
AI21 Labs Startup Credits: Free Credits
AI language models and APIs for text generation and understanding.
Premium: $79/year for unlimited deals
Already have an account? Log in
Deal Highlights
Our AI21 Labs listing: Free Credits
Offer: Free Credits. Our listing sets no requirements and there is no promo code. Step 1 is to go to ai21.com. The listing gives no dollar amount, so the numbers below come from AI21's own pricing page.
We checked these plans and prices against AI21 Labs's official pricing and help pages in November 2026.
AI21's pricing page shows a free trial worth "$10 credits for 7 days." No card is required and sign-up is self-serve. That trial is open to anyone who creates an account, so our listing adds nothing on top of it. Think of this page as a plain guide to what ten dollars and one week of AI21's API will and will not tell you.
AI21's homepage no longer leads with a text API. It now describes a "lab-as-a-service" where forward-deployed researchers train specialized models for a customer's task, with the stages Evaluate, Train, Serve and Improve. It links to two products, Post-Training and Intelligent Gateway. The homepage does not mention Jamba, Maestro or Wordtune at all. The pricing page and docs still list Jamba as a self-serve API, so it works today, but the company's own front door points toward custom work for larger customers.
$10 and seven days is the whole trial
AI21's pricing page lists three plans:
| Plan | What it is | Access |
|---|---|---|
| Free Trial | $10 credits for 7 days, no card required | Self-serve sign-up |
| Pay As You Go | Usage-based pricing, foundation model APIs and SDK, no seat limits | Self-serve |
| Custom Plan | Everything in Pay As You Go plus volume discounts, "premium API rate limits," private cloud hosting, priority support, a dedicated account manager and expert AI consultancy | Contact sales |
AI21 does not say what happens to unused credit, how the trial turns into Pay As You Go, whether trial credit may serve production, which license the weights use, or what Maestro costs.
Jamba Mini and Jamba Large prices per million tokens
The pricing page lists two Jamba models, with prices per 1 million tokens:
| Model | Input | Output |
|---|---|---|
| Jamba Mini | $0.20 | $0.40 |
| Jamba Large | $2.00 | $8.00 |
Jamba Large costs ten times more per input token and twenty times more per output token than Jamba Mini. The page does not say whether these prices apply during the trial or only after it. We assume they do, because the credit is stated in dollars.
AI21 also says an average token is roughly one English word or six characters, and claims this gives up to 30% more text per token than other providers. That is the vendor's claim, and we did not test it.
What the trial buys in requests
Take a typical chat or extraction request: 3,000 tokens in and 500 tokens out.
- Jamba Mini: 3,000 x $0.20 / 1M = $0.0006, plus 500 x $0.40 / 1M = $0.0002. Total $0.0008 a request. $10 covers 12,500 requests.
- Jamba Large: 3,000 x $2.00 / 1M = $0.006, plus 500 x $8.00 / 1M = $0.004. Total $0.010 a request. $10 covers 1,000 requests.
Over seven days, 12,500 requests is about 1,786 a day on Mini, and 1,000 is about 143 a day on Large. That is enough to test quality on a real data set, not to load test.
The rate limit is not the constraint. AI21's docs list 10 requests per second and 200 requests per minute for both models. At 200 a minute, $10 of Large requests of this size is gone in 5 minutes if you script a bulk run.
The 256K context window and what a full prompt costs
All three Jamba models in AI21's docs list a 256K token context window: Jamba Large (398B parameters, 94B active), Jamba Mini (52B, 12B active) and Jamba 3B. The docs call 256K the longest context among open models. The 3B model has no hosted API endpoint listed; it is a download.
Long context is the reason to look at Jamba, so price it. A prompt that fills the window with 256,000 tokens and returns 500 tokens costs:
- Jamba Mini: 256,000 x $0.20 / 1M = $0.0512, plus $0.0002 for output. Total $0.0514, so $10 covers about 194 full-window requests.
- Jamba Large: 256,000 x $2.00 / 1M = $0.512, plus $0.004 for output. Total $0.516, so $10 covers about 19 full-window requests.
Nineteen requests is a small test. If you want to check whether Large can answer questions across a whole contract set or a long codebase, plan the test on paper first. Run Mini on the same documents and compare answers before spending credit on Large.
