Picking a model inside Google Antigravity is not the same act as picking one in a chat window. A chat window shows you every difference immediately, because you read the answer. An agent harness hides most of them behind twenty minutes of autonomous tool calls, then hands you a diff. Some differences get amplified in that gap: tool-call discipline, coherence over a long task, and the cost of finishing a job. Others stop mattering. This is a guide to the ones that survive, built from Google’s documentation as of August 2026.
What Google Antigravity Lets You Select Today
The current model list is shorter than the launch coverage suggests, and the names have moved. Google’s models documentation lists eleven dropdown entries across six reasoning models: Gemini 3.6 Flash (Low, Medium, High), Gemini 3.5 Flash (Low, Medium, High), Gemini 3.1 Pro (Low, High), Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking), and GPT-OSS-120b (Medium).
Two things stand out. First, there is no Gemini 3 Pro option any more. At launch on November 18, 2025, Google’s announcement post offered "Access to Google’s Gemini 3, Anthropic’s Claude Sonnet 4.5 models, and OpenAI’s GPT-OSS." The Pro slot is now filled by Gemini 3.1 Pro, so a tutorial telling you to select Gemini 3 Pro describes a product that no longer exists.
Second, reasoning effort is part of the selection, not a separate setting. Gemini 3.1 Pro exposes Low and High only. GPT-OSS-120b is offered at Medium and nothing else.
The selector sits beneath the conversation prompt box, and the documentation notes that the choice "is sticky between user messages within a conversation," persisting until the agent has completed its steps for that user turn or you cancel the current execution. One model is not selectable: Nano Banana 2, which the agent reaches for on its own to generate UI mockups, diagrams or page images.
Why the Harness Amplifies Some Differences and Erases Others
In a chat window, prose quality is most of the product. In an agent harness it is nearly irrelevant. What replaces it is tool-call reliability. A model that fumbles one file edit in fifty is fine for chat and miserable for a forty-step task, because the failure compounds: the agent reads back its own broken state and reasons from it.
Long-horizon coherence is the second amplified trait. Google Antigravity produces what its IDE overview calls artifacts, defined as "anything that the agent creates to allow it to get its work done or communicate its accomplishments to the human user." A model that drifts from its own plan by step thirty produces artifacts that look complete and are not, which is worse than an obvious failure.
The third amplified trait is cost, measured per task. The plans documentation states that "The rate limits are correlated with the amount of work done by the agent, which can differ from prompt to prompt." A cheaper model that needs three attempts is the expensive choice.
What the harness erases: tone, formatting preferences, and most of what people argue about on model release day.
Rule One: Gemini 3.1 Pro When the Task Crosses Files
Use the Pro tier when the unit of work is a change rather than an edit: multi-file refactors, migrations touching configuration and code and tests together, any task where the agent holds a plan in mind while the codebase shifts underneath it.
Published token pricing is a useful proxy for where Google positions this model, even though Google Antigravity subscribers are not billed per token. The Gemini API pricing page lists Gemini 3.1 Pro Preview at $2.00 per million input tokens and $12.00 per million output tokens for prompts up to 200k tokens, rising to $4.00 and $18.00 above that threshold. The 200k cliff is the signal worth reading: long-context work on the Pro tier gets materially more expensive at exactly the point where an agent starts dragging a large repository into context.
Practical consequence: run Gemini 3.1 Pro on High for planning and the hard middle of a task, then switch away for the cleanup. The stickiness of the selector makes that mistake easy.
Rule Two: Let the Flash Tiers Carry the Volume
Flash is the default for a reason. Google’s documentation home page describes Gemini 3.5 Flash as "Powering all local agents with SOTA speed, reasoning, and context window capacity." That is a vendor claim about a vendor model, self-reported until an independent evaluation says otherwise, but the positioning is clear.
Choose a Flash tier for the bulk of the work: test generation, dependency bumps, log triage, writing the fiftieth similar handler, browser verification passes. Those tasks are shallow and numerous, the profile where per-task cost dominates and a longer reasoning budget buys nothing.
A pricing curiosity is worth noticing. On the same pricing page, Gemini 3.6 Flash lists at $1.50 input and $7.50 output, while the older Gemini 3.5 Flash lists at $1.50 input and $9.00 output. The newer Flash is cheaper on output. If you pinned 3.5 Flash months ago out of caution, that pin now costs you.
Rule Three: When a Third-Party or Open Model Earns the Slot
Third-party models are still offered, which answers the question most people arrive with. Claude Sonnet 4.6 (thinking) and Claude Opus 4.6 (thinking) are both in the dropdown, as is OpenAI’s open-weight GPT-OSS-120b. Note what GPT-OSS is here: a model served inside Google’s product, not one you are hosting.
Reach for a Claude tier when the task is dense editing inside code you already have, where the agent needs to be conservative and surgical. That preference travels from other harnesses, so our comparison of Claude Code and OpenAI Codex reads as useful background here, as does the primer on what Claude Code is.
