Gemini 3.6 Flash turns model migration into an API-behavior audit

Google’s July 21, 2026 release of Gemini 3.6 Flash and Gemini 3.5 Flash-Lite is a model-migration event, but the operational risk for agent teams is request semantics: Google’s latest-model migration guide says temperature, top_p, and top_k are now deprecated, ignored by the API for these models, and will become HTTP 400 errors in future model generations.

That is a bigger practical change than a new model string normally implies. A team can update gemini-3.5-flash to gemini-3.6-flash, rerun a benchmark, and see acceptable latency, cost, or task-completion numbers. But if the same request still carries sampling controls that production code assumes are active, the migration has already changed behavior. The service may continue returning responses, yet a guardrail the application owner thought was shaping randomness is no longer doing anything. Later, the same stale field becomes an availability problem when a future model rejects it.

Google announced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber in a July 21 post. The first two are the relevant public API releases for most developers. Google’s developer guide lists both Gemini 3.6 Flash (gemini-3.6-flash) and Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite) as generally available and ready for production use in the Gemini API. The same guide says the Interactions API is generally available and recommends it for access to the latest features and models.

The pitch from Google is straightforward. Gemini 3.6 Flash is the new workhorse Flash model, aimed at complex agentic, coding, knowledge-work, and multimodal tasks. Google says it reduces output token usage compared with Gemini 3.5 Flash and costs less per output token. The developer documentation lists Gemini 3.6 Flash pricing at $1.50 per million input tokens and $7.50 per million output tokens, with a default thinking level of medium. In the launch post, Google compares that output price with Gemini 3.5 Flash’s $9.00 per million output tokens, while input pricing remains $1.50 per million tokens.

Gemini 3.5 Flash-Lite is positioned differently. Google calls it the fastest, lowest-cost model in the 3.5 family for high-throughput execution, document processing, routing, classification, and subagent-style work. The latest-model guide lists its default thinking level as minimal and its price as $0.30 per million input tokens and $2.50 per million output tokens. Google’s launch post says Artificial Analysis measured it at 350 output tokens per second. Those are the sorts of figures that will push teams to run migration tests quickly, especially where a model sits behind repeated extraction, routing, or tool-dispatch calls.

The migration guide’s API section is the part that deserves equal attention. Google says that, starting with Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, two API changes apply to these models and to all future Gemini model releases. First, temperature, top_p, and top_k are deprecated. The guide says the API ignores those parameters and that future model generations will return an HTTP 400 error when they are supplied. Second, prefilled model turns are no longer supported: if the last non-empty turn in the request is a model turn, the API returns a 400 error.

For agent builders, the sampling-parameter change is not a cosmetic cleanup. Temperature and nucleus or top-k sampling fields often appear in shared wrappers, YAML presets, SDK helper functions, evaluation harnesses, and per-task policy files. Sometimes they are explicit product controls: use a lower temperature for extraction, a higher one for brainstorming, or a narrower sampling policy for customer-facing support. Google’s new guidance means those fields are not levers on the new models. If a workload’s deterministic behavior improves or worsens after migration, the explanation will not be “we kept temperature at 0.2.” On these models, according to the guide, that field is ignored.

That can create silent regressions in systems that do not treat the request payload as part of the migration surface. A customer-support agent may have a compliance-focused prompt and a low-temperature setting in its configuration. A code-modification agent may keep a strict sampling profile for patch generation. A document-extraction system may depend on retry logic tuned around older response variance. After migration, teams should assume the system instruction, structured-output schema, tool contract, conversation state, and thinking level are the live control surfaces—not the deprecated sampling fields.

Google gives an explicit replacement direction for determinism: define a system instruction with rules for the use case. That is a different engineering practice from turning a numeric knob. A system instruction can say the model must output only valid JSON, must not include commentary, must cite fields by source span, or must ask for missing required inputs. But those instructions must be tested as behavioral contracts. For extraction workloads, that means schema validation, invalid-output counts, and fixture-based regression checks. For coding agents, it means measuring unwanted file changes, failed test runs, extra tool loops, and patch reversions. For routing, it means confusion matrices and escalation behavior, not just aggregate success rates.

The prefilled-turn rule is the second migration trap. Older Gemini API patterns and raw REST payloads sometimes ended a conversation with a model role turn to force a prefix such as “Translation:” or to suppress a preamble before a JSON object. Google’s guide now says requests ending with a non-empty model role turn are disallowed and return HTTP 400. The recommended migration is to use system_instruction in the Interactions API or structured outputs instead of manually prefilling a model turn.

