Update README.md

Author: Yohjisakamoto

README.md

@@ -18,7 +18,7 @@ Tura reduces model round trips and repeated context through its runtime and macr
 
 ### Benchmark
 Long-horizon task [benchmarks](https://turaai.net/benchmark) are one way to measure coding-agent performance beyond isolated prompts. The published comparison uses harness-based development tasks with archived prompts, per-round tool calls, token usage, patches, and verifier results.
-> On 20 DeepSWE v1.1 tasks run three times per agent (180 sessions), Tura Balanced passed 48/60 task verifiers (80.0%) versus Codex CLI's 38/60 (63.3%), a 16.7 percentage-point improvement. Tura Direct used 75.1M aggregate tokens versus Codex CLI's 333.5M, or 77.5% fewer, while passing 39/60 verifiers (65.0%), 1.7 points above Codex CLI. In short: Balanced used 31.1% fewer tokens and improved verifier success by 16.7 points; Direct used 77.5% fewer tokens and improved verifier success by 1.7 points.[^debug-figure][^debug-manifests]
+> Across 20 DeepSWE v1.1 tasks run three times per agent, Tura first creates a substantial token-budget advantage by reducing repeated context and model round trips. Users can then choose how to spend that advantage: Direct converts most of it into lower cost, using 77.5% fewer aggregate tokens than Codex CLI while achieving a comparable verifier success rate of 65.0% versus 63.3%; Balanced reinvests part of the saved budget into deeper reasoning, investigation, and verification, reaching an 80.0% success rate—16.7 percentage points higher than Codex CLI—while still using 31.1% fewer tokens.[^debug-figure][^debug-manifests]
 
 The published artifacts compare the named Tura Balanced, Tura Direct, and Codex CLI configurations on the same benchmark tasks.[^debug-figure]