01What Is Claude Sonnet 5?

Claude Sonnet 5 is the latest model from Anthropic, released as an upgrade to Sonnet 4.6. It is supposed to be a more agentic model that can plan, use tools like browsers and terminals, and run autonomously. But early tests show it underperforms compared to Anthropic's own Opus 4.8.

Julian Goldie ran several benchmarks and found Sonnet 5 scoring lower on almost every task while costing more to use. The model is available through the Claude API and Claude Code, but many users are questioning why it was released in its current state.

02Main Features of Sonnet 5

Claude Sonnet 5 brings a few improvements over Sonnet 4.6. It can create plans, use browser terminals, and run tasks autonomously at a level that previously required larger and more expensive models. It supports agentic coding and can generate visual projects like games and graphics.

However, the actual performance gains are small. On GoldieBench, Sonnet 5 scored 63% on agentic coding, while Opus 4.8 scored 69%. The model also produces outputs with darker colors and less detail compared to fused models like Sakana Fugu.

The main selling point was supposed to be speed and cost, but early tests show it is slower and more expensive than Opus 4.8.

03How to Set Up Claude Sonnet 5

  1. Go to the Anthropic website or open Claude Desktop / Claude Code.
  2. Select the Sonnet 5 model from the model dropdown menu.
  3. If you are using an API, update your endpoint to point to the Sonnet 5 model ID.
  4. Adjust your usage settings if you want to limit costs — Sonnet 5 is more expensive than Opus 4.8.
  5. Test a simple prompt first to confirm the model is responding correctly.
  6. If you have an agent operating system, you can plug Sonnet 5 in, but Julian recommends sticking with Opus 4.8 for now.

04How to Use Claude Sonnet 5 With Examples

Julian tested Sonnet 5 on several tasks to see how it performed. He asked it to create a raycaster maze game. The output looked good — nice colors and smooth gameplay.

However, the controls felt backwards when he played it. He also tested an orbit simulation for a galaxy, and it failed completely — the planets didn’t orbit. Another task was a synthwave background, which came out fine.

He also built a crypt game that looked smooth and easy to use. For comparison, he gave the same prompts to GLM 5.2. The maze from GLM 5.2 was buggy and not smooth.

But the orbit test worked perfectly on GLM 5.2, while Sonnet 5 couldn’t do it at all. That shows Sonnet 5 is inconsistent — good at some things, bad at others.

05Results, Tests and Comparisons From GoldieBench

Julian compared Sonnet 5 against Opus 4.8, GLM 5.2, and Sakana Fugu Ultra on GoldieBench. Here are the key findings:

  • Agentic coding: Sonnet 5 scored 63%, Opus 4.8 scored 69%.
  • On every single benchmark shown, Opus 4.8 outperformed Sonnet 5.
  • Sonnet 5 costs 1.2x more than Opus 4.8, and up to 5x more than GLM 5.2.
  • The orbit test: Sonnet 5 failed; GLM 5.2 and Sakana Fugu both succeeded.
  • Sakana Fugu’s outputs had more interesting colors and details than Sonnet 5.
  • A liquid-in-a-bowl simulation from Sakana Fugu looked great, but Sonnet 5’s version looked worse.
  • Crypt game: Sakana Fugu produced a more visually interesting result than Sonnet 5.

06Tips, Limits and Mistakes to Avoid With Sonnet 5

Based on Julian’s tests, here are the main things to watch out for when using Sonnet 5:

  • Don’t expect it to be faster or cheaper than Opus 4.8 — it is actually more expensive and often worse.
  • It fails on physics-based tasks like orbit simulations, so avoid using it for such projects.
  • The model’s token efficiency is poor, meaning you spend more money for less output quality.
  • Do not assume it is a direct upgrade to Sonnet 4.6 — the improvements are marginal.
  • Ignore hype from the announcement; test it yourself on your specific use case before switching.
  • Best advice: keep using Opus 4.8 until Fable 5 arrives.

07Who Is Claude Sonnet 5 For?

Claude Sonnet 5 is for developers and users who want to test the latest model from Anthropic. However, based on Julian’s tests, it is not recommended for most people. If you are building serious projects, stick with Opus 4.8 — it is better and cheaper.

If you are experimenting, you can try Sonnet 5 for creative tasks like game graphics or synthwave backgrounds, but avoid it for anything requiring reliable physics or complex logic. The real lesson is to focus on building flexible systems that let you swap models. Julian recommends using an agent operating system that can plug in any model so you are never stuck when a new release underperforms.