Freeze the documented 3.8 facts
| Field | Current Google documentation | Migration action |
|---|---|---|
| Model ID | gemini-3.8-flash, listed as stable | Pin the exact ID in test and release records. |
| Inputs | Text, image, video, audio and PDF | Test only the modalities the workflow actually uses. |
| Output | Text with documented 65,536-token limit | Test truncation, schema and export on the real packet. |
| Thinking | Low, medium and high supported; minimal returns an error | Replace unsupported settings and record the chosen level. |
| Tools | Function calling, search grounding, file search and other capabilities are listed | Enable only approved tools and retest every schema. |
Pin the current baseline before editing the request
Save the current model ID, product surface, project, region, request body, system instruction, source packet, conversation history, tool schemas, safety settings, output, latency, token use, cost and reviewer corrections. If the production route uses a moving alias, record the resolved version when the platform exposes it.
Do not change the model, prompt, retrieval logic and tool schemas in one experiment. Isolate the model migration first. A combined refactor can produce a better answer while making the cause impossible to identify or roll back.
Validate parameters against the 3.8 contract
Inventory every request field instead of copying a 3.7 payload and assuming silent compatibility. Google’s 3.8 model page explicitly says the minimal thinking level is unsupported and returns an error. Confirm thinking level, output constraints, structured-output schema, safety settings and any feature-specific fields in the current documentation and in a live non-sensitive request.
Treat an HTTP success as transport evidence only. The response must also satisfy the expected schema, source behavior, refusal and stop conditions. Preserve the exact request and response used for the migration record.
Test multi-turn state and replay
Legal workflows often carry instructions, sources and tool results across turns. Replay a representative conversation from the first user input rather than testing only the last prompt. Verify role order, context construction, required citations, abstention behavior and how the application handles a rejected or malformed turn.
If the application uses stored state, identify which API or customer database owns it and apply the Gemini ZDR feature checklist. A model migration must not silently expand storage or reuse a stateful route that the approved workflow did not include.
Revalidate function calls and tool authority
Google lists function calling as supported for Gemini 3.8 Flash. Support does not prove that an existing tool schema, ordering or retry path behaves the same. Test required and optional arguments, invalid values, duplicate calls, timeouts, permission denial and adversarial instructions inside retrieved material.
Keep tools read-only during the first regression unless the workflow specifically requires a controlled write. A successful tool call must still be authorized, logged and reviewed. The model’s decision to call a function is not authority to change a system.
Run one fixed legal regression set
- Ordinary task with a complete approved source packet.
- Long input near the normal workflow limit.
- Conflicting sources that require an uncertainty note.
- Missing authority that requires abstention or escalation.
- Malformed tool output and a denied permission.
- Adversarial instruction inside a source document.
Score material omissions, unsupported assertions, citation fidelity, tool authority, structured-output validity, reviewer corrections, latency and completed-task cost. A style improvement cannot offset a failed material guardrail.
Pilot and roll back deliberately
Use a bounded pilot with named users, approved data classes, logging and support. Keep the prior model, request configuration and deployment path available. Roll back on a material accuracy failure, schema break, permission expansion, unexpected storage, unusable export, cost boundary breach or review-time regression.
Close the record with the chosen model ID, test date, result, accepted limits, reviewer, release owner and next review trigger. A future model alias or documentation update reopens the migration decision.
FAQ
Is gemini-3.8-flash a stable model ID?
Google’s model page lists `gemini-3.8-flash` as stable as of September 2, 2026. Pin it and preserve the checked date because model catalogs can change.
Does Gemini 3.8 support minimal thinking?
No. Google’s current model page lists low, medium and high and says `minimal` returns an error.
Can a team replace Gemini 3.7 without regression testing?
A stable release is not evidence that a specific legal workflow preserves its sources, tools, schema, data route or review burden. Run the same-task regression first.
Sources checked
- Gemini 3.8 Flash model page, checked 2026-09-03.
- Gemini API model catalog, checked 2026-09-03.
- Gemini 3.8 release announcement, checked 2026-09-03.
- Gemini 3.8 Flash model card, checked 2026-09-03.
- Gemini API Additional Terms, checked 2026-09-03.
Operational information, not legal advice. Verify current terms, account configuration and applicable professional duties before use.