SHANTI GREENE
Tokenomics
I think therefore
I'm expensive.
ABOUT ME
Shanti Greene
AI & data executive. Builder. Educator.
Head of Data Science & AI Innovation
AnswerRocketAdjunct Professor, Graduate Analytics
Washington University in St. LouisCTO (Fractional)
DriveLogicFounding Partner
ZenithLogix
TOKENMONSTER / EXPERIMENT 01
Anthropic models
Claude Code · High effort · One pilot per model
| Model | Total tokens | Build time | Est. API cost | Passed / 27 |
|---|---|---|---|---|
| Claude Haiku 4.5 | 11.65M | 13.7 min | $1.68 | 0 / 27 |
| Claude Sonnet 5 | 129.42M | 76.5 min | $30.58 | 8 / 27 |
| Claude Opus 5.5 | 47.04M | 66.3 min | $21.22 | 8 / 27 |
Opus used 64% fewer tokens than Sonnet
Both passed 8 of 27 requirement groups. No Anthropic pilot qualified.
Estimated API equivalents, not subscription charges. Total tokens include cached input. Haiku failed; Sonnet and Opus completed. Remaining groups include failures and unevaluated checks.
TOKENMONSTER / EXPERIMENT 01
OpenAI models
Codex CLI · High effort · One pilot per model
| Model | Total tokens | Build time | Cache share¹ | Passed / 27 |
|---|---|---|---|---|
| GPT 5.6 Terra | 4.02M | 18.2 min | 96.7% | 3 / 27 |
| GPT 6 Sol | 22.42M | 39.3 min | 99.1% | 3 / 27 |
| GPT 6 Luna | 39.78M | 62.0 min | 98.7% | 3 / 27 |
9.9× the tokens. Each passed 3 of 27 groups.
Luna versus Terra in these pilots. All three failed the security gate.
¹ Cache reads ÷ total input tokens. Costs unreported for these subscription runs. Total tokens include cached input. No OpenAI pilot qualified. This is a pilot observation, not a model ranking.
TOKENMONSTER / EXPERIMENT 01
Open weight models
OpenCode · API runs · Evaluation pending
| Model | Tokens so far | Elapsed | Effort | Evaluation |
|---|---|---|---|---|
| Qwen3.8-27B | 215K | 3.6 min | High | Pending |
| DeepSeek 4.1 Flash | 374K | 3.6 min | High | Pending |
| MiMo v2.6 Pro | 314K | 3.6 min | Default | Pending |
The experiment is still running
All 27 requirement groups await evaluation for each model.
Partial telemetry at the snapshot time. Qwen uses the dashboard’s “Stock Qwen counterpart” row, identified by its telemetry model ID. Costs and qualification are pending. Runs used hosted APIs, not local inference.