Reach for GPT-OSS-120b when you want a second opinion from a different model family on a bounded question, or when testing whether a workflow survives a model swap. It is offered at Medium reasoning only, so it is not a candidate for the hardest tasks.
Where Model Choice Stops: Subagents and Antigravity 2.0
This is the part most coverage misses. Your dropdown selection governs the main agent. It does not fully govern what runs beneath it.
The subagents documentation defines a model frontmatter property with a default of inherit and allowed values of inherit, flash, or pro. There is no value for Claude and no value for GPT-OSS. If you select a third-party model for the main agent and it delegates, the documented subagent vocabulary does not include that model. Anyone building a workflow around "I run Opus for everything" should test what their subagents actually run.
Two further constraints matter. A subagent "runs using the specified model tier but does not inherit the parent’s existing conversation history (context window), starting with a clean slate," so delegation is a context reset, not a context extension. Nesting is bounded at a maximum depth of 10 layers beneath the primary agent, and a parent "can invoke multiple subagents concurrently" with no published concurrency number.
The surface itself has moved. Google’s developer blog post of November 20, 2025 described the Manager as "A dedicated interface where you can spawn, orchestrate, and observe multiple agents working asynchronously across different workspaces." Today the documentation describes Antigravity 2.0 as a standalone desktop command center that, unlike the previous Agent Manager, does not require an IDE to run. When you orchestrate several agents at once, isolation matters more than model: Projects offer Local Mode, where agents share your active folders, and New Worktree Mode, which provisions a separate Git worktree per conversation. Teams thinking about agent permissions in a CMS context will recognize the containment problem.
Quota, Credits, and the Cost That Is Actually Metered
On a subscription, Google Antigravity meters agent work rather than tokens, on a refresh cadence that varies by plan. The plans page states that Google AI Ultra users receive "The highest, most generous quota, refreshed every five hours," Google AI Pro users receive "High, generous quota, refreshed every five hours until weekly limit reached," and users on neither plan receive "Meaningful quota, refreshed weekly."
Google is candid about why those numbers are not published: "The baseline rate limits are primarily determined to the degree we have capacity, and exist to prevent abuse." Do not build a team process that assumes a fixed daily budget of Pro-tier runs.
Overage has a price. Pro and Ultra users can spend purchased AI credits, and the docs say "AI credits are consumed at standard Gemini Enterprise Agent Platform consumption pricing." Once you exceed quota, the per-token differences between Pro and Flash return in full.
The rule that falls out is simple: optimize for the probability that a task finishes correctly on the first attempt, not for the sticker price of the tier. A Flash run that fails twice and then needs a Pro run has cost three units of agent work plus your attention. Teams running long asynchronous sessions, a pattern we examined in cloud agent sessions, tend to learn this the expensive way.
Frequently Asked Questions
Is Gemini 3 Pro still available in Antigravity?
No. The models documentation lists Gemini 3.1 Pro (Low, High) as the Pro-tier option. Gemini 3 was named in the November 18, 2025 launch post, but the selectable list has moved on. Guides that still say Gemini 3 Pro are out of date.
Are third-party models still offered?
Yes. The models page lists Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking) and GPT-OSS-120b (Medium). The two documentation pages disagree on eligibility: the plans page says third-party access belongs to Google AI Ultra users, while the models table marks them available on Free, Plus, Pro and Ultra.
Which models can subagents use?
The subagents documentation defines a model property with allowed values of inherit, flash, or pro, defaulting to inherit. Third-party models are not in that documented vocabulary, so a subagent cannot be pointed at Claude or GPT-OSS through that field.
Does my model choice persist between messages?
Yes. The documentation says the reasoning model choice “is sticky between user messages within a conversation” and holds until the agent has completed its steps for that user turn, or until you cancel the current execution. That makes it easy to leave an expensive tier selected for cheap follow-up work.
How is quota consumed?
Not per token, on a subscription. The plans page says rate limits “are correlated with the amount of work done by the agent, which can differ from prompt to prompt.” Google also states the baseline limits are set by available capacity and to prevent abuse, and does not publish fixed numbers.
What happens if I run out of quota?
Google AI Pro and Ultra users can use purchased AI credits for overage. The documentation says those credits “are consumed at standard Gemini Enterprise Agent Platform consumption pricing,” which brings per-token cost differences between model tiers back into play.
Is GPT-OSS-120b something I self-host?
Not inside this product. It is an open-weight OpenAI model served through Antigravity at Medium reasoning effort. The open weights mean the model is portable elsewhere, but the deployment you are using in Antigravity is Google’s.
Which model should I pick for a large refactor?
Start on Gemini 3.1 Pro at High for planning and the difficult middle of the change, then drop to a Flash tier for repetitive cleanup. Published API pricing shows Pro input rising from $2.00 to $4.00 per million tokens above a 200k-token prompt, so long-context Pro work is where costs climb fastest.
