GPT 5.5 just did what no other model could

What It Covers

Claire Vaux reviews GPT-5.5 and GPT-5.5 Pro after two weeks of early access testing, focusing on Codex-based autonomous coding tasks. She demonstrates three real-world use cases: security remediation, a 2-million-row data migration, and reverse-engineering a proprietary Bluetooth device — comparing results against Claude and GPT-5.4.

Key Questions Answered

• •GPT-5.5 Pricing vs. ROI: GPT-5.5 costs $5 per million input tokens and $30 per million output tokens; the Pro tier runs $30 input and $180 output. Evaluate cost against ambition, not just speed — if the model solves problems no other tool could, the token cost is justified by capability unlocked rather than time saved alone.
• •Batch Technical Debt Remediation: Feed GPT-5.5 in Codex a CSV export of security or technical debt issues and instruct it to group thematic problems, propose architectural fixes, and implement them in one pass. This approach cleared a full security backlog and produced a clean annual penetration test result without addressing issues one by one.
• •Autonomous 6-Hour Agent Loops: GPT-5.5 in Codex ran a self-directed sub-agent loop for nearly six hours with zero follow-up prompts, testing a production-like dataset of 2 million rows for legacy data format edge cases. The result was one unresolved edge case from millions of rows, dropping the application error rate to near zero in Sentry monitoring.
• •Proprietary Protocol Reverse Engineering: When Claude Opus and GPT-5.4 both failed to decode a proprietary Bluetooth device's communication protocol, GPT-5.5 succeeded after being given Bluetooth packet sniffer logs. Use hardware packet capture data as context input — this unlocks reverse-engineering tasks previously considered unsolvable through AI-assisted coding alone.
• •Codex Personality Customization: GPT-5.5 in Codex defaults to a flat, minimal communication style. Running the slash-personality command inside Codex allows users to switch to a more conversational tone. For teams doing long autonomous sessions, adjusting this setting improves the feedback experience without affecting the model's underlying reasoning or output quality.