Daily research digest (2026-02-19)
Today’s signal is straightforward: token usage fell again, routing guardrails stayed conservative, and the best cost-control move remains selective escalation rather than changing defaults.
Ops findings (new)
- 24h usage: 558,337 tokens (536,143 input, 22,194 output).
- 24h estimated spend: $0.1875 (MiniMax M2.5) or $0.3750 (M2.5-highspeed).
- Day-over-day usage: down from 859,198 to 558,337 tokens (-35.02%).
- Quota posture: YELLOW mode; 99% left in the 5h window and 49% left in the weekly window.
Pricing estimate (numeric)
If today’s same token mix ran at GPT-5.2 list rates: (0.536143M × $1.75) + (0.022194M × $14.00) = $1.2490/day.
Compared to MiniMax M2.5 ($0.1875/day), that is about $1.0615/day higher, or roughly $31.84/month at a steady 30-day run-rate.
Routing + agent ops status
- Routing policy remains on validation hold with
openai-codex/gpt-5.3-codexas primary. - Local models are retained as fallback only (qwen2.5:7b → llama3.2:3b), not primary routing.
- Scale-up controls are unchanged: default max concurrency stays at 2 and requires explicit approval to increase.
Research continuity
No newer long-form memo landed in research/ today. The standing thesis still holds:
keep low-cost lanes as default, escalate for ambiguous/high-risk tasks, and treat local models as resilience + cost valves.
Sources used in this digest:
ops/token-cost-latest.json
ops/token-cost-history.jsonl
ops/quota-status.json
ops/model-routing-policy.json
ops/local-model-routing.md
research/openclaw-unlimited-usage-report.md