OpenAI's GPT-5 is here. What can you expect from it in Cline?

The most intelligent and reliable coding model from OpenAI is now available in Cline – and it's half the price of Sonnet 4.

GPT-5 is now available in Cline. After testing early with OpenAI, we're seeing a model that's broaching the cutting-edge in software engineering while delivering the reliability developers actually need. For teams building production applications, GPT-5 present a shift in what's possible in agentic coding.

Exceptional Coding Performance

Let's start with the benchmarks. GPT-5 is a significant improvement on o3, OpenAI's previously most-powerful model

On SWE-bench Verified, which tests real software engineering tasks, GPT-5 achieves 74.9% accuracy with thinking enabled – more than doubling GPT-4o's 30.8%. This isn't about toy problems; SWE-bench tests the ability to solve actual GitHub issues, the kind of real-world debugging and feature implementation developers face daily.

On factuality benchmarks, GPT-5 makes ~80% fewer factual errors than o3. On LongFact-Concepts, GPT-5's hallucination rate is just 1.0% compared to o3's 5.2%. For FActScore, GPT-5 achieves 2.8% versus o3's 23.5%. This dramatic reduction in hallucinations means fewer broken implementations, fewer subtle bugs, and more code that works the first time.

Benchmarks can be telling, but real-world usage data is telling as well. So far, GPT-5 has been hovering around 7% in its diff edit failure rate – competitive with Sonnet 4 and better than 2.5 Pro.

GPT-5 was built for coding agents

OpenAI trained GPT-5 specifically for coding agents by identifying what developers actually wanted. They boiled it down to four personality traits:

• Autonomy – takes initiative, runs long chains of reasoning and tool calls without getting stuck
• Collaboration – works like a teammate, not just a tool
• Communication – explains its approach when needed, stays quiet when executing
• Context Management & Testing – maintains project context, runs builds and tests before declaring completion

This training shows in practice. GPT-5 doesn't just fix bugs; it checks for lints, realizes which ones are relevant, runs builds, and ensures code is shippable before it's done. All of this is supported by Cline – a harness that is well-suited to squeeze the most out of this model.

Superior Tool Intelligence

GPT-5 sets new records on tool-calling benchmarks, achieving 96.7% on τ-bench telecom -- a benchmark where no model scored above 49% just two months ago. This translates directly to better performance in Cline:

In Cline's Plan Mode, GPT-5 is methodical about understanding your codebase:

• asks specific clarifying questions with options
• maps out dependencies and affected files
• confirms the implementation approach

Then in Act Mode, it goes quiet and ships:

• chains together dozens of tool calls reliably –both in sequence and parallel
• clean diffs that respect existing code style
• multi-file refactors handled systematically
• keeps working through long task lists without interruption

Long Context Excellence

GPT-5 can handle up to 256,000 input tokens with strong performance even at maximum context. On OpenAI-MRCR at 256k tokens, GPT-5 achieves 86.8% accuracy compared to o3's inability to handle that length. This means GPT-5 can persistently churn through longer tasks while remembering details from earlier in the conversation – this validates its personality as "persistent".

Getting the Most from GPT-5 in Cline

The good news: GPT-5 works exceptionally well in Cline out of the box. We've optimized the prompting based on OpenAI's guidance and our testing, incorporating their new features like verbosity control and preamble messages.

Pure, Unfiltered GPT-5

Here's what matters: with Cline, you're getting pure, unfiltered GPT-5. Our open-source architecture means complete transparency – you can see exactly what prompts we use, and there's no hidden layer between you and the model.

Here's what you can expect using GPT-5 in Cline:

• Planning phase – GPT-5 will thoroughly analyze your codebase before making changes. It uses Cline's file exploration tools to understand structure and dependencies.
• Execution phase – 0nce it has a plan, GPT-5 executes efficiently. With its improved tool intelligence, it can chain together complex sequences of file modifications without losing track.
• Complex debugging – the model excels at tracing through code paths and identifying issues. It follows stack traces, examines related files, runs tests, and proposes targeted fixes.
• Long-running tasks – GPT-5 maintains focus across extended coding sessions. With 272k input token support, it can keep entire projects in context while working through task lists systematically.

Pricing That Makes Sense for Development Teams

At $1.25 per million input tokens (with 90% cache discount) and $10 per million output tokens, GPT-5 costs roughly half of what you'd pay for Claude Sonnet 4 ($3/$15).

How to Use GPT-5 in Cline

Getting started is straightforward. Simply select gpt-5 from the model dropdown in Cline's settings from the OpenAI, Cline, or OpenRouter providers. The model is available through OpenAI's API with your standard API key.

Ready to experience the most reliable coding model from OpenAI? Update Cline and switch to GPT-5 today. Share your experiences with complex refactoring and debugging on Reddit or Discord.