Google presented two new AI models on July 21, 2026: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite replace the previous Flash variants. The output price of Flash drops from nine to $7.50 per million tokens, while Flash-Lite starts at just 30 cents. Both models are said by Google to need fewer computation steps for the same task.
Flash gets faster and cheaper for programming tasks
Gemini 3.6 Flash replaces the previous Flash model as Google DeepMind’s fast all-round AI for programming tasks, research, and image processing. As Google announces in its blog post, the price for output tokens drops from nine to $7.50 per million, while the input price stays at $1.50. According to market researcher Artificial Analysis Index, the model needs 17 percent fewer output tokens for equivalent tasks than its predecessor - so the price cut and the efficiency gain add up to a larger overall cost saving.
In programming tasks, the progress shows up in the DeepSWE benchmark, which measures software-engineering tasks drawn from real GitHub projects: the score rises from 37 to 49 percent of tasks solved. In the OSWorld-Verified benchmark, which tests autonomous operation of desktop programs, Flash reaches 83 percent, up from 78.4 percent. The model’s knowledge cutoff moves from January 2025 to March 2026. Developers reach Gemini 3.6 Flash via Google AI Studio, the Google Antigravity coding environment, and the Gemini Enterprise Agent Platform, while consumers get it through the Gemini app.
Flash-Lite targets high-volume workloads
Gemini 3.5 Flash-Lite is built for tasks that need large amounts of text processed quickly and cheaply, such as automated document review or search agents. The model costs 30 cents per million input tokens and $2.50 per million output tokens. According to the Artificial Analysis Index, Flash-Lite delivers up to 350 tokens per second - enough to generate a full page of text in roughly a second and a half.
In the Terminal-Bench 2.1 benchmark, which tests command-line tasks performed by AI agents, the success rate climbs from 31 to 54 percent compared with the March predecessor, 3.1 Flash-Lite. On the SWE-Bench Pro coding benchmark, Flash-Lite reaches 54.2 percent, edging out the larger, regular Flash model’s 49.6 percent - an unusual result for a lightweight model variant. Beyond the familiar access routes through AI Studio and the Gemini API, Google says it is also rolling the model out in Google Search. For users in Germany and the rest of the EU, both new models are available without restriction, unlike the simultaneously unveiled security variant Flash Cyber, which for now remains limited to governments.
Flagship Gemini 3.5 Pro stays absent as Gemini 4 training begins
One thing was notably missing from the announcement: Google’s flagship Gemini 3.5 Pro, promised for the second quarter of 2026, remains delayed and is currently only in testing with select partners, according to the company. Google gives no new launch date. Instead, the company said it has begun training its next major model, Gemini 4 - describing it as its “most ambitious pre-training run” yet.
The new Flash generation lands amid intensifying price competition among major AI providers: earlier in July, OpenAI introduced three new pricing tiers with GPT-5.6, with the cheapest tier, Luna, priced at one dollar per million input tokens. Google also adds tightened safeguards to Flash 3.6 against misuse for chemical, biological, or nuclear purposes as well as cyberattacks, according to its model report.
What matters now is whether Google can show similar efficiency gains once the overdue flagship Gemini 3.5 Pro ships - its repeated delays increasingly look like a structural problem with Google’s internal coding benchmarks. Until a new date is set, the leaner Flash line remains the only visible progress on Google’s current model roadmap.
