Gemini 3.6 Flash Hits Antigravity, 17% Fewer Tokens [2026]

Google now lists Gemini 3.6 Flash as a selectable model in Google Antigravity, with the Gemini API documentation giving the model value as `gemini-3.6-flash` and confirming it supports the Antigravity agent inside Gemini Managed Agents. Google states the model used 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index and takes fewer reasoning steps and tool calls on multi-step workflows, describing it as a workhorse for coding, knowledge work and multimodal tasks. Google's pages do not establish the testing window that press reports from nokiapoweruser.com, Tom's Guide, Android Authority and TechRadar described, so the claim that the model had been testing for several days remains unconfirmed.

Google now lists Gemini 3.6 Flash as a selectable model in Google Antigravity, according to the Antigravity model documentation https://antigravity.google/docs/models? gl=1%2A3t0ldl%2A up%2AMQ..%2A ga%2AMTAyMTc5MTQ4My4xNzYzNTU3NzE5%2A ga 47V54ZJ3EV%2AczE3NjM1NTc3MTkkbzEkZzAkdDE3NjM1NTc3MTkkajYwJGwwJGgw . Coverage from nokiapoweruser.com, Tom’s Guide, Android Authority and TechRadar framed the listing as a model that surfaced in the agent platform and had been testing for several days. That framing is useful context, but Google’s own pages do not establish a testing window, so this analysis treats the “testing for days” part as a report rather than a fact. What Google does state is more interesting than the timeline. It describes the model as a workhorse for coding, knowledge work and multimodal performance, and it says the model used 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. For anyone running agents against Flash-tier models, that single efficiency claim matters more than the leak-style story around it. This piece sorts what is confirmed from what is reported, explains why a token-efficiency number can matter more than a benchmark headline, and lays out how developers can test the model before they commit a pipeline to it. For the wider Gemini picture, our earlier look at Gemini 4 Argon reaching AI Mode in 19 countries https://shattered.io/gemini-4-argon-ai-mode-19-countries-2026/ shows how many model lines Google now runs at once. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F What Google Has Actually Confirmed About Gemini 3.6 Flash Start with the short list of facts that Google’s official material supports. Google Antigravity officially lists Gemini 3.6 Flash as a model. Google’s pages state the model is available in Antigravity. The company describes it as a workhorse for coding, knowledge work and multimodal tasks. It reports that the model used 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index, and it says the model takes fewer reasoning steps and fewer tool calls on multi-step workflows. The Gemini API documentation adds two technical details. The model value for API calls is gemini-3.6-flash , and the model supports the Antigravity agent inside Gemini Managed Agents. Google also publishes a Gemini 3.6 Flash model card https://deepmind.google/models/model-cards/gemini-3-6-flash/ on the DeepMind site, which is where safety and evaluation notes for the release live. Google also mentions Google Antigravity 2.0 in its product material as a platform used with the new Flash model. That pairing signals where Google expects the model to earn its keep: inside agent loops, not in one-shot chat prompts. | Claim | Status | Basis | |---|---|---| | Gemini 3.6 Flash is listed in Google Antigravity | Confirmed | Google’s official pages | | Model value is gemini-3.6-flash | Confirmed | Gemini API documentation | | 17% fewer output tokens than Gemini 3.5 Flash | Google’s claim | Artificial Analysis Index, per Google | | Fewer reasoning steps and tool calls in multi-step workflows | Google’s claim | Google’s product description | | Supports the Antigravity agent in Gemini Managed Agents | Confirmed | Gemini API documentation | | The model “just appeared” in Antigravity | Unconfirmed | Press reports only | | The model had been testing for days | Unconfirmed | Press reports only | Why the “Just Appeared” Story Spread So Fast Model names that show up in a product picker before a formal announcement draw attention for a simple reason. They reveal Google’s roadmap a few days early. When a name like Gemini 3.6 Flash surfaces in a developer tool, enthusiasts check the dropdown, compare it with the old entry and write it up. That pattern explains why four separate outlets covered the same listing within a short window. The “testing for days” angle fits the same pattern. Labs often expose a new checkpoint to a slice of users or to select tools first, then widen access. Google has not said it did that here, and no official page gives a start date, a duration or a rollout schedule. Readers should treat any exact day count they see as an outlet’s estimate. The distinction matters for practical decisions. If a model has been quietly serving traffic, developers may already have seen its behavior without knowing it. If it only just landed, the first week of real usage will define its reputation. Either way, the confirmed facts above are what teams should plan around, and the rest is color. The 17% Output Token Claim, Read Carefully Google states that Gemini 3.6 