tasks that require frontier reasoning and tasks that require consistency;
- private data that should stay local and data safe enough for a hosted API;
- prompts that should be compressed and context that should be cached;
- model upgrades that improve quality and model upgrades that only improve vibes;
- token savings that are real and token savings that create hidden failure risk.
That person is not just a prompt engineer. They are a model-routing operator, eval owner, cost accountant, and workflow designer at the same time.
The model-router mindset
Serious AI systems will increasingly look like portfolios, not monoliths. A small local model classifies the request. A cheap hosted model extracts structured fields. A deterministic validator checks the schema. A retrieval layer pulls only the documents that matter. A frontier model handles the hard judgment call. An eval suite measures whether the cheaper path is still good enough.
I wrote in Deep Models Are the Future of Enterprise AI that the winning enterprise systems will be narrow, observable, and deep in the workflow they serve. This is the cost version of the same argument. Depth is not only about quality. It is about knowing which step deserves which model.
The companies that get this right will maintain a cost ledger per workflow: tokens by user, model, route, tenant, task type, retry rate, cache hit rate, tool-call count, latency, and accepted output. That telemetry will matter for pricing, access tiers, abuse detection, and product design. It will also prevent the worst kind of AI failure: a system that looks magical in a demo and silently becomes unaffordable in production.
What teams should do now
The practical move is not to panic about tokenizers. It is to stop treating model choice as a branding decision. Every workflow that uses AI should have a few basic controls.
First, log token use at the workflow level, not just the vendor-bill level. Second, run evals before and after every model migration, because cost and quality both move. Third, keep a cheap model path for routine work. Fourth, identify which data can run locally or on a private open-source model. Fifth, define a clear escalation policy: when the cheap path is uncertain, route up to the stronger model or route out to a human.
This is the adult version of AI adoption. No magic. No vague replacement fantasy. Just measured systems that convert model capability into business value without pretending compute is free.
The real future of work
AI will replace tasks. It will reshape roles. It will make some jobs smaller and some jobs more valuable. But the short-to-medium-term story is not a clean substitution of humans with software. It is a messy repricing of intelligence.
Tokenizers are a small window into that repricing. They remind us that a model is not an oracle in the sky. It is a machine with inputs, outputs, constraints, and a bill. The people who understand that bill will be the ones who keep AI useful when the easy subsidies fade.
The future job function is not merely “use AI.” It is knowing when not to use the biggest AI available. It is understanding when open-source models on local devices are enough, when a cheaper hosted model is the right default, and when frontier intelligence is worth every token.
