Google Releases Gemini Flash Models Amid Pro Delay

Gemini 3.6 Flash replaces Gemini 3.5 Flash, which Ars Technica reported had already been deprecated. Google said 3.6 Flash improves coding, knowledge-work, and multimodal performance while using up to 17% fewer tokens than its predecessor, according to TechCrunch and CNBC.

Ars Technica reported Google benchmark figures showing 3.6 Flash at 49% on DeepSWE, versus 37% for 3.5 Flash, and 83% on OSWorld computer-use testing, versus 78.4%. Google also made computer use a standard Gemini API capability for 3.6 Flash, Ars reported.

The other releases target more specific deployment profiles:

• •Gemini 3.5 Flash-Lite is aimed at high-volume, lower-complexity work. CNBC described it as Google's fastest and least expensive model in the 3.5 family.
• •Gemini 3.5 Flash Cyber is fine-tuned to identify and remediate software vulnerabilities. TechCrunch and CNBC reported that access initially is limited to governments and trusted partners through a limited-access pilot.
• •Gemini 3.6 Flash is the general-purpose model for coding, multimodal, and agentic workloads where latency and token consumption are material operational constraints.

VentureBeat reported API pricing of $1.50 per million input tokens and $7.50 per million output tokens for 3.6 Flash, compared with $1.50 and $9.00, respectively, for 3.5 Flash. It reported pricing of $0.30 per million input tokens and $2.50 per million output tokens for 3.5 Flash-Lite.

Reuters reported that 3.5 Pro's delay was linked to the model falling short of internal goals, particularly in coding. TechCrunch likewise reported that Google had previously indicated the Pro model was in internal use and expected the following month, but the July release did not include it.

The absence matters because Pro models are Google's highest-capability Gemini tier for more complex reasoning and coding workloads, TechCrunch reported. Reuters framed the delayed release as closely watched by Wall Street and by observers assessing whether Google DeepMind can keep pace with OpenAI and Anthropic in frontier-model development.

For ML teams, the immediate practical change is not a new flagship capability tier but another cost-performance option for production routing. Organizations running multi-step agents can evaluate whether the reported token reduction, computer-use support, and coding benchmark improvements translate to lower end-to-end task cost in their own traces. Comparable model transitions commonly require testing at the workflow level, since benchmark gains and per-token pricing do not by themselves determine tool-call reliability, retry rates, latency, or total agent spend.

Google has not publicly provided a release date for Gemini 3.5 Pro in the reporting cited here.