Cost Reduction¶
How G6 reduces LLM costs through algorithmic compression.
Status: a design goal of the self-training loop, not yet a verified result
Cost reduction via compression is the intended behaviour of the T2/T3 loop, not a measured outcome. Current single-run benchmarks show G6 costs more per task; the trajectories below describe how the loop is designed to behave once synthesis reuses its distilled library (a capability not yet wired in). What is independently verified today is higher accuracy on hard problems: GAIA L2 +28.8pp (exact McNemar p≈7×10⁻⁵, 95% CI [16.7, 40.9]pp) and Omni-MATH +9.4pp (p≈5×10⁻⁷, 95% CI [5.9, 12.9]pp).
The Economics¶
LLM API calls are priced by token. A complex reasoning task might cost \(0.50-\)2.00 per call. If you solve the same class of problem 100 times per month, that's \(50-\)200/month — for one task type.
How Compression Reduces Cost¶
Week 1: Learning Phase¶
Every problem requires fresh LLM reasoning. Costs are at baseline.
Weeks 2-4: Recognition Phase¶
G6 starts recognising recurring patterns. Some sub-problems are solved with cached algorithms instead of LLM calls.
Month 2+: Compression Phase¶
Most routine work is handled by the component library. LLM calls are reserved for genuinely novel problems.
graph LR
A[Week 1: 100% LLM] --> B[Week 4: 60% LLM]
B --> C[Month 2: 30% LLM]
C --> D[Month 3+: 15% LLM]Designed trajectory once library reuse is active — not yet measured (see the status note above).
Concrete Example¶
Task: Parse and validate JSON schemas from API documentation. (Illustrative of the design target, not a measured run.)
| Period | Approach | Token Cost |
|---|---|---|
| Day 1 | Full LLM reasoning | ~50K tokens |
| Day 5 | LLM + cached parser | ~20K tokens |
| Day 15 | Cached parser + LLM for edge cases | ~5K tokens |
| Day 30 | Standalone algorithm | ~0 tokens |
Cost Reduction by Adaptive Level¶
Cost savings are designed to compound as you move up the T0–T3 hierarchy (the T2/T3 "decreasing" behaviour depends on library reuse that is not yet wired into synthesis):
| Level | Cost Behaviour | Why |
|---|---|---|
| T0 | Flat — every call costs the same | No learning, no caching |
| T1 | Step drop, then flat | One-time harness optimisation reduces token waste |
| T2 | Continuously decreasing | Self-training loop caches more solutions each iteration |
| T3 | Decreasing + compounding | System learns better compression rules — each iteration is cheaper than the last |
G6 Pricing vs. Savings¶
| Tier | Monthly Cost | Typical Token Savings |
|---|---|---|
| Researcher | $20 AUD | Moderate — T0–T1 retrieval + pipelines |
| Builder | $50 AUD | High — T2–T3 evolution + full compression |
The break-even point depends on your usage volume and on the library-reuse loop being active. We do not yet publish a verified payback period — the independently verified benefit today is higher accuracy on hard problems, not lower cost.
FAQ¶
Q: What if my work is always novel? A: G6 still helps through pipeline orchestration and formal verification. Cost reduction from compression is a bonus, not a requirement.
Q: Can I see my compression ratio? A: Not today. We don't publish a per-user compression metric — system_status reports operational status and the available tool list, not usage analytics, and there's no dashboard panel for it. We'd rather say so than point you at a number that isn't there.