AI Price War and Token Costs in 2026: Margin Implications
List prices per million tokens make good headlines and poor budgets. What you pay depends on caching, batching, model choice and how many retries your agents make. This guide shows how to measure the cost that matters, what has driven prices down and where the risk of increases sits.
TL;DR
- Token prices have fallen sharply since 2023 and competition from cheaper and open models keeps pressure on.
- Effective cost per completed task matters more than list price per million tokens.
- Caching, batching and routing can cut costs substantially without changing the product.
- Price cuts are not guaranteed to last, because some pricing is subsidised by investor capital.
- Build a price-sensitivity model into your margins, and keep a second vendor warm.
The headline versus the invoice
A price list says a model costs a certain amount per million input tokens and more per million output tokens. Your invoice depends on how many tokens each task uses, how many times an agent retries, how much context is cached and which model handles which job. Two teams with the same list price can have very different costs.
What has driven prices down
- Competition. Several labs and cloud providers compete for developers, and each price cut forces responses.
- Efficiency. Smaller models trained better, distillation and hardware improvements reduce the cost to deliver a given level of quality.
- Open-weight models. Releases such as Meta’s Llama, Alibaba’s Qwen, Mistral’s models and OpenAI’s gpt-oss in August 2025 give buyers an alternative to proprietary APIs. See open-weight models.
- DeepSeek in January 2025. The release of its R1 reasoning model showed that strong results were possible at far lower prices, and markets reacted sharply.
Kelsey Piper has argued that GPT-4-level capability now costs about a thousandth of its original price. Treat the precise figure with care, since it depends on the comparison, but the trend is well established.
The cost that matters: per completed task
Measure cost per successful outcome, not per call.
| Scenario | Model | Calls per task | Success rate | Cost per call | Cost per success |
|---|---|---|---|---|---|
| A | Large model | 1.2 | 92% | 0.050 | 0.065 |
| B | Small model | 2.5 | 70% | 0.008 | 0.029 |
| C | Router: small first, large on failure | 1.7 | 90% | 0.016 | 0.030 |
In this illustration, the cheapest list price (B) has a lower success rate, which creates rework. Routing (C) combines the savings with quality. Your numbers will differ, so measure your own tasks.
Levers that cut cost
- Prompt caching. Stable context such as a policy document or tool definitions can be cached for a discount.
- Model routing. Use a small model for classification, extraction and simple replies. Escalate to a larger model only when needed.
- Batch processing. Many providers discount work that can wait.
- Shorter context. Retrieve only relevant passages, trim histories and remove boilerplate.
- Output limits. Set maximum lengths and structured formats.
- Agent guardrails. Cap retries, tool calls and loop depth so a bad run cannot burn money.
Teams often find that these steps save more than waiting for the next price cut. A deeper look at the energy and capital side is in data-centre electricity costs.
The risk on the other side
Cheaper tokens are good news, but do not assume a straight line.
- Some prices are subsidised by investor capital. If funding tightens, discounts can shrink.
- New, more capable models are often priced higher than their predecessors.
- Capacity constraints can raise prices or limit access.
- Heavy agent usage can grow volume faster than unit prices fall.
Critics argue the whole structure is fragile. One of the best-known sceptics speculated in 2024 that prices might need to rise 10 to 100 times if subsidies ended. That is speculation, but a 3 times stress test is a sensible discipline. We discuss it in the bubble sceptic case.
Pricing your own product
If you resell AI capabilities, make sure pricing covers real costs, including heavy users.
- Use included usage with overage or credits, not unlimited plans.
- Track cost per customer and flag accounts above a threshold.
- Review plan economics quarterly, as in pricing and packaging.
A monitoring routine
- Monthly: compare provider price lists and new model releases against your current choices.
- Per release: run your evaluation set on candidate models for quality and cost.
- Quarterly: recalculate margin per task at current prices and at 3 times.
Read how the labs make money in OpenAI’s business model and Anthropic’s business model.
Contract and vendor tips
When you negotiate or choose plans, a few terms matter more than headline price:
- Price protection. Ask whether prices for your committed volume can change during the term.
- Model lifecycle. Check how long a model stays available before it is retired and what notice you get.
- Rate limits. Confirm throughput at your peak, not only your average.
- Caching and batch discounts. Verify how they apply to your usage pattern.
- Data terms. Check retention and whether inputs are used for training.
- Commit flexibility. Avoid large minimum commitments until your volumes are predictable.
Keep a second vendor integrated and tested. Even if you never switch, it gives you leverage and a fallback.
Your next step
Take your busiest AI workflow, run it through a cheaper model on your test set and record quality and cost side by side. If quality holds for even a third of requests, a simple router will pay for itself within a month.
Bottom line
Token prices have fallen and competition is intense, but your invoice depends on how you use the models. Measure cost per successful task, apply caching, routing and batching, and stress-test your margin against higher prices. Treat today’s cheap tokens as a gift, not a promise.
Frequently Asked Questions
How much have AI prices fallen?
A lot. Kelsey Piper has argued that GPT-4-level capability now costs roughly a thousandth of its original price, though the exact comparison depends on the model and date. The direction is clear, and the pace has varied.
What triggered the price pressure in 2025?
Competition across labs, cheaper efficient models, and open-weight releases. The release of DeepSeek's R1 model in January 2025 showed that strong capability could be offered at far lower prices and shook markets.
What is prompt caching?
A feature where providers store repeated parts of a prompt, such as a long system message or document, and charge less when it is reused. It can cut cost and latency for applications with stable context.
Could prices rise?
Yes. Some providers price below cost, financed by investors. If funding tightens or capacity stays scarce, prices for the newest models may rise or discounts shrink. Plan for it.
