The pass rate and cost of Alma by independent developer Yetone are better than Claude Code

The locally developed AI Agent Alma, led by independent developer Yetone, ranked third in the latest round of cross-evaluation initiated by the AI Agent infrastructure company Runta, both in task success rate and cost of success, outperforming tools like Claude Code and OpenCode. The current leaderboard snapshot shows that Alma has a task success rate of 72.7%, just behind Codex CLI at 77.9% and Pi at 72.8%, while Claude Code stands at 70.0%.

Alma costs approximately $0.80 on average for each successfully completed task, compared to Claude Code's $0.94, making it about 16% cheaper. This FrontierHarness v1.1 comparison evaluates the performance of the same large model under different Agent execution frameworks, covering 21 sets of Harness configurations, 5 large models, and 62 tasks, with each task repeated 4 times, planning to run a total of 26,040 times, which is 72 times the scale of the previous version. All tools operate in a unified cloud environment, comparing task success rates, token costs, and execution speeds.

Alma emphasizes local operation and long-term memory, capable of reading and modifying code, executing terminal commands, searching the web, and supporting Skills and MCP, while also allowing for switching underlying models. Recently, Alma integrated with Herdr, allowing programming Agents like Claude Code, Codex, and Pi to continue operating within Herdr, and enabling users to view execution processes, respond to questions, and approve operations in a centralized chat interface. Users can also remotely manage multiple Agents on their computers via iPhone. The evaluation is still ongoing, and rankings may change; the models being tested are limited to open-source models like GLM, DeepSeek, Kimi, and Muse, and do not cover closed-source models from the GPT and Claude series.