Methodology & Data Sources
Glance compares frontier LLMs by what tasks actually cost to run, not headline $/1M rate cards.
Why List Prices Mislead
Reasoning models generate tens to hundreds of thousands of internal thinking tokens before answering, billed at output rates. A cheaper $/1M model that overthinks can cost several times more per task than a pricier, concise one.
Glance shows list prices beside thinking-vs-answer token splits and measured USD per task so real spend stays obvious.
Two-Layer Data Lineage
Direct Source ↗
Unmodified scores, per-task bills, token counts, and API rates from upstream leaderboards and APIs, always linked with ↗ citations.
Agent Synthesis
Workflow picks (Default / Budget / Max) and delta notes synthesized from the facts layer, each paired with clear caveats.
Inclusion Criteria (Closed + Open)
- No geographic filtering. Closed APIs and open-weights models face the exact same bar.
- Models must place near the top of Artificial Analysis, CursorBench, or Terminal-Bench, or sit on the capability-vs-cost Pareto frontier.
- Snapshots stay capped at 8–14 configurations to remain scannable.
Tracked Labs
| Lab | Weights | Focus |
|---|---|---|
| Anthropic | Closed | High-ceiling reasoning and terminal engineering |
| Closed | Coding cost-efficiency, long context, low hallucination | |
| OpenAI | Closed | Compact reasoning output, low per-task bill, modeling |
| xAI | Closed | Large-context hosted API throughput |
| Meta | Closed | Low-cost agentic Pareto floor and multimodal reasoning |
| DeepSeek | Open Weights | Open-weights per-task cost floor and reasoning efficiency |
| Moonshot AI | Open Weights | Open-weights frontier intelligence and long context |
| Z.ai / Zhipu | Open Weights | Open-weights agentic coding and terminal reliability |
| Alibaba (Qwen) | Open Weights | Open-weights multilingual and coding ecosystem |
Upstream Sources
Artificial Analysis Intelligence Index v5.0.1
Artificial Analysis · v5.0.1
Artificial Analysis Model Leaderboard
Artificial Analysis · v5.0.1
AA-Omniscience v1.1 Hallucination Benchmark
Artificial Analysis · v1.1
CursorBench 4.1
Anysphere · 4.1
Terminal-Bench 4.1
Snorkel AI / Harbor · 4.1
FrontierCode 2.0
Cognition · 2.0
Google DeepMind Model Evaluation Report
Google DeepMind · 2026-11
FAQ
- Why can an $8/$40 model cost less per task than a $2/$10 model?
- Your bill is unit rate times total billed tokens. If a $2/$10 model outputs 285k tokens while an $8/$40 model finishes in 41k tokens, the higher list-price model costs less.
- How is the Pareto frontier calculated?
- At build time, Glance marks every model where no alternative scores equal or higher at equal or lower per-task USD cost.
- How often are snapshots updated?
- New frozen snapshots are published after major frontier releases or notable benchmark and pricing shifts. Older snapshots stay live at /s/YYYY-MM-DD/.