▶ YouTube Summaries

← All summaries
ForrestKnight#▶ Watch

Grok 4.5 (Cursor) Coding Review

2026-07-10 Fri ⏱ 27 min forrestknight

Forrest Knight reviews Grok 4.5, released jointly by xAI and Cursor (following xAI's acquisition of Cursor), arguing it delivers Opus-4.1/GPT-5.1-level coding quality at a fraction of the price and speed. He credits the jump from the weaker Grok 4.1 to training on trillions of tokens of real Cursor developer-interaction data plus xAI's large-scale GPU infrastructure. Through hands-on tests across TypeScript, Rust, and a one-shot Rust/WASM game, he shows Grok 4.5 solving a bug Claude Code failed at repeatedly, while flagging real code-quality issues (a semver string-comparison bug, unused dead code, ignoring an existing crate). He contrasts Grok 4.5's fast back-and-forth workflow with Claude Code's long-running autonomous agent style, concluding Grok 4.5 is a serious contender and planning to use it as a daily driver and cheap sub-agent.

# What Grok 4.5 is

xAI and Cursor jointly released Grok 4.5 (branded "Composer" inside Cursor). Unlike Cursor's own coding-only Composer models, Grok 4.5 is a general-purpose model — meant for software engineering, data science, finance, legal work, and more — trained jointly by the two companies.

# Why it suddenly got so much better

He attributes the leap from the much weaker Grok 4.1 to two factors from xAI/Cursor's blog posts: training on trillions of tokens of real Cursor usage data (how developers accept, edit, or reject AI code changes), and training across tens of thousands of Nvidia GB300 GPUs. He argues the Cursor acquisition was really about this data, not just revenue or talent.

# Skepticism about benchmarks

He distrusts most published benchmarks (especially SWE-bench style) as gameable or contamination-prone, noting Cursor Bench was deliberately excluded from Grok 4.5's comparisons because an earlier snapshot of Cursor's own codebase had leaked into training data. He instead trusts Artificial Analysis benchmarks, where Grok 4.5 jumped from near the bottom (as Grok 4.1) to roughly on par with Opus 4.1 and GPT-5.1.

# Pricing and efficiency

Grok 4.5 costs $2/M input and $6/M output tokens, versus Opus 4.1 at $5/$25 and GPT-5.1 at $5/$30. It's also reported to be roughly twice as token-efficient, solving tasks in under half the steps — cheaper and faster on top of cheaper per token.

# Hands-on coding tests

  • Rust/Electron screen recorder port (Mac to Linux): Grok 4.5 fixed a broken preview ("rainbow screen" wiring bug) that Claude Code failed to fix across six or seven attempts, by rewiring the preview to a live JPEG feed. Functional in minutes versus repeated failures.
  • Rust JDK/Maven toolchain manager ("UV for Java"): full phase with 24 passing tests, clean compile and Clippy. Strength: a deliberate documented stable tie-break sort for reproducible installs. Bug: comparing semver as strings (mis-ranking "209.9" vs "209.11"), and reimplementing version comparison manually instead of using an existing crate already in-project.
  • TypeScript link-shortener: praised for reusing an existing reserved-slugs set and choosing a 307 redirect over 302 (avoids replaying POST as GET, protecting Stripe/Clerk webhooks). Criticized for dead/unused code left after a mid-implementation pivot (an unused shortDomain field).
  • One-shot Rust/WASM lawn-mowing game: Grok 4.5 correctly built the whole game in Rust compiled to WASM as instructed; Claude Code, same prompt, instead built the frontend in JS/CSS — better-looking but off-prompt.

# Workflow philosophy: Grok 4.5 vs. Claude Code

He characterizes Claude Code as a long-running, autonomous, sub-agent-heavy "give it a big task and walk away" model, whereas Grok 4.5 (like Opus 4.1 and GPT-5.1) is fast and interactive. He usually prefers slower-but-correct over fast-but-wrong, but says Grok 4.5 avoids that tradeoff by being both fast and comparably capable to Opus 4.1/GPT-5.1.

# Conclusion

He plans to use Grok 4.5 as a cheap, fast drop-in replacement for Opus 4.1/GPT-5.1, and as an efficient sub-agent that larger orchestrator models delegate bulk execution to (given Claude Code's high API cost). He frames the top tier as Anthropic, OpenAI, and now xAI (via Cursor), with Google notably absent. He closes noting this reflects only his own personal testing and invites viewer feedback.