What’s Under the Prompt?
Inspect context, cache reuse, output, and tools in a sample session—or your own.
Illustrative sample · 3 sessions · GitHub Copilot CLI + Codex CLI + Claude Code · Captured calls only; completeness cannot be inferred
Context, call by call
Hover, tap, or focus the chart and use arrow keys to inspect each call.
- Cached input
- Fresh input
- Output
Token signature
What stands out
- 55% of input was served from cache: most turns re-read context the model had already seen.
- The model read 15.8 input tokens for every token it wrote.
- 20% of output was reasoning rather than visible text or tool calls.
Show more findings
Same task. Lower cost?
Same token workload, different prices. Quality may differ. API estimates—not your Copilot bill.
Assumptions
One run repeats the entire loaded workload, even if it contains several sessions. Token counts and cache reuse stay fixed; models may tokenize differently, need more attempts, or not support the same context and tools. Lines show midpoint estimates; ranges show pricing uncertainty. Savings exclude switching costs and are not measured profit.
Every model, same workload
Provider-published API rates, last checked . Recorded models are highlighted.
Exact rates
| Provider | Context / rate notes |
|---|
| Model | Estimated total | Input | Cached input | 5-minute write | 1-hour write | Output |
|---|
Rates are USD per 1 million tokens. A “+” on a token total means some records did not report that category. Cache writes without a reported TTL are shown as a cost range when the provider has different TTL rates. Prices exclude batch discounts, regional uplifts, taxes, storage fees, and tool fees unless noted. Reasoning tokens are included in output.
How each model was used
| Model | Calls | Input tokens | Cached | Output tokens |
|---|
Tools the agent reached for
Coverage notes
Analyze your own session
Drop JSON or JSONL logs from Copilot, Codex, Claude Code, or OpenTelemetry. Files are processed locally.
Sample report loaded. Add a session to replace it.
Supported files and paths
- GitHub Copilot CLI
~/.copilot/session-state/<session>/events.jsonl- Codex CLI
~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl- Claude Code
~/.claude/projects/<project>/<session>.jsonl- OpenTelemetry
- GenAI chat and tool spans as OTLP JSON
Having trouble finding the session file?
These folders are hidden. On macOS, press ⌘⇧G in
the file picker and paste the folder path; on Windows, paste the path with
%USERPROFILE% in place of ~. Codex and Claude Code
files can be dropped directly. Copilot names each session folder by ID, so these
helpers list IDs and dates, then copy the chosen raw file locally.
bash copilot-export.sh --list
bash copilot-export.sh --session SESSION_ID --output copilot-events.jsonl.\copilot-export.ps1 -List
.\copilot-export.ps1 -Session SESSION_ID -Output copilot-events.jsonlReview raw logs before sharing; they may contain private conversation or tool data.
Enter token counts instead
No session log? Enter counts to compare model prices.