Open Trials · EXPECTED · Q3 2026
Benchmarks On Paper.Measured Numbers Coming Q3.
The competitor numbers below are observable from public documentation, these tools are stateless, so their behavior is predictable. Anvaya’s numbers are projected from the compression ratios measured in our product repos (12-25x per node type), not from a benchmark run. Full reproducibility package arriving end of Q3 2026.
Methodology
How We’ll Measure.
Most benchmarks measure throughput. We measure compounding , whether a tool gets better at your specific codebase the more you use it.
The Industry
Every Tool Is The Same On Day 1 And Day 100.
Claude Code, Aider, Copilot, OpenCode, jcode, and Codex CLI are stateless. They don’t improve with use. The context they need on session 50 is the same as session 1. These numbers are observable, no benchmark required.
* Competitor figures are estimated from public docs and code inspection. These tools have no persistent memory layer, so their behavior across sessions is predictable and stable.
Expected To Improve With Use.
These numbers are derived from the compression ratios measured in our product repos (12-25x per node type, verified against the Mind README) and projected across session milestones. They are on-paper estimates, not benchmark results. Final measured numbers arrive Q3 2026.
Best Case vs Anvaya At Session 50.
The best any competitor achieves is Aider’s ~7% savings at session 10, from a static repo-map, with no learning loop. Anvaya at session 50 is projected to reach ~92% savings from a calibrated, experience-driven graph. These are on-paper estimates based on measured compression ratios.
Static repo-map, no learning
7% savings, flat from here on
Calibrated, experience-driven graph
~92% savings, still improving
* Anvaya projected from 12-25x compression per node type (verified) applied across session milestones. Actual benchmark results pending.
When Do The Real Numbers Land?
We’d rather show you measured numbers than promise them. Here’s exactly what’s coming and when.
Stop Starting From Zero.
One binary. 11+9 Rust crates. 545 tests. Hand-written HNSW index. Three transport modes. Four providers, Ollama, Anthropic, OpenAI, Siemens. Zero API keys required to start. Mind remembers everything after the first session.