OpenAI Codex has changed substantially with the most recent model update. Here is what developers think after spending real time with the new version — where it shines, where it frustrates, and who should actually reach for it.
What developers love:
Sandbox execution is Codex’s unique advantage. Codex runs generated code in isolated containers before showing results. Developers working with untrusted code or experimenting with unfamiliar libraries feel markedly safer with Codex than with an agent that executes directly on their machine.
Subagent orchestration is genuinely strong for complex parallel work. Codex can spin up many subagents to attack different parts of a problem at once — for example, splitting a large refactor across modules and reconciling the results afterward. For the right workload, that parallelism is a real productivity multiplier.
What frustrates them:
Tool-call reliability has been a persistent sore spot after the update. The model sometimes hallucinates function calls or picks the wrong tool for a step. The community has documented recurring bug reports around this, and it tends to show up most on long, multi-step tasks where one bad tool call cascades.
Pricing is higher than competitors for heavy usage. Codex’s per-request pricing adds up fast for developers running frequent agentic sessions, and teams doing all-day autonomous work will feel it more than occasional users.
Where it fits versus the competition:
If you want the full picture on head-to-head usage, the Codex vs Claude Code and Codex vs Cursor breakdowns cover the tradeoffs in detail. On economics specifically, Codex cloud execution costs digs into what heavy agentic use actually runs you. And for the parallel-work angle, Codex’s parallel-agent approach is worth a read.
The verdict:
Codex is the best choice for security-conscious developers and teams running untrusted code. For general day-to-day development, Claude Code and Cursor often offer better value. The sandbox is genuinely unique, but most developers don’t need it on every single task — which is why many keep Codex in the rotation rather than making it their only agent.
Related articles
- Coding Agents in 2026: Three Hard Lessons HN Developers Learned the Expensive Way
- What Developers Think About Coding Agents — The Skill Atrophy Crisis, Interface Wars, and Trust Divide From 500 HN Comments
- Codex 0.145–0.146: Voice Coding, Agent Plugins, Multi-Agent V2, and More
The smartest developers don’t pick one AI — they use them all. aiFiesta brings 9+ premium models into one chat for $12/mo. Your AI toolkit, simplified.