That is the bar every comparison has to clear. Look at the cached-input discount in the table, the part of the bill that shrinks most when you resend the same instructions each call. The rivals' discounts sit in a narrow band, so the gap between vendors comes down to list price and context, not caching. | Model | Vendor | Discount on cached input | |---|---|---| | Claude Fable 5.1 | Anthropic | 98% | | Claude Opus 5.5 | Anthropic | 95% | | Claude Sonnet 5.5 | Anthropic | 95% | | GPT-6.1 Sol | OpenAI | 95% | | Gemini 4 Argon | Google | 95% | | Claude Haiku 5.5 | Anthropic | 90% | Source: [3AM Marketer.com](https://[3ammarketer](https://3ammarketer.com/data/model-pricing/prompt-caching)/data/model-pricing/prompt-caching). ### Small, cheap models compared At 90 percent off cached input, Claude Haiku 5.5 has the lowest discount here, and it is [Anthropic](https://www.anthropic.com/news/claude-3-5-sonnet)'s current fast tier. You send the most calls to a small model, so a 90 percent discount still means one tenth of the input price on every repeat. Jamba Mini has to beat that figure on its own list price to win. ### Large models compared At 98 percent off cached input, Claude Fable 5.1 makes a long, unchanging prompt cost almost nothing to resend. Opus 5.5, Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon all sit at 95 percent, which means buyers cannot choose between them based on caching. If your prompts are long and repeated, Jamba Large faces a cached price one-twentieth of fresh input at four rivals, and one-fiftieth at Fable. ### When AI21 is the cheaper choice AI21 wins where your prompts change on every call, because caching then saves you nothing and only list price counts. Where the same large prompt repeats, run both bills before committing.  ## Run recurring AI21 calls with Activepieces If the same large prompt keeps going out, the cheapest fix is to control when it goes out and how many times. Activepieces, an automation tool that chains steps together, does that job, with unlimited flows on every plan including the free one, and a run costs 1 credit however many steps it has.  It doesn't lower AI21's price. It sits in front of the API and decides what gets sent. In the screenshot you can see the kind of workflow this suits: a five-step flow called "Competitor Pricing Change Alert," set to run every Friday at noon UTC. It reads a list from Google Sheets, clears a storage step, loops through each row, and scrapes a page per row. Swap the last step for an AI21 call and you have a weekly job with a fixed rhythm instead of someone pasting prompts by hand. The point to take from it is the schedule: the bill is one run a week, not whenever someone remembers.  You reach AI21 through a generic HTTP step that posts to its API with your key, since it has no dedicated step here. That works, but you build the request yourself. Triggers can just as easily be a form submission, a new email or a new ticket. ### Cap and log spend per workflow Without a ceiling, a loop that misfires can send the same expensive call hundreds of times before anyone notices. Put a counter in the flow's storage and stop the run when it passes your limit. After each call, write the model name, the input size and the output size to a sheet row. Multiply those by the list price once a month and you have your own bill, per workflow, before the invoice arrives.  ### Batch inputs before the API call When calls are triggered one ticket at a time, each carries the full instructions again. Instead, let the flow collect items for a day or a week, then send them together in one request where the model's context allows. You pay for the shared instructions once rather than per item. Check the result against the cost log, because a batch that overruns the context window fails entirely. ## Checking your AI21 bill on Monday Follow five steps in order, starting from the cost log you just checked, and your first AI21 bill holds no surprises. Skipping them means a month of guessing; doing them costs an afternoon and a week of patience. 1. Re-check the pricing page on signup day. 2. Run a 50-example test and pick the cheapest model that passes. 3. Set a budget alert. 4. Log tokens per call for a week. 5. Call sales only once your volume justifies it. Each step feeds the next. The price you confirm sets what the test costs, the test picks the model, and the model sets the spending level your alert should watch. ### Re-verify the price page Prices change. Any table in an article, this one included, is a snapshot. Open AI21's own pricing page the day you sign up and write down the date next to each rate you rely on. If you budget from a stale number, every estimate after it's wrong by the same margin. ### Test the cheapest model first Collect 50 real examples from your own work, not invented ones. Run them through the cheapest model, then step up only if it fails. A pass means the output is something you would actually send or use. Paying for a larger model when a smaller one passes is the most common way a bill grows with no gain in quality. ### Measure a week of real tokens Set the budget alert before real traffic starts, so an overrun reaches you as a message instead of an invoice. Then record the tokens in and out of every call for a week. Expect your real prompts to run longer than your estimate, because instructions and pasted documents ride along each time. Once you have a week of data, you can multiply by your expected volume. If that total is large enough that a discount would matter, call sales, with your own numbers in hand. Before that point, the call tells you little. ## Frequently asked questions Here are the questions the pricing body leaves open; each answer points you to the page or test that settles it. ### Is there a free AI21 API plan? New accounts have generally received trial credit to test with, but the amount and the models it covers can change. Check your account dashboard before you build a plan around it. Treat the credit as a test budget. Once it runs out, calls bill at the normal rates, so find out what happens at that point before you ship anything to real users. ### Does AI21 charge for input and output tokens separately? Yes, each is priced on its own line. Output tokens usually cost more than input tokens, so the mix matters. A job that sends long documents and gets short answers back bills very differently from one that sends a short question and gets pages of text. Estimate both sides for your own work. ### Is AI21 cheaper than OpenAI? It depends on the model you compare and on what you send. A lower list price can still lose if the cheaper model needs longer prompts or more retries to do the same job. The fair test is to run the same real requests through both vendors and compare the resulting monthly bills, not the rate cards. ### Can I get a discount for committed volume? Don't assume so from the public price list. Volume arrangements are normally negotiated with the vendor's sales team, and they depend on how much you can commit to. If you're spending enough that a discount would matter, ask, and get the terms in writing before they go into your budget. ### How often does AI21 change its prices? When models are added, replaced or retired, prices move, and there's no fixed schedule. Date every estimate to the day you read the pricing page, and re-check it before you renew a budget or sign a contract. ## Related reading - [Gemini API Pricing 2026: Every Model, $1.25 per 1M Tokens](https://www.activepieces.com/blog/gemini-api-pricing-2026-every-model-tokens-limits) - [DeepSeek API Pricing 2026: Cost per 1M Tokens](https://www.activepieces.com/blog/deepseek-api-pricing-2026-cost-per-1m-tokens) - [Claude API Pricing 2026: Rates for Every Model](https://www.activepieces.com/blog/claude-api-pricing-2026-rates-for-every-model) ## References - [CloudCostDB](https://cloudcostdb.com/ai/ai21-labs/jamba-1-5-large/) - [PeakInfer](https://www.peakinfer.com/blog/how-failed-requests-inflate-your-bill) - [3AM Marketer](https://3ammarketer.com/data/model-pricing/prompt-caching)Only when its price per call, after any cache discount, beats a rival's is AI21 cheaper.