What the API supports and where it stops
From the Jamba API reference:
- Models are called
jamba-largeandjamba-mini, with no version suffix needed. - Messages accept system, user, assistant and tool roles.
- Tool calling supports the "function" type only, and every function must be one you define yourself.
- A
documentsfield lets you pass content with optional key and value metadata, which is the hook for grounding answers in your own text. response_formatset tojson_objectturns on JSON mode.max_tokenshas a maximum of 4,096. That caps every answer. A long report generated in one call will be cut short.temperaturedefaults to 0.4 and ranges from 0.0 to 2.0.top_pdefaults to 1.0.nreturns between 1 and 16 completions.stopsequences can be up to 64K each.
Streaming must be off when you use tools. And if stream is on, n must be 1. If your product needs streamed tokens and tool calls in the same turn, this API does not do it.
Python and TypeScript libraries exist. The docs show Python only: pip install ai21, then pass the key to AI21Client(api_key=...) or set the AI21_API_KEY environment variable. You get the key from AI21 Studio, though the studio address now redirects to the AI21 homepage, so confirm where sign-up lives when you create the account.
Rate limits and what Custom adds
Both Jamba models have the same limits in AI21's docs: 10 requests per second and 200 per minute. For more, the docs tell you to email sales@ai21.com. The pricing page ties "premium API rate limits" to the Custom Plan but gives no numbers. Cloud marketplaces such as Amazon SageMaker and Amazon Bedrock set their own limits, which you check in their documentation.
If you plan to launch on Pay As You Go, ask sales about limits before launch, not after the first rate-limit error.
Maestro, Mistral models and the Together AI link
Beyond raw Jamba calls, AI21's changelog describes AI21 Maestro, a system for building knowledge agents. It launched on July 8, 2025, and combines retrieval, real-time validation and execution planning. Its "Validated Output" feature enforces requirements you define.
Later entries add detail:
- August 7, 2025:
mistral-small,mistral-8x7bandmistral-7bbecame available inside Maestro, using AI21's own API key. - September 18, 2025: Requirements now work with File Search and Web Search tools.
- October 8, 2025: Jamba Reasoning 3B launched, a 3B-parameter open reasoning model that runs locally through llama.cpp or LM Studio. AI21 claims 2 to 5 times efficiency gains over competitors.
- November 27, 2025: Maestro connected natively to Together AI's cloud for open-source models.
The newest changelog entry we could read is dated December 1, 2025, about 11 months before today (November 8, 2026). Ask AI21 what has shipped since if you are betting a product on Maestro. The pricing page does not list Maestro prices, so we cannot say what an agent run costs.
Open weights, self-hosting and data handling
AI21's docs call Jamba "open models" and say weights can be downloaded into your own environment. The page we read does not name the license and says only that use must comply with AI21's Terms of Service and Usage Guidelines. We did not find a separate license page. Read the terms before you ship a self-hosted model in a paid product.
The homepage says training, evals and serving can run in the customer's VPC or on premises, and that AI21 keeps no copies of customer data. It names no certifications in text. The docs link to a Trust Center at trust.ai21.com, which is where a security review should start.
What the free trial leaves out
- Time. Seven days, then the credit goes away on a schedule the page does not explain.
- Higher limits. The Custom Plan owns "premium API rate limits," private cloud hosting, priority support and a dedicated account manager.
- Long answers. Output is capped at 4,096 tokens per call on any plan.
- Streaming with tools. Not supported together.
- Post-training. The lab-as-a-service work on the homepage is a sales-led engagement, not part of the self-serve plans.
- Startup program. We found no startup credit program on ai21.com. Our listing's "Free Credits" does not point to one either.
Who it fits
- Teams testing long-document tasks. Summarizing or querying a 200,000-token set of filings, contracts or transcripts is the use case the 256K window was built for. Mini at $0.0514 for a full window is cheap enough to run a few dozen experiments.
- Developers who want open weights. If you might self-host later, Jamba Mini or the 3B model gives you a path, subject to the license terms.
- Founders comparing cost per request. Jamba Mini at $0.20 and $0.40 per million tokens is in the same range as the small models we priced below.
Who should skip AI21 Labs
- Anyone who needs answers longer than 4,096 tokens in one call. Pick a model with a bigger output cap.
- Apps that stream tokens and call tools in one turn. The API forbids that combination.
