Today, Apex is generally available. Built by Callstack, Apex is a cost-effective model for everyday agent work, with particular strength in React Native and Next.js. It works inside the coding tools you already use to build features, debug failures, review changes, and write tests.

If you’ve been using Pixel Canary on Vercel AI Gateway, you were already stress testing Apex's infrastructure. The public stealth preview gave developers a chance to put it to work before we revealed the team behind it.

We’re building Apex into a family of specialized coding models. Today’s release starts with our React Native expertise and a model that also performs strongly on Next.js tasks. Dedicated models for native iOS and Android development are next.

Apex is available through the Apex API with Vercel AI Gateway availability following soon, with commercial self-hosting for companies that want to run it on their own infrastructure.

Built on our React Native experience

We’ve spent years contributing to React Native, maintaining libraries, and helping companies ship applications across platforms. That work involves decisions about navigation, native dependencies, animations, and differences between iOS and Android. Our experience with those decisions shapes what we train Apex on and how we evaluate its answers.

Apex delivers frontier-level performance on our React Native Evals, which measures success across tasks involving navigation, animations, asynchronous state, lists, and React Native APIs.

That quality and low price is important when agents are iterating to achieve a gaol. A task can require several rounds of reading code, making changes, running checks, and correcting mistakes. Cost and waiting time accumulate with every round. We’re building Apex to make that process more affordable while continuing to improve its speed and reliability.

A code review in practice

We compared Apex and Claude Opus 5.5 on a code review task in the Expensify React Native repository. In this recorded run, Apex delivered comparable review outcomes in 2 minutes 16 seconds for $0.09. Opus 5.5 took 13 minutes 30 seconds at $0.60.

That’s 85% lower inference cost and approximately six times faster completion on this task.

One recorded comparison. Costs are calculated using launch-day API pricing. Apex’s recorded session displays $0.17 under the previous cached-input rate; the same usage costs $0.09 at launch pricing.

Pricing that leaves room to iterate

Apex’s direct API pricing starts at $0.50 per million input tokens, with cached input at $0.20 and output at $3.00.

Here’s how those rates compare with GPT-6 Sol and Claude Opus 5.5:

Standard provider API rates, checked September 30, 2026. Excludes cache-write charges, tools, batch discounts, premium processing, and long-context surcharges. Sources: OpenAI, Anthropic. Gateway pricing is listed separately by the provider.

Apex’s output tokens cost 70% less than GPT-6 Sol’s and 85% less than Opus 5.5’s. The cost of completing a task also depends on token usage, caching, and retries. That’s why we evaluate task cost alongside the quality of the result.

For developers, lower costs create room to investigate another approach, review another change, or run another round of verification.

You may already know Apex as Pixel Canary

Before today’s launch, we released Apex anonymously on Vercel AI Gateway under the name Pixel Canary. We wanted developers to try it without knowing who built it, to learn from what they did with it, and to stress test our infrastructure. And it was a good test before going live!

The preview also gave us an external view of its capabilities. On Vercel’s Next.js evaluation, Pixel Canary completed 28 of 31 tasks, matching GPT-6 Astra at high effort. With Next.js documentation supplied through AGENTS.md, it completed 30 of 31. The tasks covered work such as App Router migrations, data fetching, caching, and image and font optimization. Read Vercel’s evaluation results.

React Native is our first specialization, and these results show useful capability beyond it. They also reinforce the importance of evaluating the context and tools around a model alongside the model itself.

Pixel Canary vs Space Bunny is not even a debate . Space bunny took 398 minutes while Pixel Canary took only 105 minutes pic.twitter.com/diZANWxtrP

We've served over 200B tokens over 5 days of the stealth period. It put a real stress on our infrastructure, which started rough, and we managed to get over it at a performance we're happy with. We'd like to thank Vercel to allow us for this opportunity and supporting us during this time.

An open foundation, with our own training pipeline

When we first introduced Apex, it was based on Gemma 4. We’ve since moved to Qwen. Our contribution is the domain data, training, evaluation, and deployment work that turns an open-weight foundation into a specialized coding model.

We curate the dataset around the React Native ecosystem. It draws on documentation, source code, public APIs, libraries, and conversations grounded in real repositories. Knowing which material belongs in that dataset is part of the work. So is checking whether training on it improves the model.

Our pipeline gives us a repeatable way to test new foundations as they become available. We can apply our dataset, evaluate the resulting model, and deploy it when the results justify the change.

We’re also developing tools that let agents interact with mobile applications and check their work. Tools such as agent-device give an agent a way to operate an application and collect evidence of what happened. We believe specialized models paired with domain-specific tools and verification will make coding agents more useful in everyday development.

The same training and evaluation process lets us expand the Apex family into other areas. Models for native iOS and Android development are coming next, with datasets and evaluations built for those domains.

Run Apex on your own infrastructure

Apex is available for commercial self-hosting in your own cloud or data center. You can decide where inference happens and how the deployment fits your infrastructure and data-handling requirements.

Customer-hosted deployments are arranged with Callstack under a commercial agreement. If you’re evaluating models for your organization, we can help assess Apex against your tasks and deployment needs.

Visit apex.callstack.com for API access and setup instructions. To configure a supported coding tool, run:

npx @callstack/apex init

Try Apex on a task from your own project. Inspect the result, run your checks, and compare the time and cost with your existing workflow. We want to hear where it helps, where it gets stuck, and what it needs to understand better.

Tell us where hosted models fall short for your team today. We'll come back with the right approach for your data and workload, and what the rollout path looks like.