Real tasks. Measured evidence.
Compare quality, speed and cost. Inspect the evidence, then rerun the same task yourself.
Prepare a three-model comparison without signing up. Use trial quota or your own API keys.
How a comparison works · demo
Quality is marked unassessed until an explicit task check passes.
2026-10-31 · Animated explainer
26 animated diagrams of Tao Zhang's essay: Mono, the bug-fix loop, the merge gate, the 200-item weekend, PMs shipping code, and what isn't working yet. Plain-language rewrite, not a speed benchmark.
Open the diagram guide →2026-10-31 · Architecture diagram explainer
Animated diagrams of Router/Worker, triggers, and durable sessions—from public discussion, not a speed benchmark.
Open the diagram guide →2026-10-30 · Open-topic knowledge videos
Same Claude Code conditions. Watch both models expand a topic, then judge for yourself. Unified edge-tts narration.
Open the compare page →2026-10-27 · Creative agent tests
Watch the actual outputs and inspect editable scenes, scientific explanations, dimensions and failures.
View outputs and commentary →2026-09-27 · 15 attempts · hkg1 · Retest pending
JSON extraction, instruction constraints and grounded QA — pass counts, failure cases and original outputs.
Inspect results →Use a versioned check or inspect the original output.
Keep provider, region, settings and measurement window explicit.
Export results and return to the same report version.