AI Launch Profile
Gemini changes sampling and turn-validation rules for new models
Google deprecated temperature, top_p and top_k for Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and later Gemini models. The named models ignore those parameters; future generations will return HTTP 400, while requests ending in a non-empty model turn already return HTTP 400.
At a glance
Launch Snapshot
- Company
- Launch date
- July 21, 2026
- Launch type
- Major Update
- Category
- AI Developer Tools
- Audience
- AI Engineers, Developers
- Pricing
- The compatibility changes have no separate charge. Both named models offer free-tier input and output tokens. Paid standard rates are $1.50 per million input tokens and $7.50 per million output tokens for Gemini 3.6 Flash; Gemini 3.5 Flash-Lite costs $0.30 and $2.50 respectively.
- Free plan
- Yes
- API
- Yes
- Open weights/source
- No
Launch Context
Use these links to move from this record into the broader Launch Intelligence database.
Verification & Sources
- Status
- Verified
- Source links
- 2
- Freshness
- Verified July 27, 2026
- Last verified
- July 27, 2026
- Last updated
- July 27, 2026
Key source checks
Suggest a correction
Creator Coverage Next Steps
This launch has signals that may support demos, reviews, creator education, founder storytelling, or practical product explainers.
Launching an AI product that needs clear demos, creator education, and buyer trust? Sponsor a Kingy AI video or launch feature.
Kingy AI Take
Remove these legacy controls before migrating to Gemini 3.6 Flash, Gemini 3.5 Flash-Lite or later releases. Sampling parameters are ignored on the current named models and may fail on future generations, while a request ending in a non-empty model turn fails with HTTP 400 now.
Who it is for
Developers, AI engineers, and platform teams migrating Gemini API workloads to Gemini 3.6 Flash, Gemini 3.5 Flash-Lite or later models, especially applications that override sampling or prefill model turns.
What feels promising
Google separates the two behaviors and provides replacement patterns: use system instructions when sampling overrides were shaping output, and use the Interactions API or structured outputs instead of prefilled model turns.
What feels unproven
The guidance does not show whether those replacements reproduce existing output distributions, formatting or determinism for a given workload. Migration tests need separate assertions for ignored parameters, future parameter errors and current turn-validation failures.
Source-backed record