Google Gemini in 2026: Models, NotebookLM and Workspace Integration

AI models in this article
Gemini is Google’s generative AI ecosystem, and by February 2026 it had become the default tool for companies that run on Google’s services. Its edge comes from two things: it works inside Google Workspace (Docs, Sheets, Gmail, Drive) and it pulls current data from Google Search.
The flagship models take up to 2 million tokens of context – enough to analyze years of correspondence with a client or a product’s entire technical documentation in one go.
Think of Gemini as an interface to your own data. It sees changes in your spreadsheets and documents as they happen and takes part in work where the work already lives – in Docs, Sheets and Gmail.

Gemini Models (September 2026)
From the release of 3.5 Flash in mid-May to September 2 is 106 days. In that window Google shipped four more Flash models: 3.5 Flash-Lite, 3.6 Flash, 3.7 Flash and 3.8 Flash. The flagship Gemini 3.5 Pro is not on that list. Pichai promised it for June; in September it is still “coming soon.” According to The Wall Street Journal, the internal flagship candidates didn’t beat Flash by enough to justify a launch.
For anyone buying, that leads to one concrete conclusion. If you need a frontier model specifically from Google, there is nothing to buy right now. Google’s strongest model in the API came out in February; its newest one costs a fraction of that and covers most day-to-day work.
| Date | Release |
|---|---|
| July 21 | Gemini 3.6 Flash and 3.5 Flash-Lite go GA: better token efficiency, stronger planning in code and agentic workflows. The temperature, top_p and top_k parameters are deprecated. |
| July 30 | Gemini Robotics ER 2 in preview. |
| August 13 | Gemini 3.7 Flash GA – three weeks after 3.6, at half the introductory price. |
| August 26 | Gemini 3.5 Transcribe and Transcribe Live GA – speech recognition in 85+ languages. |
| August 27 | Gemini Omni Flash GA (gemini-omni-1.1-flash) – video generation and editing. |
| September 1 | Agentic video understanding for 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. |
| September 2 | Gemini 3.8 Flash GA and 3.8 Flash Cyber. |
| September 3 | Lyria 3.5 in public preview – full song generation. |
Gemini 3.8 Flash (September 2, 2026)
The model isn’t in our benchmark yet, so the figures below come from Google and independent leaderboards.
Google calls it its best model for reasoning and code at the same speed and cost as 3.7 Flash. Pricing is $0.75 per 1M input tokens and $3.75 per 1M output tokens – but that is an introductory rate. It runs until December 31, 2026, and Google hasn’t published the regular price from January 1. A 2027 plan built on today’s price is a bet that the price won’t move.
On the Artificial Analysis Intelligence Index the model ranks seventh out of 250+, with a score of 47 at a blended price of $0.74 per 1M tokens. On DeepSWE v1.1 – a test of autonomous engineering tasks – it beats most larger frontier models, per the same source. The index is published in versions and rescaled from time to time, so only numbers from the same snapshot are comparable.
One caveat worth passing on to your team before rollout: on real tasks a 3.8 Flash request costs about 40% more than the price list suggests, because the model produces more output tokens and takes more agentic steps. $0.75 is the price of a token. The price of a task is higher.
Speed is 334 output tokens per second, against 40–90 for the rest of that index’s top ten. That line never makes it into a procurement report. In practice it decides whether people read the answer to the end or switch to another tab.
On the Arena.ai text leaderboard (formerly LMArena), 3.8 Flash is eighth at 1,494 ± 9. The top nine overlap within their confidence intervals – statistically a tie. Why that keeps happening is covered in AI Benchmarks Are Losing Their Meaning, and the scoring mechanics in a separate piece. What makes eighth place interesting is the company it keeps: a model at $0.74 blended sits alongside flagships from Anthropic and Meta that cost several times more.
