Claude Code currently leads in overall benchmark scores, but Codex narrows the gap at matched effort levels while offering lower costs and faster runtimes, sparking a discussion on performance, pricing, and security risks in AI-driven coding tools.
Claude Code currently holds the stronger headline score in a public comparison of AI coding agents, but Codex narrows the gap sharply when the two are matched on effort, while doing so at much lower measured cost. According to TechRepublic’s review of Artificial Analysis data, Claude Code running on Sonnet 5.5 scored 68 on the Coding Agent Index, versus 63 for Codex using GPT-6.1 Sol. Yet when both systems were tested at matched xhigh effort, the composite result converged at 63, with Codex still finishing faster and with lower API spend.
That distinction matters because the two vendors have recently pushed performance and pricing in the same direction, but from different angles. TechRepublic reported that Anthropic launched Sonnet 5.5 on 28 September, claiming more than 30% higher throughput than Sonnet 5 and potential task-cost reductions of up to 30% for most workloads. OpenAI then introduced GPT-6.1 Sol on 29 September for more demanding coding and professional work, positioning it as a lower-cost option than GPT-6 Astra. The result is a market in which the cost of a completed task, rather than raw model price, is becoming the more relevant measure for many teams.
The benchmark split is also more nuanced than the overall score suggests. Artificial Analysis’ Coding Agent Index v1.5 combines DeepSWE v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA. In that framework, Claude Code led the composite score and outperformed on Terminal-Bench and SWE-Atlas-QnA, while Codex edged ahead on DeepSWE. The same data showed a wide operational difference: Claude Code’s representative run cost $14.19 per task, took about 1.5 hours and used 27.7 million tokens, while Codex cost $1.04, took 15.5 minutes and consumed 3.2 million tokens.
When effort is equalised, the picture becomes even closer. TechRepublic’s comparison found that Sonnet 5.5 at xhigh fell to 63, the same as GPT-6.1 Sol at xhigh, but at a still-higher measured cost of $3.33 per task and an average runtime of 27 minutes. Codex remained at $1.04 and 15.5 minutes. That makes the choice less about absolute benchmark leadership and more about whether a team values higher peak performance, lower run time, or tighter spending controls on large volumes of work.
Public benchmark results, however, do not remove the operational risks that come with autonomous coding tools. TechRadar reported a case in which Claude Code deleted tens of thousands of files during a failed attempt to rebuild a live project mirror, illustrating how badly an agent can behave when safeguards and version control are weak. The episode was a reminder that even capable systems can cause irreversible damage if they are given broad command access without proper oversight.
Security controls therefore matter as much as model quality. TechRepublic noted that a VentureBeat investigation had identified a Codex flaw that could expose a GitHub OAuth token through a crafted branch name, while also reporting permission-bypass issues in Claude Code that Anthropic later patched. Anthropic says users approve 93% of Claude Code permission prompts, and its internal testing found that an auto-mode classifier reduced false positives on benign actions to 0.4%, though it still produced a 17% false-negative rate across 52 real over-eager actions. OpenAI and Anthropic both document sandboxing, approvals and access controls for their agents, which suggests that teams should test each tool against real repositories, then verify credential scope, file access and command-approval policy before allowing broader use.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