This matters because agent failures often hide in edge payloads rather than the happy-path demo. A backend may build a normal prompt for most calls, then add a prefilled model turn only for translation, strict JSON, or assistant-message continuation. A chat product may preserve historical roles and accidentally send the last stored assistant turn as the final non-empty turn. A tool runner may replay traces into a new model without normalizing the final role. Those cases should be found with payload linting before production traffic reaches the new model.

The new models also introduce choices around thinking configuration. Google’s migration checklist for Gemini 3.6 Flash says to replace thinking_budget with the string enum thinking_level, set to medium or high, and to remove candidate_count, which it says is unsupported in Gemini 3.x. For Gemini 3.5 Flash-Lite, Google recommends leaving thinking_level at minimal for high-volume extraction, routing, or classification, while setting it to medium or high for autonomous subagents with tool calls, code execution, or multi-step reasoning. That is another reason a migration should be treated as an API-behavior audit: performance changes may come from model capability, but also from different defaults and different accepted configuration fields.

Practitioners should start with an inventory, not a leaderboard. Search every call path that can reach the Gemini API for temperature, top_p, top_k, candidate_count, thinking_budget, and handcrafted model role turns. Do not stop at application code; check prompt registries, saved evaluation configs, notebooks, background jobs, customer-specific overrides, and SDK wrapper defaults. Then add a request-shape test that fails the build if deprecated sampling parameters or terminal model turns appear in a payload targeting gemini-3.6-flash or gemini-3.5-flash-lite.

Next, split migration testing by workload class. For Gemini 3.6 Flash, test the agent tasks where Google is claiming the model is strongest: coding loops, multimodal reasoning, and multi-step workflows. Google reports DeepSWE at 49% for Gemini 3.6 Flash versus 37% for Gemini 3.5 Flash, MLE-Bench at 63.9% versus 49.7%, OSWorld-Verified at 83.0% versus 78.4%, and GDPval-AA v2 at 1421 versus 1349. Those are useful orientation points, not substitutes for local evals. If your agent’s risk is unwanted repository edits, measure unwanted edits. If your risk is tool-loop spiraling, measure loops and tool-call counts. If your risk is malformed structured output, measure parse failures.

For Gemini 3.5 Flash-Lite, separate throughput economics from autonomy. Google’s documentation presents it as a low-latency, cost-effective model for high-volume execution, and the launch post cites 350 output tokens per second. That makes it attractive for routing, extraction, and subagent fan-out. But the migration guide also says more complex subagent work may need medium or high thinking levels to prevent premature tool termination. A cheap model can become expensive if it fails early, retries often, or hands bad intermediate state to a master agent.

The release is still a meaningful product update. Google’s model pages list both models with text, image, video, audio, and PDF inputs; text output; a 1,048,576-token input limit; and a 65,536-token output limit. The launch post says Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are available starting July 21 for developers in the Gemini API through Google AI Studio and Android Studio, with Gemini 3.6 Flash also available in Google Antigravity. It also says both are available for enterprises through Gemini Enterprise Agent Platform, and that consumer availability includes the Gemini app, with Flash-Lite rolling out in Google Search.

But the safest read is that the models are only half the migration. The other half is the contract between your application and the API. Remove dead sampling controls now. Replace hidden prefill tricks with explicit system instructions or structured outputs. Move thinking configuration onto supported fields. Add linting so future model generations cannot turn ignored parameters into surprise 400s. Then evaluate the real workload with the actual request shape you intend to ship.

In other words: treat Gemini 3.6 Flash as a reason to audit behavior, not just swap a name. The teams that get the most from the release will be the ones that notice which controls no longer control anything.

Sources

• Google: Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
• Google AI for Developers: Using the latest Gemini models
• Google AI for Developers: Gemini 3.6 Flash model page
• Google AI for Developers: Gemini 3.5 Flash-Lite model page
• Google AI for Developers: Gemini Developer API pricing
• Google DeepMind: Gemini 3.6 Flash model card
• Google DeepMind: Gemini 3.5 Flash-Lite model card

---

Shadowfetch is an independent software company publishing evidence-based journalism. Explore Shadowfetch Linux — our own Linux build — and the Shadowfetch apps on the App Store.