Flash used 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. Two things are worth separating here. The first is what the number measures. The second is what it does not. What the number measures Output tokens are the tokens a model generates, including any visible reasoning it produces before an answer. On an index run, a model that reaches the same or better answers with fewer generated tokens finishes faster and costs less per task, assuming per-token output pricing stays constant. A 17% drop is a relative figure against the predecessor on one benchmark suite, so it describes that suite’s workload. What the number does not tell you It does not state a price. The fact sheet for this story contains no confirmed pricing, plan availability or launch date, so any claim about savings in dollars would be invented. It also does not tell you how your own prompts behave. Agent workloads with long tool outputs, large code diffs or heavy retrieval can show a very different ratio from a standardized index. A hypothetical shows the arithmetic. If a job on the older model generated 10,000 output tokens, a 17% reduction would bring it to 8,300. That is an illustration of the math, not a measured result for any real workload. Teams should replace it with their own logs. Fewer Reasoning Steps and Tool Calls in Multi-Step Workflows The second efficiency claim may carry more weight for agent builders. Google says Gemini 3.6 Flash uses fewer reasoning steps and fewer tool calls for multi-step workflows. In an agent loop, each tool call adds latency, each reasoning step adds tokens, and each extra round trip adds a chance for the run to drift off course. Consider a coding task that reads three files, runs a test, edits a function and reruns the test. A model that plans the sequence cleanly might finish in six calls, while a more hesitant one might take ten. The cost difference compounds across hundreds of runs a day. Google does not publish a per-workflow call count in the material we reviewed, so the gain remains a stated direction rather than a measured figure. This is also where reliability shows up. Fewer tool calls mean fewer chances to hit rate limits, malformed arguments or timeouts. If the claim holds up under independent testing, it would make Flash-tier agents cheaper and steadier at once, which is a rarer combination than raw speed. How Antigravity 2.0 Fits the Picture Google’s product material names Antigravity 2.0 as a platform used with Gemini 3.6 Flash. The Wikipedia entry on Google Antigravity https://en.wikipedia.org/wiki/Google Antigravity offers background on the product line for readers who want the overall history. Reports around Google I/O 2026 described Antigravity 2.0 as an agent-oriented development environment, with a workspace for running several agents in parallel and for scheduling background tasks. That design explains why a Flash-tier model is the natural default. Parallel agents multiply token spend, so the cheap, quick tier of a model family does most of the work while a larger model handles planning or review. When the Flash tier gets better at finishing tasks in fewer steps, the whole orchestration pattern gets cheaper without any change to the platform itself. Google has been pushing agent governance alongside agent capability. Our report on Google Chrysalis and its five guardrails https://shattered.io/google-chrysalis-governed-self-evolving-ai-agents-2026/ covers the safety side of that push, and it pairs naturally with an agent platform that will run unattended tasks. Gemini Managed Agents and the Antigravity Agent The Gemini API documentation says Gemini 3.6 Flash supports the Antigravity agent in Gemini Managed Agents. In plain terms, developers can call a hosted agent through the API instead of building their own harness, and the new Flash model is among the models that can sit under it. Research gathered for this story, drawing on Google’s developer pages, describes Managed Agents as a public-preview capability that can reason, use tools and run code in an isolated environment. We treat those details as reported rather than confirmed for the 3.6 Flash release, because the fact sheet only guarantees the model’s support for the Antigravity agent. The practical takeaway is that the model reaches developers through two doors. One is the Antigravity product, where users pick it from a model list. The other is the API, where the gemini-3.6-flash value routes calls to it. Having both doors open at once reduces the lag between a consumer-facing listing and programmatic access, which is useful for teams that prototype in the app and then script the same behavior in production. The Model String and What Developers Should Check Today The documented identifier is gemini-3.6-flash . A minimal request looks like any other Gemini API call, with the model value swapped in. The snippet below shows the shape of a call using the Python SDK. Treat it as a template and confirm parameter names against the current documentation before you ship. python from google import genai client = genai.Client reads the API key from your environment response = client.models.generate content model="gemini-3.6-flash", contents="Summarize the failing test output and propose a fix.", print response.text print response.usage metadata