The news
Alphabet's (GOOGL) Google will start converting Gemini Gems, the custom assistants users build inside its Gemini app, into Gemini skills on November 17, 2026, according to an in-app notice reported on September 27. Separately, two Google blog posts on September 28 pointed builders toward Gemini 3.8 Flash and open Gemma models.
The notice says Gems will become skills starting on that date and that Google will automatically migrate them, 9to5Google reported, and users can keep using Gems until then. TechCrunch reported that the company said users will not have to do anything to make the transition. Gems launched in 2024 as a way to give the AI standing instructions for tasks such as coaching, brainstorming or editing.
Skills do much the same job, holding custom instructions, but are called up by typing a slash in the prompt box and several can be used at once, according to 9to5Google. The outlet reported that Google introduced skills with Gemini Spark, that Google AI Pro and AI Ultra subscribers can make skills from the Gemini Spark tab, and that the linked help article did not yet work when it checked.
On September 28, a post on The Keyword, Google's main company blog, described Gemini 3.8 Flash, launched earlier in September, as its "most intelligent workhorse model." Google said the model improves on Gemini 3.7 Flash in coding, agent-style work and multistep reasoning, and frequently comes close to pricier frontier models. It is available in Google AI Studio and Google Antigravity, the post said.
The same day, Darren Mowry, Google's vice president for global startups and investor ecosystem, argued on the Google Cloud blog for a compound AI stack, in which startups reserve large frontier models for the hardest reasoning work and hand routine jobs to smaller open-weight models they can customize and host where they choose. He pointed to Gemma 4, released under the Apache 2.0 license in five sizes, and cited Cue, a voice-activated desktop assistant that runs Gemma 4 on local hardware; he said its makers cut latency 44%, from 876 milliseconds to 488, after making Gemma their default engine.
The numbers
- Date Gems start becoming skills
- November 17, 2026
- Gemma 4 model sizes
- 5 (from E2B to a dense 31B)
- Cue latency after moving to Gemma (Google figure)
- 876 ms to 488 ms, a 44% drop
- MedGemma score on MedQA benchmark (Google figure)
- 87.7%
- MedGemma inference cost vs frontier models (Google claim)
- Roughly one-tenth
Why CEOs should care
For chief information officers and IT teams, the Gems change is a small migration with a licensing question attached. Inventory the Gems your staff rely on for recurring work, retest them as skills after November 17 and document the prompts they contain, because any automated conversion may change how they behave. Ask your Google account team whether Workspace business accounts follow the same schedule, since 9to5Google did not address them, and whether creating skills will require AI Pro or AI Ultra seats in your plan.
For CFOs and heads of AI platforms, Google's own messaging now favors matching the model to the job. Mowry wrote that sending high-frequency, structured tasks such as intent routing or JSON extraction to general-purpose frontier endpoints spends capital that could fund product work. His suggested steps are practical for any company: find the three highest-volume, predictable tasks on your most expensive model, test a smaller model on them and measure latency and margin. Treat Google's figures, such as MedGemma's roughly one-tenth inference cost, as vendor claims to verify.
For chief information security officers (CISOs), open-weight models change who carries the risk. Running Gemma on your own hardware keeps data local, and Google cited K-Dense, whose Gemma 4-based scientific assistant runs fully air-gapped for pharma and biotech work, but your team then owns patching, access control and monitoring. Mowry also warned that self-hosting models above 70 billion parameters turns early-stage teams into infrastructure providers, a cost worth pricing before you commit.
The bigger picture
The two moves point in the same direction. On the user side, Google is turning one-off custom assistants into reusable skills that can be stacked. On the builder side, two separate Google posts, read together, sketch a tiered model lineup: Flash as the workhorse, frontier models for the hardest tasks and open Gemma models where cost, speed or data control matter most. TechCrunch noted that even as skills, the feature is less simple than typing a request into a chatbot such as Meta Platforms' (META) Muse, and argued that Google is too quick to give every new AI feature its own brand and a spot in the app's navigation.
What’s next
Migration starts November 17, 2026. Before then, watch for Google's help documentation, which 9to5Google found was not yet live, and for any statement on Workspace accounts and skill-creation limits. Teams building on Gemini should run their own cost and quality tests on Gemini 3.8 Flash and Gemma 4 against current workloads rather than relying on Google's comparisons.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
