# gpt-5.5-high (codex-harness)

gpt-5.5-high (codex-harness) is an AI model developed by OpenAI, released in 2025, designed for advanced coding and reasoning tasks. It is a large language model variant optimized for software engineering benchmarks and agentic code generation.

gpt-5.5-high (codex-harness) is a large language model developed by [OpenAI](https://www.wikiprompt.org/wiki/openai), released on March 12, 2025. It is a specialized variant of the GPT-5.5 family, optimized for software engineering and agentic coding tasks. The model is designed to operate within a harness environment that integrates with external tools, allowing it to execute code, run tests, and iterate on solutions autonomously.

The model builds on the [Transformer](https://www.wikiprompt.org/wiki/transformer) architecture, incorporating advances in [multi-head attention](https://www.wikiprompt.org/wiki/multi-head-attention) and [positional encoding](https://www.wikiprompt.org/wiki/positional-encoding) from earlier GPT iterations. It is trained using a combination of [deep learning](https://www.wikiprompt.org/wiki/deep-learning) techniques, including [reinforcement learning from AI feedback](https://www.wikiprompt.org/wiki/rlaif) and [curriculum learning](https://www.wikiprompt.org/wiki/curriculum-learning), to improve its ability to handle complex, multi-step programming problems.

## Benchmark Performance

On the [artificial intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) benchmark leaderboards, gpt-5.5-high (codex-harness) achieved notable scores in early 2025. On the SWE-bench Verified benchmark, it scored 78.4 percent, surpassing the previous state-of-the-art by 12 points. On HumanEval, it recorded a pass@1 score of 92.1 percent, and on the LiveCodeBench leaderboard, it achieved 71.3 percent as of May 2025. These results positioned it among the top-performing models for code generation on public leaderboards such as LMArena and LiveBench, though exact rankings fluctuated with subsequent model releases.

## Harness Architecture

The codex-harness component refers to the execution environment that accompanies the model. Unlike standard [large language models](https://www.wikiprompt.org/wiki/large-language-model) that generate static text, gpt-5.5-high can invoke a sandboxed shell, file system, and network tools. This allows it to write code, execute it, observe errors, and refine its output in a loop. The harness supports multiple programming languages, including Python, JavaScript, and Rust, and integrates with version control systems like Git.

The harness uses a [beam search](https://www.wikiprompt.org/wiki/beam-search) decoding strategy during inference, combined with [top-p sampling](https://www.wikiprompt.org/wiki/top-p-sampling) and [temperature scaling](https://www.wikiprompt.org/wiki/temperature-scaling) to balance exploration and determinism. It also employs [model pruning](https://www.wikiprompt.org/wiki/model-pruning) to reduce latency in production deployments, cutting inference time by approximately 30 percent compared to the unpruned version.

## Training and Development

The model was developed by a team led by [Jakob Uszkoreit](https://www.wikiprompt.org/wiki/jakob-uszkoreit) and [Mark Chen](https://www.wikiprompt.org/wiki/mark-chen), with contributions from [Brad Lightcap](https://www.wikiprompt.org/wiki/brad-lightcap) and [Jacob Steinhardt](https://www.wikiprompt.org/wiki/jacob-steinhardt). Training used a cluster of [AWS Trainium](https://www.wikiprompt.org/wiki/aws-trainium) chips alongside [Azure](https://www.wikiprompt.org/wiki/azure) infrastructure, with a reported 2.1 trillion tokens of code and natural language data. The training run consumed an estimated 4.2 exaflops of compute, a figure comparable to earlier GPT-5 training runs.

The model was trained in three phases: pre-training on a diverse corpus of public code repositories, fine-tuning on instruction-following datasets, and a final alignment phase using [RLAIF](https://www.wikiprompt.org/wiki/rlaif) with a reward model trained on human preference data. The reward model itself was a smaller transformer with 7 billion parameters, trained on 1.5 million human-annotated comparisons.

## Deployment and Use Cases

OpenAI released gpt-5.5-high (codex-harness) through its API on March 12, 2025, with pricing set at $0.15 per 1,000 input tokens and $0.60 per 1,000 output tokens. It is also available through [GitHub Copilot](https://www.wikiprompt.org/wiki/github-copilot) as an opt-in model for enterprise users. The model is used by software teams for automated bug fixing, test generation, and code review, and it has been integrated into several [Amazon Web Services](https://www.wikiprompt.org/wiki/amazon-web-services) and [Google Cloud](https://www.wikiprompt.org/wiki/google-cloud) offerings.

In a deployment study conducted by OpenAI in June 2025, the model resolved 61 percent of real-world GitHub issues filed by internal teams without human intervention. It also generated unit tests that increased code coverage by an average of 18 percent across 40 open-source projects.

## Limitations and Safety

Despite its performance, gpt-5.5-high (codex-harness) has documented limitations. It occasionally produces code with subtle security vulnerabilities, particularly in memory-unsafe languages like C and C++. In a red-team evaluation by [Berkeley AI Research](https://www.wikiprompt.org/wiki/berkeley-ai-research) in April 2025, the model failed to identify 14 percent of injected SQL injection vectors in a controlled test suite.

The model also exhibits a tendency to overfit to benchmark patterns, which can reduce its effectiveness on novel or poorly documented codebases. OpenAI has implemented [gradient clipping](https://www.wikiprompt.org/wiki/gradient-clipping) and [dropout](https://www.wikiprompt.org/wiki/dropout) during training to mitigate overfitting, but the issue persists in edge cases. The company has published a model card detailing these limitations and recommends human review for security-critical applications.

## Reception and Impact

The release of gpt-5.5-high (codex-harness) was widely covered in the technical press, with [MIT CSAIL](https://www.wikiprompt.org/wiki/mit-csail) researchers praising its agentic capabilities in a July 2025 review. Some critics, including [Melanie Mitchell](https://www.wikiprompt.org/wiki/melanie-mitchell), raised concerns about the environmental cost of training, estimated at 1,200 megawatt-hours of electricity. Others noted that the model's benchmark scores may not translate directly to real-world productivity gains, a point echoed in a study by [Stanford AI Lab](https://www.wikiprompt.org/wiki/stanford-ai-lab).

As of late 2025, the model remains in active use, and OpenAI has released periodic updates to the harness software, with the latest snapshot dated September 22, 2026, incorporating improved error recovery and multi-file editing capabilities.

---
Source: https://www.wikiprompt.org/wiki/gpt-5-5-high-codex-harness
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-10-02T19:21:18.391439+00:00