Prices across AI APIs vary by roughly an order of magnitude depending on which tier you pick, and the tier matters far more than which vendor you pick. A flagship model from any major provider costs meaningfully more per million tokens than that same vendor's small, fast model — often enough that routing most traffic to a smaller tier and reserving the flagship for genuinely hard tasks is the single biggest cost lever available. Sticker price per million tokens is also only the starting point: caching, batching, and volume commitments routinely change the effective price by a large margin.
What changed in 2026
- The price gap between tiers widened, not narrowed. Vendors kept pushing small, fast models down in price while flagship models absorbed more compute for harder reasoning tasks, so the spread between a vendor's cheapest and most expensive tier stayed wide.
- Prompt caching became a default expectation, with most major providers offering a steep discount on tokens that repeat across calls — system prompts, long context documents, and tool definitions in particular.
- Batch APIs matured across the board, typically offering roughly half the real-time price for workloads that can tolerate a delay of minutes to hours instead of an instant response.
- Effective price became harder to compare on sticker rate alone, since discounts for volume commitments, regional hosting, and enterprise agreements now meaningfully change the number a small team actually pays versus the published rate.
Representative pricing by tier
Treat these as ballpark figures to reason about relative cost, not as current exact quotes — check each vendor's pricing page directly before budgeting, since rates shift often.
| Tier |
Typical input cost (per million tokens) |
Typical output cost (per million tokens) |
Best for |
| Flagship / frontier |
Roughly $10-15 |
Roughly $50-75 |
Hard reasoning, complex agent planning |
| Mid-size |
Roughly $2-4 |
Roughly $10-15 |
General chat, summarization, most production traffic |
| Small / fast |
Roughly $0.10-0.30 |
Roughly $0.40-1.00 |
High-volume classification, simple extraction, routing |
| Self-hosted open-weight |
GPU cost only, no per-token fee |
GPU cost only, no per-token fee |
Very high volume where infrastructure cost beats API fees |
Where the real savings are
- Route by task difficulty, not by default. Send classification, extraction, and formatting tasks to a small model, and reserve the flagship tier for the fraction of requests that actually need deep reasoning.
- Cache anything that repeats. A long system prompt or reference document sent on every call is close to free money left on the table if your provider offers prompt caching and you are not using it.
- Batch what can wait. Nightly summarization jobs, bulk classification, and any workload without a real-time user waiting on it belongs on a batch endpoint at roughly half the live price.
- Cap output length deliberately. Since output tokens cost several times more than input tokens, a system prompt that encourages concise answers has a direct, measurable effect on your bill.
- Revisit vendor choice at renewal, not just at launch. Relative pricing between vendors shifts over time; a comparison done a year ago may no longer reflect the cheapest option for your workload today.
Common mistakes
Benchmarking only the flagship tier. Comparing vendors solely on their most expensive model gives you a number that has little to do with what most of your production traffic should actually run on.
Ignoring output token cost. Teams optimize prompt length obsessively while a chatty model response silently costs three to five times more per token than the input that triggered it.
Never revisiting the routing decision. A cost analysis done once at launch goes stale as vendors reprice and as your own traffic mix shifts toward harder or easier tasks over time.
Treating self-hosting as automatically cheaper. Below a certain volume, GPU and operational cost for self-hosting exceeds what a hosted API would have charged — it is a real crossover point, not a given.
FAQ
Which AI API is cheapest overall?
There is no single answer — it depends entirely on which tier you compare and your specific input-to-output token ratio. Small, fast tiers from any major vendor are usually within a similar range of each other.
Why do output tokens cost more than input tokens?
Generation happens token by token in sequence, while an input prompt can be processed largely in parallel, making output meaningfully more compute-intensive per token — which is why providers price it several times higher.
Is it worth self-hosting to cut API costs?
Only past a certain volume threshold, once GPU utilization is high enough to beat the per-token API price. Below that threshold, a hosted API is usually both cheaper and less operational work.
How much does prompt caching actually save?
For workloads with a long, repeated system prompt or context document, caching can cut the effective input cost substantially on repeat calls, since cached tokens are typically billed at a small fraction of the standard input rate.
Where to go next
For the mechanics behind these numbers, see token pricing explained 2026. If you are managing calls across more than one provider, best AI API gateways in 2026 covers the tooling, and vLLM vs Ollama in 2026 covers the self-hosting alternative referenced above.