GPT-5.1-Codex-Max

Hans Reinl's December 7, 2025 benchmark pits GPT-5.1-Codex-Max against Gemini 3 Pro and Claude Opus 4.5 on a full-stack MVP build called Speakit, executed locally inside Cursor to isolate model capability from platform tooling. The test covered URL and PDF text extraction, text-to-speech with playback controls, and a reading interface, followed by one iteration cycle to verify completeness. GPT-5.1-Codex-Max required post-generation fixes to start but completed iteration in 9 minutes 20 seconds total, including 8m 20s for initial generation and 1m for verification, outpacing Gemini 3 Pro's 15m 30s and Claude Opus 4.5's 22m. Lighthouse scores reached 74 mobile and 98 desktop. The article frames the comparison as a check on whether headline SWE-bench numbers translate into shipping product quality.

A November 20, 2025 analysis positions GPT-5.1-Codex-Max as a shift from chatbot-style pair programming to agentic coding, where the model writes, executes, debugs, and fixes code autonomously over extended sessions. The release announcement was made on November 19, 2025, specifically targeting the memory limitations of earlier models. The model introduces native context compaction, automatically summarizing essential state such as variable definitions and architectural decisions when sessions near the context limit, enabling theoretically continuous 24-hour development workflows. This approach addresses context drift failures common in prior models, where fixes in one file inadvertently broke dependencies in others due to lost session memory.

OpenAI introduced GPT-5.1-Codex-Max on November 19, 2025 as a frontier agentic coding model available immediately in Codex across the CLI, IDE extension, cloud, and code review, with API access planned. The release emphasizes long-horizon work via native compaction across multiple context windows, enabling millions of tokens per task. Training spans software engineering, math, and research tasks. Benchmark results show GPT-5.1-Codex-Max reaching 79.9% on SWE-Lancer IC SWE and 52.8% on Terminal-Bench 2.0 versus GPT-5.1-Codex (high) scores of 66.3% and 58.1% respectively. The model is the first trained for Windows environments and offers improved Codex CLI collaboration. On SWE-bench Verified, medium reasoning beats its predecessor while using 30% fewer thinking tokens, and a new Extra High reasoning effort targets non-latency-sensitive work.

Vals AI lists GPT-5.1 Codex Max as an OpenAI model released December 4, 2025, with a 400k context window, 128k max output tokens, and token pricing of $1.25 per million input and $10.00 per million output. The default provider is OpenAI with high reasoning effort configured as the standard hyperparameter setting. Benchmark results show 52.86% overall accuracy with 83.56% on LiveCodeBench (ranked 48 of 143) and 22.17% on Vibe Code Bench v1.1 (ranked 77 of 112). Input modalities include text, and file support, while video input is not supported, giving developers a quantitative view of the model's coding performance relative to peers.

Microsoft announced general availability of OpenAI's GPT-5.1-codex-max in Microsoft Foundry on December 5, 2025, making the coding-focused model accessible to enterprise developers through the unified platform alongside offerings from Anthropic, xAI, Meta, and others. Developers can also access it via the Foundry extension in Visual Studio Code. The model targets complex, long-running coding projects with a reported SWE-Bench score of 77.9, emphasizing multi-agent coding workflows, enterprise app modernization for legacy .NET and Java systems, and secure API development. The announcement positions GPT-5.1-codex-max as an enterprise-grade coding agent designed for sustained context across extended engineering sessions.

OpenAI published the official GPT-5.1-Codex-Max System Card on November 18, 2025, providing first-party safety documentation for the model. The card covers baseline safety evaluations including disallowed content, jailbreaks, and vision capabilities testing across the model's product surface. The document also details product-specific risk mitigations such as the agent sandbox and network access controls, plus model-specific mitigations addressing harmful tasks, prompt injection attacks, and data-destructive actions through targeted safety training. A dedicated preparedness section assesses biological and chemical capabilities using benchmarks like long-form biological risk questions and TroubleshootingBench.