AI API pricing compared
What 7 ai api tools actually cost — free tiers, the cheapest paid plan, and where each free plan runs out. Each figure is taken from the vendor’s own pricing page and dated, so you can see when we last checked.
| Tool | Free tier | Cheapest paid plan | Verified |
|---|---|---|---|
| Replicate | Yes | $0/sec · Pay-as-you-go — CPU (cpu-small) | 14 Jul 2026 |
| OpenAI API | No | $1.75/per 1M tokens · Codex (gpt-5.3-codex) | 14 Jul 2026 |
| Anthropic API | Yes | $17/per month · Pro | 14 Jul 2026 |
| Ollama | Yes | $20/mo · Pro | 14 Jul 2026 |
| Groq | Yes | Custom / contact | 14 Jul 2026 |
| OpenRouter | Yes | Custom / contact | 14 Jul 2026 |
| Together AI | No | Custom / contact | 14 Jul 2026 |
Prices are as published by each vendor and may be region-specific — many vendors price by country. Confirm on the vendor’s page before buying.
Sorted cheapest paid plan first. A “free tier” means there is a genuinely free plan — tap through to any tool to see the full tier breakdown and exactly where the free plan stops. We list no fabricated prices: where a vendor only quotes “contact sales” we say so.
What drives ai api pricing
This is the most usage-pure category on the site: you pay for compute you consume, and the meter takes two forms. Token-based inference is billed per million tokens split into input, output, and a discounted cached-input rate, and output almost always costs several times more than input. OpenAI's chat-latest runs $5 input against $30 output per million; audio models like gpt-realtime hit $32 input and $64 output because audio tokens are dense. The second form is time-on-hardware: Replicate bills GPUs by the second, from $0.000025 on a small CPU to $0.001525 on an H100, so a slow model on a big card costs real money per call. Aggregators add a spread instead: OpenRouter passes list price through free up to $25,000 of monthly inference, then takes 5%. Ollama inverts the whole model, keeping local inference unlimited and charging only for cloud overflow.
How to choose on price
Optimize the output-to-input ratio first, because that is where token bills actually balloon. A verbose model that emits long completions can cost more than a nominally pricier one that answers tersely, so measure real prompt and completion lengths, not headline per-million rates. Cached input is the biggest lever most buyers ignore: reused system prompts and context can drop input cost by an order of magnitude, so structure calls to hit the cache. For non-language workloads, watch the hardware, not the model, since a $5.49/hr H100 billed per second rewards fast inference and punishes slow. Route through OpenRouter when you want provider flexibility without per-vendor contracts and your spend sits under the fee-free ceiling. Choose Ollama's local path when data residency or steady high volume makes any per-token rate a false economy. Reach for dedicated endpoints only when latency guarantees, not raw price, drive the decision.
AI API pricing FAQ
Why does output cost so much more than input?
Generating tokens is sequential and compute-heavy, while reading input is parallel and cheap, so vendors price them separately. OpenAI's chat-latest is $5 per million input but $30 output, a six-to-one gap. This means the shape of your workload matters: summarization is input-heavy and cheap per call, while long-form generation is output-heavy and costs far more even at the same total token count.
What is cached input and how much does it save?
When you resend an identical prefix, such as a fixed system prompt or shared context, providers serve it from cache at a steep discount. OpenAI charges $0.50 cached versus $5.00 fresh on chat-latest, roughly a tenth. If your app reuses long instructions across calls, structuring requests to hit the cache is often the single largest saving available, larger than switching models.
When is per-second GPU billing cheaper than per-token?
When you run open or custom models where you control the runtime. Replicate bills hardware by the second, so a fast model on a cheap card beats per-token pricing, but a slow model on an A100 or H100 can cost more than a hosted API. It only pays off when your inference is quick or your throughput is high enough to amortize the card. Aggregators like OpenRouter sit between the two, passing list price with no fee up to $25,000 of monthly inference, then adding 5%.
Is running models locally with Ollama actually free?
Local inference is unlimited and free once you own the hardware; you only pay for cloud usage overflow. Ollama's paid tiers ($20 and $100/mo) buy 50x more cloud capacity for models too big to run yourself. For steady high volume or sensitive data, local is a genuine escape from per-token metering, but you absorb the hardware and ops cost instead.
See the full breakdown
Open any tool for its complete pricing tiers, free-tier limits, screenshot and honest review.