Independent developer Yetone's Alma has broken into the top three of Agent, surpassing Claude Code in both pass rate and cost.

Beating AI News Flash: Alma, led by independent developer Yetone, has secured third place in both task pass rate and cost per success in the latest AI Agent benchmark, surpassing well-known tools such as Claude Code and OpenCode.

This evaluation was initiated by AI Agent infrastructure company Runta. In the current leaderboard snapshot, Alma achieves a task pass rate of 72.7%, trailing only Codex CLI at 77.9% and Pi at 72.8%. Claude Code stands at 70.0%.

Alma not only has a higher pass rate but also completes tasks more cost-effectively. It spends an average of approximately $0.80 per successfully completed task, compared to $0.94 for Claude Code—about 16% lower in cost.

This edition of FrontierHarness v1.1 specifically investigates one question: with the same large model, how much can performance differ when switching to a different Agent execution framework?

The evaluation covers 21 Harness configurations, 5 large models, and 62 tasks, with each task repeated 4 times, planning for 26,040 runs—72 times the scale of the previous edition. All tools run in a unified cloud environment, comparing task success rates, token costs, and execution speed.

Alma is an AI Agent focused on local operation and long-term memory. It can directly read and modify code, execute terminal commands, search the web, and supports Skills and MCP. It can freely switch underlying models and does not depend on any single model vendor.

Recently, Alma also integrated Herdr. Coding Agents such as Claude Code, Codex, and Pi can continue to run within Herdr, while Alma consolidates them into a single chat interface for viewing execution processes, answering questions, and approving operations. Users can even remotely manage multiple Agents on their computers via iPhone.

The evaluation is not yet complete, and rankings may still change. The models tested are also limited to open-source models such as GLM, DeepSeek, Kimi, and Muse, and do not cover closed-source models from the GPT and Claude series.