Where to get it: the Gemini app for Pro and Ultra subscribers, AI Mode in Google Search, and Google Sheets. For developers – AI Studio, Android Studio and Google Antigravity. For companies – Gemini Enterprise.
Gemini 3.8 Flash Cyber
A build for security work. On CyberGym it solves more than 70% of tasks that involve finding real vulnerabilities across 20 programming languages – Google’s numbers, not independently verified. It isn’t on general sale: it is distributed only through the Fairwind program, for governments and critical infrastructure operators. You can’t plan a commercial project around it.
3.7 Flash, 3.6 Flash and 3.5 Flash-Lite
3.7 Flash shipped on August 13, three weeks after 3.6, at half its predecessor’s introductory price. Google claims gains in debugging, incident resolution and first-try code accuracy. 3.6 Flash and 3.5 Flash-Lite launched on July 21, straight to GA.
None of the three is in our benchmark yet; the figures above are the vendor’s.
Four Models in 106 Days: the Operating Cost
Release cadence is a cost line that doesn’t appear on the price list. Prompts tuned on 3.6 Flash aren’t guaranteed to behave the same way on 3.8: both the models and the API change. Along the way Google deprecated temperature, top_p and top_k – the parameters teams normally use to lock in the style and predictability of answers.
Standardizing a team on Google’s models today means standardizing on a lineup that changes every few weeks. What works is splitting a project into two layers: workflows and prompts that you test and keep on your side, and the model underneath, which you can swap in a day. Before every swap, rerun your task set and compare it with the previous run.
Gemini 3.5 Flash (May 2026)
Gemini 3.5 FlashDigs into details
score 7.7/10 rank 23 of 47 $0.027 per task VPN needed
Tested in the mysummit.school benchmark: May 2026About the model →Compare →
Gemini 3.5 Flash was the headline of Google I/O 2026 (May 19). It shipped straight to GA (General Availability), with no preview period. Google pitched it as a Flash model with Pro-level intelligence.
Key characteristics:
- Speed – ~289 tokens per second, roughly 4× faster than comparable frontier models
- Context – 1M tokens of input, up to 65,536 tokens of output
- Multimodal – text, images, audio, video and PDF on input (but no image or audio generation)
- 4 thinking levels – minimal, low, medium (the new default) and high. They replace the old numeric
thinking_budget - Thinking context retained between turns in a conversation (on by default)
- Built-in tool orchestration – Google Search, Maps, URL context, code execution and custom functions run in parallel within a single API call
According to Google and independent leaderboards, Flash leads on agentic and multimodal tasks but trails GPT-5.5 on reasoning (ARC-AGI-2) and hard coding tasks (SWE-bench). We don’t copy external benchmark scores here – they go stale fast. The current comparison with competitors lives in the model card above and in the comparison block below.
Via the API, 3.5 Flash costs more than the previous Gemini 3 Flash and less than Gemini 3.1 Pro – current rates are in the model card above. Subscriptions for end users didn’t change. Since the fall, 3.5 Flash is no longer the best value in the lineup: 3.8 Flash is half the price on both input and output.
Gemini 3.5 Pro: Promised, Not Shipped
Google showed its main model at the same I/O on May 19 and promised it “in June.” In September there is still no public release: Pro is available to Google employees and a small circle of enterprise testers, and the official status is “coming soon.”
What’s promised once it ships:
- 2M-token context and a Deep Think mode for the hardest reasoning.
- Pro will take over work that used to require the expensive Ultra tier: deep multimodal analysis, very long context, heavy reasoning.
- Expected API price is around $15 / $60 per 1M tokens, roughly 10× 3.5 Flash. It will reach the Pro and Ultra subscriptions first.
You can’t plan a rollout around that date, because the vendor won’t name one. If you need a frontier model by a specific deadline, you’ll have to buy it from someone else. If the deadline can wait, the sensible move is to build the workflow on 3.8 Flash and treat the move to Pro as a separate task with its own retesting budget.