- Teams that need a free tier past one week. Some rivals offer ongoing free or trial usage, which we cover below.
Jamba against Cohere, Together AI and Anthropic
We priced three rivals in November 2026 and reuse those figures here. Compare against the same example: 3,000 tokens in, 500 out.
| AI21 Jamba Mini | Together AI gpt-oss-120B | Anthropic Haiku 5.5 (prompts up to 100,000 tokens) | |
|---|---|---|---|
| Input per 1M tokens | $0.20 | $0.15 | $0.10 |
| Output per 1M tokens | $0.40 | $0.60 | $0.50 |
| Cost of the example request | $0.0008 | $0.00075 | $0.00055 |
| Trial or startup credit | $10 for 7 days | Startup accelerator, up to $15K, $30K or $50K by tier | Startup program cohort was closed when we checked |
On this mix of tokens, Mini lands within a fraction of a cent of the others, and it is not the cheapest in the table. Its edge is the 256K window and open weights, not price.
- Cohere: trial keys are free but cannot be used commercially, with a limit of 1,000 calls a month. Production is pay as you go. Cohere also sells embedding and rerank models, which AI21 does not list on its pricing page.
- Together AI: hosts many open models on one API. Its accelerator is the larger startup program, but it is tiered by funds raised and credits are "up to" figures.
- Anthropic: a different class of model, priced higher at the top end. Its startup program was a closed cohort at last check, so do not plan on credits from it.
If your decision is "which API has the biggest startup credit," Together AI's published program is larger than a $10 trial. If it is "which one can I test on a 200,000-token document this afternoon," AI21 gets you there with no card.
How to get the AI21 Labs offer
- Go to ai21.com and find the sign-up or Studio link.
- Create an account. The pricing page says no card is required for the trial.
- Create an API key. The docs point to a guide called "Create an API Key."
- Install the Python library with
pip install ai21, setAI21_API_KEY, and calljamba-minifirst. - Write down your start date. The trial runs 7 days, so a signup on a Monday ends the following Monday.
- Log token counts per call so you can compare your real cost to the $0.0008 example above.
- Before day 7, decide between Pay As You Go and a rival. Ask support whether unused credit survives.
Getting set up on AI21 Labs
- Start on
jamba-miniand move tojamba-largeonly for the prompts where Mini fails. At ten times the input price, Large should earn its place. - Set
max_tokensexplicitly. The ceiling is 4,096, and a lower cap keeps a runaway answer from eating your credit. - Turn on JSON mode for any output a program will parse.
- Do not enable streaming on calls that use tools.
- If you try Maestro, ask for its prices first, since the pricing page lists none.
Worth knowing before you commit
AI21 is shifting toward custom post-training for larger buyers, and the self-serve Jamba API sits next to that business. The API is documented, priced and open today. Your product would depend on a line of business the homepage does not mention. Keep your prompts portable. The messages format here follows the common chat shape, so moving to another provider is a small change if you need to.
A ten dollar, seven day trial is enough to learn whether long-context Jamba suits your task. It is not a runway. Finish the experiment in the week and decide with your own measured costs.
Who Is This Deal For?
Early-Stage Startups
Seed and pre-seed companies looking to move fast without overspending on tools.
Growing SaaS Teams
Series A+ companies scaling their stack and optimizing software costs.
Solo Founders
Indie hackers and bootstrapped founders who need enterprise tools at startup prices.
Get Free Credits off AI21 Labs
Premium deal. Upgrade once, unlock everything.
Frequently Asked Questions
Everything you need to know about this startup deal.
Get the weekly deals digest
New verified startup deals every week. No spam, ever. Unsubscribe anytime.
Related Offers
Replicate
Used by 1,772 members
$500 in credits
Run open-source ML models in the cloud, deploy Llama, Stable Diffusion, and custom models via API without GPU management.
View offerConsensus
Used by 1,544 members
Free Plan
AI-powered research platform that finds, summarizes, and synthesizes scientific papers.
View offerLaxis
Used by 2,176 members
Free Plan
AI meeting assistant that records, transcribes, and generates actionable meeting notes.
View offerZoomInfo
Used by 2,014 members
Free Trial
B2B data platform for sales and marketing intelligence with contact databases.
View offerDeal Summary
Looking for more startup deals?
Browse all offers