Tura v0.1.33 Released and Benchmarked Against Codex CLI

SGM_Finance · reddit · 2026-07-16

Tura has released v0.1.33, a local AGPL-3.0 coding agent.

Key Changes in This Version

Added a macro tool commandrun, allowing models to build inspectable multi-step execution trees.
Support for patches, builds, tests, and parallel steps.
Introduced explicit task states and context compression, retaining code locations, patches, tests, and execution statuses.
Available via CLI, TUI, Web, and Desktop GUI.
Supports multiple model providers, with credentials fully controlled by the user.

Public Benchmark Comparison

The author used DeepSWE v1.1 for 20 tasks, 3 reproduction rounds, totaling 60 runs. All configurations used GPT-5.6 SOL / High reasoning:

Tura Balanced: 48/60 passed, 80.0%, 229.7M tokens, 2,017 rounds
Tura Direct: 39/60 passed, 65.0%, 75.1M tokens, 969 rounds
Official Codex CLI High: 36/60 passed, 60.0%, 455.7M tokens, 6,074 rounds

Author's Relative Results

Balanced used 49.6% fewer tokens and 66.8% fewer rounds than Codex CLI High, with a pass rate 20 percentage points higher.
Direct used 83.5% fewer tokens and 84.0% fewer rounds than Codex CLI High, with a pass rate 5 percentage points higher.

The author cautions that this is strictly a full-configuration comparison and does not prove that any single feature is the sole cause of the performance difference.

Original post →

Tura v0.1.33 Released and Benchmarked Against Codex CLI

More from coding & agent