Claude Sonnet 5 becomes the first model to pass every task in GitLab Duo Agent Platform's evaluation suite

Anthropic and GitLab have announced that Claude Sonnet 5 is now on GitLab Duo Agent Platform. The model is available across all tiers and all deployment options through GitLab's AI Gateway, and it is the first model in GitLab's evaluation suite to pass every benchmark task.

How much better is Sonnet 5 than the previous model?

GitLab tests models against its own task suite, designed to reflect how AI agents actually work in software development: multi-step tasks, generating code that holds up in code review, and running workflows at an acceptable cost. Claude Sonnet 5 is the first model to complete every task in the suite; its predecessor, Sonnet 4.6, completed 93.8% of the same tasks.

• Claude Sonnet 5 is the first model in GitLab's evaluation suite to pass 100% of the benchmark tasks
• Sonnet 4.6, the previous model, completed 93.8% of the same suite
• The number of resolved issues rose by 8.8%
• The model is available across all tiers and all GitLab Duo deployment options via the AI Gateway

What changes for development teams

For GitLab Duo Agentic Chat users, this changes the familiar prompt–wait–evaluate loop. Multi-file refactors now produce output that's ready for review instead of stalling halfway through. Test generation returns coverage you can actually use. Security investigations dig deeper into repository history. Duo's foundational agents handle more of their assigned work without human intervention, so teams spend their time reviewing results rather than restarting failed tasks. GitLab points to a concrete example of this agentic workflow: in GitLab Orbit, the model is already being used to investigate pipeline failures from the past two months — analyzing the history of broken builds and tracking down the root cause directly in a conversation with the agent.

"Claude Sonnet 5 handled the full range of coding tasks we tested, and it resolves more issues while doing so.

That's a meaningful improvement in both quality and efficiency. We've made it available in GitLab Duo Agent Platform today, across all tiers and all deployment options," said Manav Khurana, Chief Product and Marketing Officer at GitLab.

What it means

The most expensive agent failure is the one that stops halfway: the cost isn't just the lost work, but also the diagnosis, the re-prompting, and checking whatever did come back. A model that can get through GitLab's entire benchmark suite without failing turns an agent from a tool you have to supervise constantly into one you can genuinely delegate work to. That threshold — reliability, not just the quality of an individual response — is where GitLab and Anthropic suggest measuring the progress of agentic models in software development.

FAQ

How does

Claude Sonnet 5 compare with Sonnet 4.6 on GitLab's tests?

Claude Sonnet 5 is the first model to pass every benchmark task in GitLab's evaluation suite, while Sonnet 4.6 completed 93.8% of the same tasks; the number of resolved issues rose by 8.8%.

Which GitLab tiers is Claude Sonnet 5 available on?

The model is available across all tiers and all deployment options of GitLab Duo Agent Platform through GitLab's AI Gateway — no special premium tier was required to access Sonnet 5.

Where is Claude Sonnet 5 already being used in GitLab?

GitLab cites GitLab Orbit, where the model is used to investigate pipeline failures from the past two months — analyzing the history of broken builds directly in a conversation with the agent.