Research investigation R0902 / comparison
The Flash-vs-Frontier Cost Curve
The published token-price gap matters only if Gemini 3.8 Flash preserves completion rates after agent token use, retries, caching, and output length are counted.
This synthesis reassesses saved evidence collected on September 2, 2026. It is not a new collection, calculation, or execution. Google’s stated standard pricing is through December 31, 2026 and is documented to double on January 1, 2027.
This synthesis reassesses saved evidence collected on September 2, 2026. It is not a new collection, calculation, or execution. Google’s stated standard pricing is through December 31, 2026 and is documented to double on January 1, 2027.
- Dataset ID
- spd:the-flash-vs-frontier-cost-curve-does-gemini-3-8-flash-keep-its-advantage-on-matched-coding-work-629530d2
- Stable URL
- /research/the-flash-vs-frontier-cost-curve-does-gemini-3-8-flash-keep-its-advantage-on-matched-coding-work-629530d2
- Version
- Live
- Coverage
- Live collection
- Records
- 0
- Fields
- 7
- Updated
| Entity | Metric | Value | Unit | Observed | Source | Transform |
|---|
Measurement technique
How to read this report
- 01Evidence matrix: listed pricing supports a 13.33-times uncached input and output rate difference and a 3.33-times cache-read rate difference through December 31, 2026; these are not all-in task-cost results.
- 02Evidence matrix: both vendors’ documentation indicates reasoning or effort can affect billed output or token use, so visible answer length is not sufficient for all-in cost accounting.
- 03Evidence matrix: use the same held-out public coding tasks, prompts, tool policy, retry policy, output ceiling, and documented effort-setting protocol.
- 04Record uncached input, output, cached-token use, retries, task success, patch-test results, and wall-clock time; calculate all-in API cost per successful completion only after those measurements exist.
- 05Use a symmetric output ceiling no higher than Gemini’s documented 65,536-token maximum.
- 06Report results by repository and task-difficulty band rather than treating a pooled result as universally applicable.
Sources
Evidence
4 publishers supporting 0 records. Expand a publisher to inspect its cited pages.