OpenAI released GPT-6.2 Astra and Sol for Codex and the API on Tuesday launching head-to-head in the terminal against Claude Code after Anthropic deployed Claude Fable 5.2 on November 9, 2026, according to Shattered. Developers testing both agent tools face a stark reality: Stack Overflow’s Developer Survey 2025 shows 45 percent of developers spend more time debugging AI-generated code than writing it by hand.
Claude Fable 5.2 spans Pro, Max, Ultra, Team, and Enterprise plans while Codex GPT-6.2 limits Free and Go tiers and unlocks Plus at $20 monthly and Pro at $100 monthly.
Evaluating AI code agents through feature lists rarely reveals how they handle unfamiliar codebases. While both Claude Code and Codex read files, execute commands, and modify code within the terminal, debugging provides a definitive measurement where a red test turns green without breaking anything else.
Shattered reported that this exact methodology anchors the bugbench harness, which runs automated tests using a disposable clone per run to score success, discipline, intervention size, and effort.
Testing Code Discretion
The discipline rule prevents agents from gaming the evaluation by simply weakening assertions or deleting failing tests. Shattered reported that the scorer compares SHA-256 hashes of all files under the tests directory before and after each run to catch alterations.
An empirical study on arXiv analyzed bugs across CLI tools, noting that 67.0 percent of recorded issues were functional bugs while 37.3 percent stemmed from API, integration, or configuration problems. Out of the evaluated issues, 2,343 bugs were tied to Claude Code and 1,192 to the Codex CLI, illustrating that tool-level failures often interrupt automated benchmarks independently of the underlying models.
| Model / Tool | Release Date | Deployment Status |
|---|---|---|
| Claude Fable 5.2 | November 9, 2026 | Available, selectable in Claude Code |
| GPT-6.2 Astra & Sol | November 17, 2026 | Preview in Codex Desktop, API available |
Because six small bugs cannot establish a universal winner, the benchmark scenarios remain intentionally concise to keep execution times under minutes and costs within single-digit dollars. Complex issues like race conditions or memory leaks require bespoke scenarios drawn directly from specific repositories.
The bugbench framework structures these tasks through a target folder containing a shop library with pre-inserted flaws alongside a Python harness script that coordinates fresh clones, agent execution, and results logging into runs.jsonl.
Execution requirements demand a local environment running on Linux, macOS, or WSL2 equipped with Python 3.11 or newer and Git. Claude Fable 5.2 operates across Anthropic plans including Pro, Max, Ultra, Team, and Enterprise, whereas OpenAI includes Codex access within ChatGPT plans with limitations on Free and Go tiers, lifting restrictions on Plus subscriptions at $20 monthly and expanding allocations on Pro plans starting at $100 monthly.
Developers can alternatively utilize API keys governed by token pricing models, creating distinct accounting paths where ChatGPT account users consume plan quotas while API keys incur direct per-token charges.
Whether API billing structures or subscription token pools ultimately dictate long-term developer preference remains unmeasured as teams run custom evaluations on their own proprietary codebases.