compare output token counts across models Research for this story also pointed to Google release notes saying calls to the older gemini-3.5-flash value are routed to the new model after the old one’s deprecation. We could not tie that behavior to the fact sheet, so check the changelog yourself. If it holds for your project, your production traffic may already be running on 3.6 Flash, which makes a before-and-after comparison urgent. Pin the model value explicitly in config, log it with every request and watch your token counts for a shift. The same Google documentation family includes a page covering what is new in Gemini 3.6 Flash and 3.5 Flash-Lite, which you can read in the Gemini API what’s-new page https://ai.google.dev/gemini-api/docs/generate-content/whats-new-gemini-3.6?hl=id . Note that the version we verified is the Indonesian-language rendering, so switch the language setting in your browser for English text. What Remains Unconfirmed Several details readers will want are simply not established by the official material behind this story. There is no confirmed testing start date, no testing duration and no formal launch date. There is no confirmed price, no confirmed plan availability, and no named Google employee quote tied to the listing. We found no verified expert quotes for this topic, so this article carries none. That gap is normal for a model that surfaces in a tool before a launch post. It also means any article that gives a precise rollout date or a per-million-token price should show its source. If the source is a screenshot of a pricing page that later changes, the number deserves suspicion. Benchmarks deserve the same caution. Google cites one index and one relative figure. It does not, in the material we reviewed, give coding scores, multimodal scores or latency numbers that we can repeat here. Independent evaluators will fill that in over the coming weeks, and early results often move as labs tune serving stacks. Historical Context: How Fast the Flash Tier Turns Over Google’s Flash line has become its most frequently refreshed tier. Reporting from Google I/O 2026 put Gemini 3.5 Flash at the center of the company’s agent push, and by October the successor name was already visible in a developer tool. The spacing between those two releases is a few months at most, which tells you how quickly a “current” model becomes the previous one. That cadence shapes how teams should build. A model name hard-coded into a dozen services becomes technical debt within a quarter. Teams that wrap model choice in a single config value, log it per request and keep a small regression suite can swap in a new Flash version within a day. Teams that do not often discover the change when a bill or a behavior shifts. The industry shows the same churn elsewhere. GitHub recently moved its assistant off older models, as covered in our piece on GitHub Copilot retiring 6 models and shifting to GPT-5.6 https://shattered.io/github-copilot-retires-6-models-gpt-5-6-2026/ . Retirement notices are now routine, and a deprecation window is the norm rather than the exception. Market Impact: Cost, Speed and the Agent Race Efficiency claims on the cheap tier hit a market that is sensitive to AI spend. Investors have been jittery about capital expenditure, as our coverage of Nasdaq futures dropping 1.7% on AI capex fears https://shattered.io/nasdaq-futures-drop-ai-capex-fears-oil-112-2026/ showed. A model that finishes tasks with fewer tokens eases the pressure at the customer end, even if it does nothing for the vendor’s infrastructure bill. For developers, the buying logic is straightforward. Agent workloads multiply calls, so a few percentage points of token savings matter more than they do in chat. If Google’s 17% figure holds for your traffic, the effect on a monthly bill scales with volume. If it does not, the model still has to justify itself on quality and speed. For Google, the bigger prize is stickiness. When Antigravity, the Gemini API and Managed Agents all default to the same Flash model, developers who build there tend to stay. Distribution through Google Cloud’s expanding catalog reinforces that, a trend we traced when OpenAI joined Google Cloud’s 200-model Garden https://shattered.io/openai-joins-google-cloud-model-garden-2026/ . Competitive Comparison: Flash-Tier Models Across Vendors Every major lab now fields a fast, cheap tier meant for agents and high-volume coding. Anthropic’s agent work, covered in our report on Claude Duet running on Opus 5.5 https://shattered.io/claude-duet-anthropic-ai-agent-opus-5-5-2026/ , leans toward a premium model for harder jobs. OpenAI’s lineup has been rearranged around GPT-5.6 in tools like Copilot. Google’s answer to both is to make the small model good enough that fewer jobs need the large one. Price comparisons across vendors move quickly, and our analysis of the Claude, ChatGPT and Gemini price gap hitting 20x https://shattered.io/claude-vs-chatgpt-vs-gemini-price-gap-2026/ shows how wide the spread can get. We are not repeating any per-token prices for Gemini 3.6 Flash because none are confirmed. The honest comparison today rests on categories rather than numbers. Google leads on integration, since the model sits in its own agent product, its API and its managed agent service. Rivals lead