Gemini 3.1 Pro
Gemini 3.1 ProFollows clear tasks
score 7.3/10 rank 30 of 47 $0.027 per task VPN needed
Tested in the mysummit.school benchmark: May 2026About the model →Compare →
Gemini 3.1 Pro (February 2026) is Google’s flagship for complex planning and multimodal analysis. Compared with 3.0, it gained a Long-Context Reasoning architecture: across the full 2M-token context it connects scattered documents with, by Google’s measurements, almost no loss of accuracy. As of September 2026 it is the strongest model Google sells through the API – and it is already seven months old.
What changed for managers:
- In Deep Think 2.0 mode the model cross-checks its conclusions against external sources in three stages, which makes it more reliable for financial and legal audits.
- 3.1 Pro can carry out multi-step tasks in Workspace on its own, for example: “Collect all invoices from my inbox for the quarter, reconcile them against the expense sheet in Sheets, and draft a report in Docs highlighting the discrepancies.”
- It can work through hours of conference recordings and pick up not only what was said but the emotional tone of the participants (sentiment analysis).
Gemini 3 Flash
Gemini 3 FlashFollows clear tasks
score 6.8/10 rank 40 of 47 $0.001 per task VPN needed
Tested in the mysummit.school benchmark: May 2026About the model →Compare →
A very fast model with a 1M-token context. It uses 30% fewer thinking tokens than earlier versions. Good for routine work: summarizing long meetings, scanning the news, processing large volumes of incoming messages. It trades depth of analysis for speed.
Flash-Lite: the Cheapest Models in the Lineup
Gemini 3.1 Flash LiteNeeds supervision
score 6.2/10 rank 44 of 47 $0.000 per task VPN needed
Tested in the mysummit.school benchmark: May 2026About the model →Compare →
3.1 Flash-Lite has been in our benchmark for a while; on July 21 it was joined by 3.5 Flash-Lite, released as GA, which we haven’t tested yet. Flash-Lite is the pick when the cost of a thousand processed requests matters more than the quality of any single answer: triaging incoming messages, first-pass labeling, rough summaries. Since September 1 all three Flash models support agentic video understanding, so even the cheapest model in the lineup can now work through recordings.
Gemini 2.5 Pro
The previous flagship, still relevant thanks to its price–quality balance. This is the model that, with Deep Think, won a gold medal at the 2025 International Mathematical Olympiad. It is often the default model in enterprise subscriptions.
Gemma 4 – Open Models (April 2026)
Google’s new generation of open models, built on Gemini 3 research. Four sizes: E2B and E4B for mobile devices, 26B MoE (Mixture of Experts, 128 small experts – only 3.8B parameters are active per token) and 31B Dense for servers and workstations.
Gemma 4 31BFollows clear tasks
score 6.9/10 rank 38 of 47 $0.001 per task VPN needed
Tested in the mysummit.school benchmark: May 2026About the model →Compare →
Gemma 4 26B MoEFollows clear tasks
score 6.6/10 rank 42 of 47 $0.001 per task VPN needed
Tested in the mysummit.school benchmark: May 2026About the model →Compare →
What’s different from Gemma 3:
- Up to 256K tokens of context on the server models (up from 128K), 128K on mobile
- Native video and image handling – OCR, chart analysis, variable resolution
- A big jump in quality – on AIME (math reasoning) and LiveCodeBench (coding), the fourth generation beats the third several times over, per Google
- Apache 2.0 license – Google dropped its own Gemma license. The terms now match Qwen and Mistral: no restrictions on commercial use or redistribution
- On Arena AI both larger versions sit near the top of the open-model list, and 26B MoE outranks models with 20× more parameters
Gemma 4 is the choice when data can’t go to the cloud: confidential documents or a requirement to keep data in-house. E4B runs on a smartphone; 31B runs on an ordinary 16 GB laptop when quantized.
