跪拜 Guibai
← All articles
Artificial Intelligence · AI Programming

Grok 4.5 Lands as a Fast, Cheap Coding Contender That Closes the Gap with Opus 4.8

By 甲维斯 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Grok 4.5 delivers Anthropic-competitive results on simpler coding tasks at a fraction of the cost and with less token waste, making it a viable daily driver for developers who want fast, cheap code generation without paying for Opus-level reasoning they may not need.

Summary

SpaceXAI released Grok 4.5, positioning it directly against Opus 4.8 and the upcoming GPT 5.6. Official benchmarks show it leading on three of five tests, including Terminal-Bench 2.1, while trailing Opus 4.8 by roughly five points on SWE-Bench Pro. Musk acknowledged the capability gap but pitched the model on speed and price: 80 TPS throughput and API pricing at $2 per million input tokens and $6 per million output tokens.

Hands-on testing with Grok Build, the model's terminal agent, produced a working 3D solar system in four minutes and a playable Super Mario clone in nine minutes. The Mario build included self-verification where the agent launched a browser, played the game, detected a jump-height bug, and patched it autonomously. A complex astronomical-watch single-file HTML challenge completed in one pass without verification, though sub-dial layout and sunrise/sunset labeling needed refinement.

Across nine examples run in a single batch, Grok 4.5 produced zero basic errors in just over nine minutes. The model appears strongest on straightforward coding tasks and weaker on complex software-engineering problems, matching the benchmark story. Its combination of speed, low token burn, and competent output makes it a practical option for developers who do not need top-tier project-level reasoning.

Takeaways
Grok 4.5 beats Opus 4.8 on three of five benchmarks, including Terminal-Bench 2.1, but trails by about five points on SWE-Bench Pro.
API pricing is $2 per million input tokens and $6 per million output tokens, with throughput up to 80 TPS.
The model's context window is 500K tokens, smaller than the 1M-token mainstream.
Grok Build, the companion terminal agent, installs with a single PowerShell command and requires an X Premium membership.
A 3D solar system built from one prompt completed in four minutes with no verification step.
A Super Mario clone took nine minutes; the agent self-verified by launching a browser, playing the game, and fixing a jump-height bug.
An astronomical mechanical watch in a single HTML file finished in one pass without verification, though sub-dial layout and sunrise/sunset labeling had issues.
Running nine coding examples in one batch took just over nine minutes with zero basic errors across all outputs.
Conclusions

Grok 4.5's real competitive edge is not raw capability but efficiency: it burns far fewer tokens by skipping the obsessive verification loops that domestic Chinese models rely on to mask weaker generation.

Musk's public concession that the model is not the strongest but is cheap marks a strategic pivot from chasing benchmarks to winning on developer economics.

The 500K context window is a deliberate trade-off that likely contributes to the model's speed and low cost, even as competitors push toward 1M tokens.

Self-verification that actually plays the generated game in a browser and patches bugs is a concrete step toward autonomous coding agents, not just a benchmark gimmick.

Grok 4.5's absence from the LMSys Arena leaderboard at launch is conspicuous given how quickly other models appear there, and it raises questions about whether xAI is avoiding direct community ranking.

Concepts & terms
Terminal-Bench 2.1
A benchmark that evaluates how well an AI agent can use a command-line terminal to complete complex, multi-step tasks.
SWE-Bench Pro
A professional software engineering benchmark that tests whether a coding agent can fix real, complex bugs in actual codebases. Claude models have historically dominated this benchmark.
Grok Build
A terminal-based coding agent built specifically for Grok models, analogous to Claude Code or OpenAI Codex, that can generate, run, and self-verify code.
Coyote time
A game-design technique that allows a character to still jump for a few frames after walking off a ledge, making platformer controls feel more forgiving.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