Gemini's new Flash models arrive cheaper, and the API drops temperature and top_p
Google shipped three Flash-tier Gemini models on July 21, all aimed at cost and throughput rather than a frontier leap. The quieter change: the latest models ignore temperature, top_p, and top_k, and future generations will reject them outright.
Google released three Gemini models on July 21, and all of them point the same direction: cheaper, faster, and tuned for agents that run many turns rather than a single frontier leap. The headline is Gemini 3.6 Flash, joined by a faster 3.5 Flash-Lite and a security-specialized 3.5 Flash Cyber. For developers, the release carries a quieter change that is easy to miss in the benchmark tables: the latest Gemini models no longer honor the sampling parameters that many pipelines still send by default.
What shipped
Gemini 3.6 Flash is the workhorse. Google prices it at $1.50 per million input tokens and $7.50 per million output, and claims it cuts output token usage by about 17% against 3.5 Flash on the Artificial Analysis Index, with larger reductions on coding-agent workloads. On the benchmarks Google published, it posts 49% versus 37% on code-editing precision (DeepSWE) and 83.0% versus 78.4% on computer-use tasks (OSWorld-Verified). For most teams the token-reduction number matters more than the accuracy points: on a long-running agent, output tokens are where the bill compounds, so a model that says the same thing in fewer tokens lowers cost without any change to the prompt.
Gemini 3.5 Flash-Lite is the throughput play, priced at $0.30 per million input and $2.50 per million output and clocked by Artificial Analysis at roughly 350 output tokens per second. Google says it beats the older 3.1 Flash-Lite on Terminal-Bench 2.1 (54% versus 31%) and edges past the larger Gemini 3 Flash on some tasks. That last claim is the more interesting one: a cheaper, faster model matching a bigger one is the kind of trade that changes which tier a team reaches for by default.
Gemini 3.5 Flash Cyber is the outlier. It is a fine-tune aimed at finding and patching security vulnerabilities, and it does not ship to the general API. It runs only inside CodeMender, Google Cloud’s vulnerability-remediation agent, in a limited-access pilot for governments and trusted partners. For most developers it reads as a signal rather than a tool: Google is packaging model plus agent harness for security work instead of selling the raw capability.
The through-line is efficiency. None of the three is a bid for the capability crown; all three are bids for the cost and latency budget of production agents.
The API change developers have to make
Alongside the launch, Google’s model documentation now states that temperature, top_p, and top_k are “deprecated and ignored” on the latest models, and that “in future model generations, supplying these parameters returns an HTTP 400 error.” The change covers Gemini 3.6 Flash, 3.5 Flash-Lite, and everything after. Prefilled model turns, a common trick for forcing a specific output format, are no longer allowed either.
The practical effect is that code setting temperature to zero for reproducible output, or tuning top_p to rein a model in, is now silently doing nothing on these models, and will start failing outright on the next generation. Google’s guidance is to remove the parameters and reach for a system instruction with explicit rules, or structured outputs, when you need determinism or a fixed format. Teams that have wrapped Gemini behind a provider-agnostic SDK should check how that layer forwards sampling settings, because a request that still carries top_k will keep working today and break on the model after this one.
It is a small migration, but it is the kind that surfaces as a production incident months later, when a parameter everyone assumed was a no-op suddenly returns a 400.
The missing piece is Pro
The gap in the lineup is Gemini 3.5 Pro, the high-capability tier that would compete directly with Claude Opus 4.8 and GPT-5.6. It was last updated in February. Google product lead Logan Kilpatrick said the team is testing 3.5 Pro with partners and expects it to “land soon,” and noted that work on Gemini 4 has begun. As TechCrunch reported, Bloomberg had earlier flagged internal delays tied to unmet performance targets.
For a team choosing a frontier model today, the read is straightforward: Google is competitive on the efficient, high-throughput tier and is not currently contesting the top of the reasoning market. That is a defensible place to ship from, given how much real agent spend lives in the Flash tier, but it leaves the highest-capability slot to Anthropic and OpenAI for now.
What’s worth watching
Three signals over the next quarter:
- Whether 3.5 Pro actually ships, and where its price and capability land against Opus 4.8 and GPT-5.6.
- Whether the sampling-parameter deprecation spreads. If Google is comfortable dropping temperature and top_p in favor of system instructions and structured outputs, other labs may follow, which would make prompt code less portable across providers than it is today.
- Independent cost numbers on 3.6 Flash. The 17% token-reduction figure is Google’s own; field billing on real agent workloads is the test.
The larger pattern is that Google is optimizing the part of the market where agents actually run up a bill, and tightening the API contract while it does. Stackmaven’s follow-up coverage will revisit when 3.5 Pro ships.