GPT-5.4 & GPT-5.4 Pro Deep Dive
OpenAI's March 2026 flagship models reviewed — native Computer Use, Tool Search, configurable reasoning, and a head-to-head comparison with Claude Opus 4.6
GPT-5.4 Key Highlights
Achieves ~80.0% on SWE-bench Verified, 95.1% on HumanEval — on par with Claude Opus 4.6.
First general model with native computer operation support, 75% on OSWorld-Verified, surpassing human baselines.
272K standard, 1M tokens via API — matching Claude Opus 4.6's context window.
New tool-calling system that saves 47% tokens in tool-heavy workflows, significantly reducing costs.
What's New in GPT-5.4?
Released on March 5, 2026, GPT-5.4 is OpenAI's first general-purpose model with native Computer Use support. It scores 75% on OSWorld, surpassing human baselines.
GPT-5.4 Pro is the enterprise-grade deep reasoning variant, priced at $30/$180 (input/output), supporting 922K input + 128K output context. Both are accessible via QCode.cc.
Configurable Reasoning
5 reasoning levels (low/medium/high etc.) for flexible control over reasoning depth vs. cost tradeoffs.
Mid-Response Correction
Mid-response Course Correction — self-corrects during generation, reducing false claims by 33% vs GPT-5.2.
Codex Security
Released 2026-03-06: AI security agent that scanned 1.2M commits, found 10,561 high-severity issues, with 50%+ fewer false positives.
Codex CLI Integration
GPT-5.4 is now the default Codex model, supporting CLI terminal, Slack integration, subagents, and more.
GPT-5.4 Model Family
| Model | Release Date | Context Window | Positioning |
|---|---|---|---|
| GPT-5.4 | 2026-03 | 1M tokens | Flagship General ($2.50/$15) |
| GPT-5.3 Instant | 2026-03 | 128K tokens | Everyday Use |
| GPT-5.3-Codex | 2026-02 | 400K tokens | Coding Optimized (Codex default) |
| GPT-5.2 Thinking | 2025-12 | 400K tokens | Deep Reasoning ($30/$180) |
GPT-5.4 vs Claude Opus 4.6
GPT-5.4 has made impressive progress, but Claude Opus 4.6 maintains several advantages in coding scenarios. Here's an objective comparison:
GPT-5.4
- Lower input pricing ($2.50 vs $5), clear cost advantage for standard tasks
- Native Computer Use support, more mature automation capabilities
- Tool Search saves 47% tokens, more economical for tool-heavy workflows
- Deep integration with Copilot + GitHub ecosystem for smoother enterprise workflows
Claude Opus 4.6 / Sonnet 4.6
- SWE-bench 80.8% (vs GPT-5.4's ~80.0%) — slightly ahead on coding benchmarks
- 1M context + Adaptive Thinking for more flexible reasoning
- Native Claude Code CLI + Agent Teams multi-agent collaboration
- Industry-leading codebase understanding depth, superior for Chinese coding
By March 2026, the AI coding landscape is mature — Claude Opus 4.6 leads in coding reasoning depth and codebase understanding, while GPT-5.4 has advantages in pricing and tool calling. Through QCode.cc, you can access both Claude and Codex model families, choosing the best tool for each task.
Codex CLI v0.104: Rust Rewrite & MCP Integration
In March 2026, OpenAI released Codex CLI v0.104, rewritten in Rust for better performance, with new MCP server integration and web search capabilities.
Built with Rust
Migrated from Node.js to Rust for significantly faster startup and memory efficiency
MCP Integration
Configure MCP servers via ~/.codex/config.toml to extend tool capabilities
Web Search
Built-in web search with automatic result caching for up-to-date information
Claude + Codex, All in One Place
Sign up for QCode.cc to use both Claude Code and Codex at full power. Stable API, transparent pricing, optimized for China connectivity.