Video, Audio and Transcription: August and September
Until September 1 the model processed a video recording whole: every frame and every second of audio went into the context, and you paid for all of it. Agentic video understanding for 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite works differently – the model goes into the recording selectively and pulls a transcript, a single frame or a slice of audio as needed. On long content Google promises up to 88% fewer tokens.
That goes straight into the budget. An hour-long recording of a client call or a training course used to cost as much as all of its footage. Now you pay for the parts the model actually looked at. Test the claim on your own recordings: 88% is the ceiling for long content, not an average you can expect across your tasks.
Gemini 3.5 Transcribe and Transcribe Live (August 26, GA) are a separate speech recognition product: 85+ languages, speaker separation, word-level timestamps and a custom vocabulary of up to 1,000 terms. For transcribing calls, interviews and field research it removes half the manual work, and the custom vocabulary also fixes the biggest source of errors: product names, surnames and internal acronyms start coming out right.
Gemini Omni Flash (August 27, GA, gemini-omni-1.1-flash) covers video generation and editing: extending a clip, transitioning between two images, resolutions from 360p to 4K. Lyria 3.5 (September 3, public preview) generates full songs from text and images in 44.1 kHz stereo.
Visual Models
Veo 3.1 is a frontier model for generating video and dynamic presentations. Version 3.1 adds “interactive directing”: you can adjust details of the video in real time.
Gemini 3 brought much better image generation, good enough to produce realistic infographics on request.


A generated clip or infographic looks ready to publish. Checking it is not nearly as fast: the model confidently draws numbers that were never in your data and adds details you didn’t ask for. The more convincing the output, the more a missed error costs – and the more the skill of framing the task matters.
NotebookLM: Answers From Your Own Documents
NotebookLM is arguably Google’s most powerful tool for deep work with knowledge. Unlike a regular chat, it works only with the documents you upload (“grounding” mode): answers are based on your materials and don’t go beyond the notebook.

For a manager, that means:
- Audio Overview (AI podcasts) – turn a dull 50-page annual report or company strategy into a 10-minute conversation between two experts you can listen to on the commute.
- Feedback analysis – upload your engagement survey results (eNPS) and NotebookLM will surface hidden patterns: “What are people in marketing actually worried about?”
- Team knowledge base – build an onboarding guide that’s easy to extend: new hires ask about internal processes and get answers that point to specific clauses in your policies.
Custom Gems: Gemini for a Specific Role
Gems are customized versions of Gemini tuned to a specific role. You explain the Gem’s job, tone and context once, and it’s ready to work.

Gems a team lead might set up:
- Onboarding Buddy – explains company culture and first tasks to new hires.
- Feedback Coach – helps a manager prepare for a difficult conversation by reframing criticism as constructive recommendations.
- Interview Assistant – generates interview questions from the job description and the candidate’s résumé, then helps compare answers across candidates.
The Gemini App: September 2026
The wave of app updates shown at Google I/O on May 19 has reached users by September. For managers, the agentic features are the most interesting part.
- Daily Brief – a personal morning summary: Gemini pulls the key items from your email, calendar and tasks and lays out the day’s priorities. Rolling out to Google AI subscribers in the US.
- Gemini Spark – a personal agent that runs in the background 24/7, even with your phone locked, and sees your workflows through to the end. Launching for Ultra subscribers.
- Gemini Omni – video generation from text, audio and visual prompts in Google Flow and YouTube Shorts; since August 27, Omni Flash is GA, with editing and clip extension.
- The new Neural Expressive interface – answers no longer look like a wall of text: the key point sits at the top in bold, and details unfold as you scroll.
- 3.5 Flash by default – since June 8, 3.5 Flash is the default model in the Gemini Enterprise app, and it can’t be turned off. With 3.8 Flash out, the model lineup in the enterprise environment is being updated, so before investigating “the answers got worse,” check which model is actually responding.
