Google almost ready to launch new Gemini AI model, may beat Anthropic and OpenAI in coding this time

Google is reportedly working on a new Gemini AI model that could be released as early as this week. The new model, codenamed Skimaki, may outperform rival models from OpenAI and Anthropic.Google is reportedly planning to launch a new Gemini model as early as this week. (Photo: Reuters)Google has fallen behind OpenAI and Anthropic in the AI race in recent months. But now, the company is planning to make a comeback, particularly when it comes to coding with a new Gemini model. This new Gemini model may bring the company back on track in the AI race, as per a new report.As per a report from Wall Street Journal, Google is preparing to launch a new AI model, internally known as “Skimaki,” with stronger coding performance. Coding is one of the most important business uses of AI, and is largely dominated by Anthropic and OpenAI.This model, the report adds, will be called Gemini 3.8 Flash, and could be released as early as Wednesday – September 2, 2026. Previous reports indicated that Google employees had begun testing Gemini 3.8 Flash internally last month.This release could be crucial for Google. The company has faced delays with its next flagship AI model Gemini 3.5 Pro as it reportedly failed to meet internal coding benchmarks. At the same time, the company has seen changes among its top leadership, with Demis Hassabis stepping aside as Google DeepMind CEO. What is Gemini 3.8 Flash?According to the report, in internal head-to-head tests inside Jetski, Google’s coding platform, some engineers preferred 3.8 Flash to Anthropic’s Opus model. Keep in mind that Gemini 3.8 Flash will likely not be a new frontier model. Gemini’s Flash models are smaller models designed for quicker responses at a cheaper cost.Flash models are usually trained on a couple hundred billion parameters, compared to trillions of parameters when it comes to the most advanced models from OpenAI or Anthropic. Though the smaller size makes them easier to modify, allowing several Google teams to test different approaches at the same time. The report adds that this is an advantage over the Pro models, where updates need far more of Google’s computing resources. Google has been moving quickly when it comes to new Flash models. Gemini 3.6 Flash rolled out in July, followed by 3.7 Flash three weeks later.Google briefly looked like a front-runner in the AI race when it released Gemini 3.0 last November, but it has since lost ground to newer models from Anthropic and OpenAI. The delays around the release of Gemini 3.5 Pro have also made things tougher for the company.As per the report, Google employees claim that internal candidates for Gemini 3.5 Pro were dropped because they were not enough of an improvement over the Flash line. On the other hand, Gemini 4, Google’s next planned flagship model, has performed well in pretraining evaluations but is still completing posttraining.The pressure to deliver has grown inside Google. Google co-founder Sergey Brin urged staff to move faster on Gemini after Anthropic released Claude Mythos in April. At the same time, several high-profile researchers, including Noam Shazeer and Jeff Dean, left earlier this summer. Meanwhile Koray Kavukcuoglu has taken over Demis Hassabis’ day-to-day responsibilities at DeepMind.As per the report, Kavukcuoglu had already been leading day-to-day Gemini decisions since at least last year, with Hassabis spending much of his time on external engagements. Google has also seen strong user adoption for Gemini, with the app crossing 1 billion users globally, but it remains under pressure in the enterprise market, where coding tools and AI agents are now central.Since the start of the year, the company has put more researcher time and computing power into coding performance. Google has also hired Barret Zoph, a former Thinking Machines Lab co-founder and OpenAI posttraining lead, as vice-president of research focused on reinforcement learning and posttraining. As per the report, Google DeepMind teams often work on several model bets at once, so weaker progress on one model does not necessarily reflect what is coming next.Anthropic and OpenAI are not Google’s only rivals in the coding space. In recent weeks, Elon Musk’s SpaceXAI has also released new Grok AI models designed for coding. At the same time, Meta has launched its own AI coding agent Muse Code, developed by Alexandr Wang’s Superintelligence Labs team.- EndsPublished By: Armaan AgarwalPublished On: Sep 2, 2026 10:24 IST