on mindshare in some coding circles. Whether Gemini 3.6 Flash closes that gap depends on independent coding evaluations that have not yet landed. How to Test Gemini 3.6 Flash Yourself Vendor claims are a starting point. A small harness will tell you more about your workload than any index. Pick 20 to 50 representative tasks from your own logs, run each on the old and new model, and record the same metrics every time. Keep temperature and tool definitions identical across runs. | Metric | How to measure | Why it matters | |---|---|---| | Output tokens per task | Read usage metadata on each response | Tests Google’s 17% fewer output tokens claim on your data | | Tool calls per task | Count calls in the agent trace | Tests the fewer-tool-calls claim | | Reasoning steps per task | Count planning turns in the trace | Tests the fewer-reasoning-steps claim | | Task success rate | Run your existing pass/fail checks | Guards against savings that come from giving up early | | End-to-end latency | Time from request to final result | Captures real speed including tool round trips | | Cost per completed task | Multiply tokens by your current rates | Turns efficiency into a budget number | Success rate is the metric people skip, and it is the one that keeps savings honest. A model can cut tokens by stopping early, which looks great on a cost chart and terrible in production. Pair every efficiency number with a quality check. Our write-up on AI cutting coding time 52% while slop hits 25% of pull requests https://shattered.io/ai-cuts-coding-time-slop-25-percent-prs-2026/ is a reminder that speed gains and quality losses often travel together. Run the comparison more than once, because agent behavior varies between runs. Three passes per task give you a rough sense of variance. If the new model wins on all six metrics, switch. If it wins on tokens but loses on success, keep the old one for hard tasks and route easy ones to the new Flash. Five Predictions for Gemini 3.6 Flash These are analyst judgments, not reported facts. They follow from the confirmed details above and from how earlier Flash releases played out. 1. Independent testers will confirm a token saving but at a different size. The 17% figure comes from one index, and real agent traces usually produce a different ratio, up or down. 2. Google will formalize the release with a launch post and pricing. A listing in Antigravity and an API value suggest the paperwork follows, though we cannot say when. 3. Developers will default to Flash for sub-agents. If fewer tool calls hold up, Flash becomes the obvious worker model, with a larger model reserved for planning. 4. Competitors will answer with efficiency claims of their own. Token economy is now a marketing axis, and rivals will publish comparable figures. 5. Model churn will keep accelerating. Expect another Flash-tier name before the end of the year, and plan config-driven model selection accordingly. What This Means for Teams Building on Gemini The practical advice is short. Pin your model value. Log token counts. Build a regression set that mirrors your real tasks. Read the changelog weekly, because routing and deprecation behavior can change your traffic without a code change on your side. Treat the “just appeared” and “testing for days” stories as an alert to check your dashboards, not as proof of anything. Google’s confirmed facts give you plenty to work with: the model is listed in Antigravity, it has an API identifier, it supports the Antigravity agent in Gemini Managed Agents, and Google claims it uses 17% fewer output tokens and fewer tool calls than its predecessor. Everything else waits for Google’s own announcement or for independent measurement. For more AI model coverage, browse our AI and machine learning section https://shattered.io/ai-machine-learning/ . FAQ: Gemini 3.6 Flash in Antigravity Is Gemini 3.6 Flash officially available in Google Antigravity? Yes. Google’s official pages list Gemini 3.6 Flash as a model in Google Antigravity and state that it is available there. Did the model really appear suddenly and test for days? Press coverage says so, but Google has not confirmed it. No official material gives a testing start date or duration. What is the API model name? The Gemini API documentation identifies the model value as gemini-3.6-flash . How much more efficient is it than Gemini 3.5 Flash? Google states it used 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. Google also says it uses fewer reasoning steps and tool calls in multi-step workflows. Your own workload may differ. What does it cost, and which plans include it? No confirmed price or plan availability appears in the official material behind this story. Check Google’s pricing pages before budgeting. Does it work with Gemini Managed Agents? Yes. The Gemini API documentation says Gemini 3.6 Flash supports the Antigravity agent in Gemini Managed Agents. Should I switch production traffic now? Test first. Run your own tasks on both models, compare tokens, tool calls, success rate and latency, then decide. Pin the model value so a routing change cannot surprise you. Where can I read the official model details? Google publishes a model card on the DeepMind site and a what’s-new page in the Gemini API docs. Both are linked in the article above.