Daily Brief plus Gemini Spark moves Gemini from a chat window to a background assistant: the morning summary shows up on its own, and routine workflows finish without you. The quality of the result, though, depends entirely on how precisely you described the workflow.
Why Managers Choose It
- Workspace integration – Gemini is built into Google Docs and Gmail: it finds the right email, summarizes a long thread in Chat, pulls figures from a spreadsheet and drops them straight into a draft document.
- Huge context (2M tokens) – you can upload the video of a long workshop or a quarter’s worth of meeting recordings and ask: “Summarize every agreement where budgets came up.”
- Up-to-date data – thanks to its direct link to Google Search, Gemini knows about the latest changes in employment law or job-market trends “as of today”: the data comes from search, not from the training cutoff.
- Google Vids – automatically produces internal training videos and team clips from text scripts and materials in Drive.

Practical Scenarios
- Gemini 3.8 Flash: the go-to model this fall – document analysis, agentic workflows (parallel search plus data processing in one request), fast multimodal processing. Budget about 40% above list price: for the same task the model spends more output tokens.
- Gemini 3.5 Flash: still a solid all-rounder for the same tasks, but since September choose it deliberately – 3.8 Flash is cheaper and, according to Artificial Analysis, stronger.
- Gemini 3.1 Pro: competitive analysis (web research plus PDF reports), Performance Review prep (pulling an employee’s achievements from Docs and Gmail over six months), auditing the security of your team’s processes.
- Gemini 3 Flash: instant summaries of missed Meet calls, team-building ideas, sorting and prioritizing your inbox. With 3.5 and 3.8 Flash out, it mostly makes sense where minimum cost is critical.
- Agentic video understanding: reviewing recordings of meetings, client calls and training courses – the model picks the frames and transcript fragments it needs instead of digesting the whole video.
- Gemini 3.5 Transcribe: transcribing interviews and calls with speaker separation and a custom vocabulary.
- Project Mariner (Ultra plan): autonomously managing calendars, booking meeting rooms and organizing team travel with complex constraints.
- NotebookLM: an interactive digest of the quarterly strategy for the whole team.
The scenarios above make it look as if all you need is to pick a model and paste a prompt. The gap opens later: a summary you can show your CFO and a summary that’s “roughly on topic” are asked for differently. The difference between “the model answered” and “the result is usable at work” comes down to how the task is framed – and that is a skill you build on your own tasks.
Try It Yourself: Onboarding Plan Plus Employee Development
You
You are an HR business partner at an IT company (300 people). Draw up a 90-day onboarding plan for a new team lead joining from another company. ## Context - Team: 6 developers, 1 QA engineer, 1 analyst - 2 people on the team also applied for this role - A new product launches in 60 days - The team lead has a strong technical background but no experience managing people ## Task 1. Break the 90 days into phases with concrete goals for each week 2. For each phase: which skills the team lead needs to develop and how to measure them 3. How to handle the two "passed-over" candidates – concrete steps in the first 2 weeks 4. What early signals will show that onboarding is going off track 5. Account for the product deadline: what to do if learning conflicts with the release
What to look for: planning is Gemini’s strong category in our benchmark, and learning and employee development is its weak one. The model will put together the week-by-week structure confidently and handle the development program worse than its competitors. DeepSeek V4 Pro is worth a look as a budget alternative with comparable planning. GPT-5.6 usually goes deeper on the people-management side.
Results in Our Benchmark
We tested Gemini models on real management tasks – planning, communication, teamwork, learning and development, and regional awareness. The results are mixed.
Google’s current workhorse – Gemini 3.5 Flash – is already in the benchmark; its current score, rank and price per task are in the model card above. The fall releases aren’t in our data: 3.6, 3.7 and 3.8 Flash will appear once they’ve been run on the same task set, and everything said about them above relies on Google, Artificial Analysis and Arena.ai.
