APEX(1) _{General Commands Manual}

apex — agentic coding model for mobile and web, fluent in React

npx @callstack/apex init
apex run <claude | codex | opencode | pi>
apex undo

A cost-effective, general-purpose model for everyday agent work, built by Callstack. Frontier-level where React runs: React Native on mobile, Next.js on the web. Plans features, writes code, debugs build and runtime failures, reviews changes, writes tests, and carries out React Native, Next.js and library upgrades.

Apex speaks OpenAI- and Anthropic-compatible APIs, so it runs inside Claude Code, Cursor, Codex, OpenCode, VS Code and GitHub Copilot. Swap the endpoint, keep the workflow.

REACT NATIVE        quality   cost of a full eval run
  Claude Opus 5.5    88.6%    ████████████████████████  $28.75
▶ Apex               87.8%    ██████▍                   $7.72   3.7× cheaper
  GPT 6.1 Sol        86.3%    █████████▌                $11.50

NEXT.JS  pass@4     w/ docs   cost per task
▶ Apex               96.8%    █▊                        $0.067* 13× cheaper
  Claude Opus 5.5    96.8%    ██████▎                   $0.234
  GPT 6 Astra        96.8%    ████████████████████████  $0.891

THROUGHPUT          up to 250 tok/s, up to 3× GPT 6.1 Sol

_{Sources: rn-evals.vercel.app · nextjs.org/evals (Sep 25, 2026, 31 tasks). *Apex's Next.js cost is estimated from the token usage of its published eval runs. Without docs in context, Apex scores 90.3%, the same as GPT 6 Astra.}

input           $0.50  / 1M tokens
cached input    $0.20  / 1M tokens
output          $3.00  / 1M tokens
on-premise      custom quote

_{USD, excluding applicable taxes. API access is available to business accounts. Dedicated and on-premise deployments: contact the team.}

• Processed in the EU, with zero data retention always on. Prompts and outputs are never stored or used for training.
• SOC 2-compliant infrastructure; security package and DPA available for enterprise.
• Dedicated, your-cloud or on-premise deployments when shared hosting doesn't fit.

Strongest in React Native and React; frontier models still do better on broad general-knowledge tasks. Like any model it can be confidently wrong, so review its output before you ship.

apex.callstack.com · models · training data summary · @callstack/apex on npm · status · discord

  _{Callstack  ·  2026-10-01  ·  APEX(1)}