Meta opened a paid model API at $1.25 per million input tokens last week, and the procurement desks that stop there will miss the three numbers that decide the bill.
Meta's $1.25-per-million-input-token rate is real. It is also a single number from a four-number budget, and the other three are usually larger.
On July 9, 2026, Meta opened its first paid model API and listed Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens. The rate sits below OpenAI's and Anthropic's flagships, and it landed in the middle of the densest pricing week of the year. SpaceXAI released Grok 4.5 on July 8. OpenAI moved the GPT-5.6 family to general availability on July 9, with Sol priced at $5 input and $30 output per million tokens, Terra at half those rates, and Luna at $1 input and $6 output per million tokens. DeepSeek had made a 75% cut to V4-Pro pricing permanent in late May.
Each announcement stressed the same message: more capability per dollar. OpenAI explicitly said GPT-5.6 was trained to get more useful work from every token. The price-per-token story is genuine, and the cloud-storage-and-compute history makes the next move predictable.
Storage and compute unit prices fell for a decade. Total cloud bills still rose. Consumption, managed services, and egress did the heavy lifting, and the same pattern is now showing up in agentic systems on a much shorter clock. A single user request inside an agentic workflow can trigger planning, retrieval, multiple tool calls, validation, and retries. Each step is metered. The rate card is one term in the bill. The rest of the bill comes from how many of those steps a workflow actually consumes.
Enterprise buyers who price agents the way they priced static API calls are doing the wrong math. The four quantities that get conflated are: price per token, which is the rate card; tokens per attempt, which is how much work one workflow draws; cost per successful task, which divides the whole workflow by the number of correct finishes; and total organizational AI opex, which is what finance actually pays. A 75% drop in the first number can coexist with a rising last number, because a model cheap enough to retry also gets retried more often, and a workflow good enough to trust gets scoped wider.
The empirical work on this is early. A Microsoft Research and academic team paper on token consumption in agentic coding tasks (arXiv preprint) measures the gap between model use and task cost in coding workflows. A separate tokenomics preprint tries to quantify where tokens actually go inside agentic software engineering. Vendor engineering teams have started publishing their own read on the same mechanism; Glean's perspective on token efficiency in agentic systems is one of the more concrete practitioner takes. The demand side is being charted too: Goldman Sachs Research forecasts AI agents will boost tech cash flow as usage scales, which is consistent with bills rising even as per-token prices fall.
The honest caveat matters. Public data on production agent costs remains thin. Whether cost per successful task is rising across the enterprise, or only in a few well-instrumented coding workflows, is not yet established as a universal trend. The budgeting lesson holds either way.
The worksheet is short. Take the per-token rate. Multiply by the average tokens a single attempt burns. Multiply by the expected number of attempts before a task succeeds, retries included. Multiply by the task volume the organization actually plans to run. The result is the number the rate card never shows. A procurement desk that runs this once with the team's real workflow data, and again with the vendor's optimistic estimate, will usually find the second number is two to five times the first.
Meta's pricing is real. So is OpenAI's tiered structure and DeepSeek's permanent cut. None of them signal a falling enterprise AI bill, because the unit is getting cheaper at the same moment the workflow is getting larger. The Q3 budget will be settled by which of those two forces a buyer's worksheet is built around.