- Information search7.3
- Data analysis7.4
- Decisions under uncertainty7.8
- Feedback analysis8.2
- Corporate requirements and risks7.3
- Automation7.2
- People management8.2
- Russia and Kazakhstan specifics7.8
The radar shows the model’s profile across eight categories: the Gemini lineup is strong in planning and communication, while learning and employee development consistently lag.
Gemini 2.5 Pro is a solid mid-table performer. It handles planning well and knows regional context surprisingly well. But on learning and employee development it posts one of the weakest results among the major models. Communication is average.
2.5 Flash is nearly on par with the Pro version, which is notable in itself: the speed-optimized model gives up very little quality. Its regional knowledge is just as good. Team management, on the other hand, is a failure – one of the worst results in the test.
The Gemini 3 Flash paradox: newer doesn’t mean better. On management tasks it is weaker than both 2.5 versions. It does reasonably on communication and planning, but learning and team management are its weak spots.
The pattern holds: people buy Gemini for the ecosystem. On the quality of answers to management tasks, competitors deliver more – you can see it in our cards and in the independent rankings above. On access to Docs, Sheets and Gmail, Google has no rival. If you’re choosing Gemini for the “smartest” AI, you’re overpaying for something that isn’t its strength. If you’re choosing it to work with your team’s documents, it is still the best option.
Gemini Pricing (September 2026)
- Free: access to Google’s Flash models (with daily limits), basic image generation, Google Docs/Drive integration.
- Google AI Pro ($19.99/month): full access to current models, including 3.8 Flash and 3.1 Pro, an extended context window (1M+), Gemini in Workspace, 2 TB of storage.
- Google One AI Premium ($100/month): a new tier introduced with 3.5. Higher limits, priority access to new Flash models and the upcoming 3.5 Pro.
- Gemini Enterprise ($30/user/month): enterprise security, priority model access, the ability to build your own agents, and the Google Antigravity 2.0 platform for IT infrastructure automation.
- Google AI Ultra ($199.99/month): down from $249.99. Maximum limits, access to Veo 3.1, unlimited Deep Think, the Project Mariner agent for automating browser actions, and exclusive access to Gemini Spark (the background agent).
The API picture is separate and harder to plan around. Until December 31, 2026, 3.8 Flash costs $0.75 / $3.75 per 1M tokens, and that is an introductory rate: Google hasn’t published the regular price. 3.7 Flash launched at half the price of 3.6. You can’t build an annual budget on terms like these, and that applies to the whole lineup.
The price–quality map from our benchmark shows where the Google models we’ve tested sit. The fall releases aren’t on it: 3.8 Flash, at $0.74 blended and 334 tokens per second, would have to be drawn in by hand from Artificial Analysis data.
Limitations
- Ecosystem lock-in – Gemini delivers its full value only inside Google Workspace. If your company runs on Microsoft 365, many of the advantages disappear.
- No flagship – Google’s strongest API model came out in February, and the vendor won’t say when the next one arrives. For frontier-level work, Google isn’t an option right now.
- Temporary prices – 3.8 Flash’s introductory rates run until December 31, 2026, there is no regular price list yet, and the real cost per request is about 40% above list.
- Release cadence – four Flash models in 106 days and deprecated sampling parameters: you’ll have to retest prompts with every move.
- Cost of the Ultra plan – at $199.99/month it’s one of the most expensive tiers on the market, justified only for heavy use of Veo 3.1, Project Mariner and Gemini Spark.
- Languages other than English – answer quality in other languages (we tested Russian) still trails English, though the gap narrows with each update.
- Availability – current access status for each model is in the model cards above, and some app features, such as Daily Brief, reach US subscribers first. If Google doesn’t fit your requirements, compare DeepSeek and GLM-5.
This fall the story has two open questions: will 3.5 Pro ship before year-end, and what will 3.8 Flash’s regular rates look like from January 1. Both affect your budget more than any leaderboard position.


















