Stop Pricing AI by the Token. Price It by the Outcome
Every AI pricing conversation I’ve sat in on this year started the same way: someone pulls up a spreadsheet comparing cost per million tokens across GPT, Claude, and Gemini. They optimize for the cheapest line item. Then they wonder why their AI features feel cheap.
OpenAI just dropped a strategy document called “Building Abundant Intelligence” that articulates something most teams haven’t internalized yet: the unit economics of AI don’t work the way we’ve been measuring them. The question was never which model costs less per token. It’s how much a successful outcome costs — retries, corrections, human review, and frustration included.
The 80% Price Cut Nobody Asked the Right Questions About
In late July 2026, OpenAI cut GPT-5.6 Luna’s price by 80% — down to $0.20 per million input tokens and $1.20 per million output. Terra dropped 20%. They introduced a Fast mode for Sol that runs 2.5x faster for double the cost. On paper, this looks like a pricing war. It’s not.
CFO Sarah Friar’s framing is the interesting part. Customers don’t buy tokens for fun. They want the support ticket resolved, the code shipped, the contract reviewed. A model that nails it on the first try at $12/M output can be dramatically cheaper than a model that costs $4/M but needs five attempts, a human in the loop, and a cleanup pass.
This sounds obvious when you say it out loud. But I’ve watched teams build pricing models that completely ignore retry rates, failure-handling labor, and the compounding cost of wrong outputs. They treat token price as the whole picture because it’s the number on the pricing page.
What ARC-AGI-3 Actually Proves
Here’s the detail that stopped me cold. OpenAI’s ARC-AGI-3 benchmark score jumped from 13.3% to 38.3% — nearly tripling — and the underlying model didn’t change at all.
What changed was the scaffolding around it. Better context management. Smarter retained-reasoning compression. The model itself was identical. The system wrapping it got better at keeping relevant context, discarding noise, and presenting problems in a way the model could actually solve. And this improvement came with a 6x reduction in output tokens.
If you’re building AI products, this is the lesson that matters more than any benchmark number. The gap between a 13% success rate and a 38% success rate on hard reasoning tasks came entirely from how you frame the problem and manage the conversation. That’s not a model choice. That’s an engineering choice. And it’s one most teams underinvest in because they’re too busy comparing API pricing.
The Agentic Token Reality
OpenAI claims agentic work through Codex now accounts for 99.8% of weekly output tokens across their platform. Read that again. Nearly all the compute being consumed isn’t chat — it’s agents running in loops, writing code, testing, iterating, and producing work that may or may not be useful.
This reframes the pricing conversation entirely. When you’re paying for agents that churn through millions of tokens autonomously, the per-token rate matters a lot more than it does for a single chat interaction. But so does the success rate. An agent that burns 2 million tokens and ships working code is worth more than one that burns 500,000 tokens and produces something that needs to be rewritten.
GPT-5.6 Sol helped optimize its own serving infrastructure, cutting end-to-end costs by 20% and boosting token generation efficiency by 15% through better speculative decoding. The model improved its own runtime. That’s the kind of efficiency gain that compounds across billions of queries.
How to Actually Price Your AI Work
Stop building token-cost spreadsheets. Start building outcome-cost models. Here’s what that looks like in practice.
Track retry rates as a primary metric. If your pipeline calls the model three times for every successful output, your real cost is 3x the per-token rate — plus the latency, plus the human review overhead. A model that costs 2x per token but succeeds on the first call is cheaper.
Measure human-in-the-loop cost. How many minutes does someone spend reviewing, correcting, or rewriting AI output? At $50-100/hour fully loaded, that dwarfs API costs for most use cases. The cheapest model often generates the most expensive human labor.
Separate exploration from execution. Use cheap models for broad search — scanning documents, triaging inputs, filtering noise. Use the expensive reasoning models for the final synthesis. OpenAI’s own pricing reflects this: Luna for high-volume low-stakes work, Sol for the hard stuff.
Build for context efficiency, not just prompt quality. The ARC-AGI-3 jump came from context management, not prompt engineering. If you’re spending hours tweaking system prompts but ignoring how context accumulates across a multi-turn conversation, you’re leaving performance on the table.
The Bigger Picture
OpenAI frames this as a flywheel: better intelligence drives adoption, adoption funds investment, investment improves the next generation. ChatGPT now reportedly reaches over a billion active users and two million businesses. Users who stick around for six months send 50% more messages daily than new users.
The flywheel argument is self-serving — of course a CFO wants you to believe the spending will pay off. But the underlying observation is sound. The teams that win with AI aren’t the ones who find the cheapest model. They’re the ones who’ve figured out which tasks need which tier of intelligence, and how to measure whether the work actually got done.
The pricing pages will keep getting more confusing. More models, more tiers, more speed-cost tradeoffs. The teams that cut through the noise will be the ones tracking a different number entirely: cost per successful outcome.
That’s the only metric that pays the bills.
This analysis draws on OpenAI’s “Building Abundant Intelligence” strategy document, published August 2026.
