The frontier model field guide: what you actually pay per million tokens

A procurement lead at a mid-size insurer asked me last month which model her team should standardise on. I told her the question was wrong, and she was visibly annoyed, so let me put it more usefully here: in March 2026 the top five models are close enough on quality for most enterprise work that the deciding variables are price per completed task, tail latency, and how badly you get hurt when you want to leave.

Start with price, because it is the one number people quote and the one they most often misread. List rates are per million tokens, but nobody consumes a million tokens in a shape that matches the price sheet. A retrieval-augmented support flow I profiled at a logistics firm averaged 8,400 input tokens and 190 output tokens per answer. On a model at $3 in / $15 out, that is roughly 2.8 cents per answer — and the output rate, the scary-looking one, contributes barely a fifth of it. Flip to a code-review agent that emits 3,000-token patches and the arithmetic inverts entirely.

The table everyone asks me for

List prices below were checked on 3 March 2026 and will be stale by the time you renew. Verify with the vendor. Context figures are the standard published maximum, not the beta tier you have to email a salesperson to unlock.

Where the spread actually shows up

On my twelve-task suite, the difference between the most and least accurate frontier model was nine percentage points. The difference in cost per completed task was 24×. That ratio is the whole story of 2026 procurement. If your workload is invoice line-item extraction, a small model at $0.25 in / $2.00 out with a strict JSON schema and one retry beats a flagship on unit economics and loses nothing you can measure.

Latency is the variable most teams under-test. Median time-to-first-token on shared endpoints ran between 380 ms and 1.4 s across the nine models. But p95 is where products break: two vendors showed p95 first-token above five seconds during European business hours, which is invisible in a batch job and fatal in a chat widget. If you are building interactive, test at 10 a.m. on a Tuesday, not at midnight.

Reasoning modes complicate the sheet further. Extended-thinking settings can multiply output tokens by six to ten on hard problems, and those thinking tokens bill at the output rate. I watched a team’s monthly spend triple in eleven days because someone enabled high reasoning effort globally rather than per-route. Set it per endpoint, log the token split, and alert on it.

The pros and cons nobody prints

Closed frontier APIs give you the best raw capability, prompt caching that genuinely halves repeat-context costs, and a support contact who answers. They also give you a unilateral deprecation calendar. I have now seen three teams forced into unplanned re-evaluation because a model version they had pinned and evaluated was retired with roughly six months’ notice. Budget an eval re-run per year per critical route; it is a real cost line.

Open-weight models — Llama, Mistral, Qwen, DeepSeek — buy you price stability, data residency and the right to keep running a version forever. The bill moves from tokens to GPUs and engineers. My rough crossover, from four deployments where I saw the real numbers: self-hosting starts to make sense somewhere north of roughly 400 million tokens a month of steady, predictable load, assuming you already employ someone who can debug a vLLM deployment at 3 a.m. Below that, you are paying a salary to save a subscription.

What I would do on Monday

Route by task, not by vendor loyalty. Put a gateway in front of everything so switching is a config change rather than a sprint. Cache aggressively — prompt caching cut one client’s input spend by 61 per cent for a two-line change. Keep a frozen golden set of 200 real examples with human-graded answers, and re-run it every time anything changes: model version, prompt, retrieval chunking, temperature. And write down your cost per completed task somewhere a finance person can see it. Three of the eleven teams I reviewed could do that. Those three were also the only ones whose AI budget survived the January planning round intact.

My personal default at filing: Sonnet-class for anything agentic, a mini/flash-class model for classification and routing, an Opus-class model reserved for the 5 per cent of requests a cheaper model has already failed. That ladder costs about a third of running the flagship everywhere and I have yet to see it lose a user-visible quality argument.