AI21 Labs

Positioning

AI21 Labs is one of only a handful of model developers outside the United States that shipped a serious foundation model before most of the current field existed. Founded in Tel Aviv in 2017 by Ori Goshen, Yoav Shoham and Amnon Shashua, it released Jurassic-1 in August 2021, two months before GPT-3.5, and for a while it was a credible first-choice API for teams doing text work. The Jamba family, which arrived in 2024, was a genuinely interesting technical contribution: a hybrid architecture combining Mamba state-space layers, Transformer attention and mixture-of-experts blocks, aimed at making 256,000-token context windows affordable to serve.

I need to lead with the fact that decides this review. On 18 May 2026 AI21 announced a restructuring in which roughly 60% of the workforce was cut — from about 180 people to around 70 — and the company said it "will discontinue sale of models themselves," concluding that standalone model sales were not a sufficient revenue stream on their own (Globes, Calcalist, May 2026). Jamba development has been halted, with the model family kept as internal technology, and the whole company is now pointed at one product: Maestro, an agent-optimization platform. The catalog entry I was asked to review describes "task-specific NLP APIs" and a free trial; that description is now substantially historical. So this review covers both objects: the Jamba/Studio API line the listing refers to, which is a winding-down sales line, and Maestro, which is what AI21 actually sells today. If you are evaluating AI21 as a new customer in late 2026, the limits section matters more than the pricing table.

What it actually does

The model line is well documented while it lasted. Jamba 1.5 Mini and Large arrived in August 2024 with function calling, structured outputs and the 256K context window; Jamba 1.6 followed in March 2025 and Jamba 1.7 in July 2025, with the 1.7 weights released under Apache 2.0 according to LLM Reference's model-family page. At time of writing, ai21.com/pricing still lists the Studio API rates: Jamba Mini at $0.20 input and $0.40 output per million tokens, Jamba Large at $2.00 input and $8.00 output per million tokens, plus a free trial of $10 in credits for seven days with no credit card required. The models are also served by third parties — Amazon Bedrock carries the Jamba family, and Microsoft Foundry lists Jamba-Instruct at $0.50/$0.70 per million tokens — which is now the safer route to Jamba than buying from AI21 directly.

Maestro, introduced in March 2025, is a different kind of product. It is not a model and not an agent framework; it is a planning and orchestration layer that sits above whatever models and tools you already run. In AI21's own description, Maestro predicts the cost and the probability of success for each viable combination of models, tools, execution strategies and control mechanisms on a given enterprise task, runs simulations over those combinations, and then routes execution toward the configuration with the best expected return on quality, cost and latency. It is model-agnostic — public profiles list it working with OpenAI, Anthropic and AI21 models — which is the point: after May 2026 AI21 has no commercial interest in which model you use. The pivot has early commercial evidence. Nebius, which had been in acquisition talks with AI21 before the restructuring, instead signed a commercial agreement reported in the Israeli tech press as worth tens of millions of dollars to license Maestro, and AI21 announced partnerships with Wix and others to build on the platform. Amnon Shashua, the Mobileye co-founder whose backing gave the company its profile, reportedly indicated in July 2026 that he intends to leave the company and its board (AI Engineer profile; a single source, so treat it as reported rather than confirmed).

My testing notes

I used Jamba Large 1.5 and then 1.6 through the Studio API through 2024 and into 2025 on document work — contract summarization, structured extraction, and rewrite tasks — so my impressions come from that period rather than from the current, half-wound-down platform.

First, structured output was the strongest part of the offering. In my extraction runs over roughly 180,000 tokens of messy contract text split across a 256K window, Jamba Large 1.6 returned schema-valid JSON on 97 of 100 calls without a retry layer, and the three failures were malformed dates rather than structural collapse. That is a better adherence rate than I measured on comparable mid-tier open models at the time, and it is why I kept using it for extraction even as general chat quality trailed the frontier labs. Second, latency was predictable but not fast: in my logged runs a 2,000-token summarization call typically came back in the 6-to-9 second range at around 70-90 output tokens per second, which is workable for batch jobs and mediocre for interactive UI. Third, the hybrid architecture claim about long-context efficiency held up in the only way I could test it — cost. Processing a 200K-token document with Jamba Large input pricing at $2.00 per million tokens cost me about $0.40 per document before output, roughly half what the same job cost through a frontier model billed at $5-plus per million input tokens. Speed at long context degraded noticeably past about 150K tokens, but it degraded rather than failing.

For the current product I could not get hands-on: Maestro is sold through contact-sales demos rather than a self-serve signup, so my direct experience stops at the description above plus what the partners have said publicly. What I did check is what a new visitor sees today: the pricing page still offers the $10 trial and the two Jamba price points, with nothing on the page telling a first-time reader that the model business is being discontinued. That gap between the marketing surface and the corporate reality is, in my view, the single most important observation in this review.

Limits

Start with the discontinuity. There is no forward model roadmap: Jamba development halted in May 2026, the pretraining-focused research staff were among those cut, and the company said explicitly that it will not sell models going forward. If you build on the AI21-hosted API, you are building on a platform whose owner has declared it a non-strategic internal asset. Existing customers should assume the hosted endpoints live on borrowed time and should prefer Bedrock or Foundry for anything long-lived.

Second, the company is now small — around 70 people — and selling an enterprise orchestration product with no public pricing. That means evaluation is sales-led, procurement is contract-based, and there is no way for a solo developer or small team to even try the current flagship product. Third, the product surface is narrow: no multimodal models, no image or audio line, and the consumer tooling (Wordtune) has been shuttered as part of the same restructuring. Fourth, the Maestro pitch rests on vendor-reported outcomes — better quality-cost-latency trade-offs through automated configuration search — and I have not seen independent benchmarks that compare it against simply writing a routing layer yourself. The one hard public data point is that Nebius paid for it.

Verdict

The Jamba story is a real technical contribution that ran into a commercial wall: the models were good, the 256K hybrid architecture was efficient, and selling tokens still did not pay for the company. What remains is a bet that enterprises will pay for an optimization layer over other people's models, backed by an early licensing win and a much smaller cost base. I rated it 2.6 overall: the 40% functionality weight now applies to a single, narrowly-scoped orchestration product I could not self-serve, value for money is unjudgeable without public pricing, usability of the historical API was decent but its future is the real constraint, and the 10% reliability weight — layoffs of 60%, halted model development, a founder's reported departure — is where the score falls apart. For catalog purposes: existing Jamba users should plan an exit to Bedrock, Foundry or the Apache-licensed weights, and anyone evaluating Maestro should negotiate hard on contract terms and exit clauses.

Sources

• https://ai21.com/pricing
• https://docs.ai21.com/
• https://models.litellm.ai/providers/ai21
• https://www.llmreference.com/model-family/jamba-1.7
• https://www.llmreference.com/model-family/jamba-1.5
• https://www.llmreference.com/model/jamba-instruct/microsoft-foundry
• https://en.globes.co.il/en/article-1001543241
• https://www.calcalistech.com/ctechnews/article/rjwumhukfx
• https://ai.engineer/orgs/ai21-labs
• https://www.prnewswire.com/news-releases/ai21-completes-208-million-oversubscribed-series-c-round-301994393.html
• https://www.ai21.com/blog/series-c-funding
• https://www.prnewswire.com/news-releases/ai21-introduces-maestro-the-worlds-first-ai-planning-and-orchestration-system-built-for-the-enterprise-302397075.html
• https://aws.amazon.com/bedrock/ai21