Short answer. Build a 2027 AI budget around cost per solved task, not price per million tokens. That means counting output tokens, retries, long-context tiers, cache behaviour and human review. Then compare a flat plan against that number at three volumes: your expected month, a 2× spike and a 5× growth case.
Four price shapes
| Shape | How it bills | Best for | Budget risk |
|---|---|---|---|
| Per token | Input and output rates per million tokens; cache and batch discounts vary | Low or spiky volume; frontier models | Retries and output length multiply the bill |
| Credits | Prepaid balance drawn down at list rates, sometimes with a platform fee | Many providers behind one key | Credits expire; fees stack on top |
| Flat / fair use | Fixed monthly or annual fee with published caps and fair-use terms | Steady production volume | Concurrency and fair-use ceilings |
| Self-hosted | GPU capex or rental plus operations | High sustained utilisation; data control | Idle capacity and on-call |
What changes in 2027
- Long-context tiers. More providers reprice the whole request above a context threshold, not just the tokens beyond it.
- Peak and off-peak rates. Time-of-day pricing makes scheduling part of the architecture.
- Cache economics. Cache hits get cheaper, but writes, storage and TTL rules can add cost.
- Routing becomes standard. A cheap model for routine work plus a frontier model for hard tasks beats one model for everything.
- Free-tier graduation. A prototype that starts free needs a priced paid path before launch.
Build the budget model
Use a simple formula, then sanity-check it against a pilot:
cost per solved task =
(input tokens × input rate
+ output tokens × output rate
+ cache writes + cache storage
+ retries × average attempt cost
+ human review minutes × loaded hourly rate)
÷ successful tasks
| Line item | What to measure | Typical surprise |
|---|---|---|
| Input tokens | Prompt, retrieved context and system instructions | RAG adds far more input than the user message |
| Output tokens | Answer, reasoning and tool arguments | Output often costs 4× input |
| Retries | Attempts per successful task | A 1.5× retry rate is a 50% cost increase |
| Cache | Hit rate, write price and storage | Writes can cost more than hits save |
| Human review | Minutes to accept, fix or reject | Usually the largest line item |
| Fixed fee | Plan price, seats, gateway or platform fee | Adds a floor even at zero usage |
Flat vs per-token math
At 2026 GPT-OSS rates of $0.15 input and $0.60 output per million tokens, an 80/20 input-output mix costs about $0.24 per million tokens. That gives a clean break-even:
| Plan | Monthly | Break-even at $0.24/M | Cap |
|---|---|---|---|
| Hobby | $8 | ~33M tokens/month | 2M tokens/day |
| Starter | $20 | ~83M tokens/month | 5M tokens/day |
| Builder | $60 | ~250M tokens/month | Fair use; 10 concurrent |
For a 120M-token workload, per-token is about $28.80 and Starter is $20. For 60M, per-token is about $14.40 and Hobby is $8. Below roughly 33M, pay-as-you-go is usually cheaper than a flat plan. The unlimited API guide has the full fine print.
Cost governance
- Daily caps. Set them per workspace or API key, not just globally.
- 80% warning. Plugsky returns an
X-Plugsky-Quota-Warningheader; alert on it before the 429. - Retry breaker. Stop a loop after repeated failures; retries are the fastest way to burn a budget.
- Model routing. Send routine work to a cheap model and escalate only when the task needs it.
- Showback. Attribute spend to a team or feature so the invoice has an owner.
- Monthly review. Compare actual cost per solved task with the budget, not just total spend.
If you run flat plans, the governance question shifts from “how many tokens?” to “which ceiling do we hit first: daily cap, concurrency or RPM?” The feature page lists the current caps.
Questions for vendors
- What is the binding limit for this model: RPD, TPD, concurrency, RPM or a trial credit?
- What happens at 100%: hard 429, queue, throttle or overage?
- How is long context priced — the whole request or only the excess?
- What are the cache write, hit and storage prices?
- Do free-tier prompts train on my data, and in which regions?
- What is the paid upgrade, and what does it cost at 2× and 5× my volume?
If a vendor cannot answer those six, the price is not comparable. The LLM API pricing comparison and the cost calculator are the starting points.
For the free-tier graduation checklist, see Free AI API in 2027.
Frequently asked questions
How much will AI API pricing change in 2027?
There is no reliable single forecast. Some model tiers have become cheaper, while long-context, peak-hour and premium-model rates have risen. Plan for a range of scenarios and keep the ability to route between models rather than betting on one price direction.
Is a flat-rate AI API cheaper than per-token?
It depends on volume and model mix. At 2026 GPT-OSS rates and an 80/20 input-output mix, the blended rate is about $0.24 per million tokens. An $8 plan breaks even around 33 million tokens per month; a $20 plan around 83 million; a $60 plan around 250 million.
What is cost per solved task?
Cost per solved task is total model spend plus retries plus human review time, divided by the number of tasks completed successfully. It is a better buying metric than price per million tokens because it captures quality and rework.
What hidden costs should I budget for?
Output tokens usually cost more than input, retries multiply the bill, long-context tiers can reprice the whole request, cache writes may cost more than cache hits, and human review is often the largest cost. Add all four to the budget.
How do I keep an AI API bill predictable?
Use daily caps, a warning at 80%, a hard 429 at 100%, per-team budgets and a retry breaker. Route routine work to a cheap model and reserve the frontier model for tasks that need it.
Will AI API prices go up or down in 2027?
Both are possible. Competition has pushed some per-token prices down, while context length, peak demand and premium models can push bills up. Keep an OpenAI-compatible interface and a model map so you can switch without a rewrite.
Should I use a free tier for the 2027 budget?
Use free tiers for evaluation, not as the production baseline. They are useful for a prototype and a model comparison, but they can change without notice and often have different data terms from paid plans.
Price the plan, not the token
Run your input/output mix, retries and context through the calculator, then compare the flat plan at your real volume.
Open the calculator → See plansReviewed 20 November 2026. This is a planning guide, not a price forecast. Rates and caps are a dated snapshot; verify current terms before committing.