Google leads Gemini 3.5 with Flash, not Pro, and ties it to the Antigravity agent harness

Flash before Pro is a deliberate ordering

The convention for frontier launches has been to lead with the flagship: the biggest, most capable, most expensive model, followed weeks later by the cheaper variant. Google inverted that here. Gemini 3.5 opens with 3.5 Flash generally available today, while 3.5 Pro is described as "already being used internally" with a rollout "next month."

That choice tells you what Google thinks the constraint is. For agentic workloads that fan out into many subagent calls and run over long horizons, per-token cost and latency dominate the economics far more than the marginal quality of a single response. By making the fast, cheap model the headline, Google is signaling that the unit of value is the completed multi-step task, not the eloquence of one answer.

It also lets Google make a specific comparative claim: 3.5 Flash outperforms the previous-generation 3.1 Pro on Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), and MCP Atlas (83.6%). Beating your own last-gen flagship with your new mid-tier model is a cleaner story than beating a competitor, and it justifies leading with Flash.

The benchmarks chosen are agent benchmarks

Look at what Google reported: Terminal-Bench, MCP Atlas, GDPval-AA, and CharXiv Reasoning (84.2%). These are not general knowledge or chat-quality scores. Terminal-Bench measures whether a model can operate in a shell across multiple steps; MCP Atlas concerns tool-calling via the Model Context Protocol; GDPval-AA is framed around economically valuable work. The benchmark selection is itself a claim about what the model is for.

The speed figure — "4 times faster than other frontier models" on output tokens per second — is presented as the enabler rather than a separate feature. Google's framing is that when a workflow requires an agent to plan, execute a tool call, read the result, and iterate dozens of times, throughput compounds. A 4x speed advantage on a single call becomes a much larger wall-clock advantage across a long-horizon task.

The MCP Atlas score is worth flagging specifically for teams building on top of this. Reliable tool-calling is the load-bearing capability for any agent that touches real systems, and it degrades ugly when it fails. An 83.6% is a number to hold Google to in practice, not just accept in a launch post.

The harness is doing as much work as the model

Almost every capability demonstration in the announcement is qualified by Antigravity, Google's "agent-first development platform" and its associated harness. The examples — two agents synthesizing the AlphaZero paper and coding a playable game in six hours, a builder-and-player self-improvement loop, migrating a legacy codebase to Next.js, deploying subagents to categorize unstructured assets — are all described as things 3.5 Flash does "powered by" or "leveraging" Antigravity.

When coupled with the updated Antigravity harness, 3.5 Flash becomes a powerful engine for deploying collaborative subagents to tackle problems at scale for the most demanding use cases.Montana Labs

This matters because it blurs the line between what the model can do and what the surrounding orchestration provides. The subagent coordination, context retention, and multi-step supervision are properties of the harness as much as the weights. Google also notes the agent runs "under supervision" and "under your direction" — a quiet acknowledgment that unsupervised long-horizon autonomy is not yet the claim being made.

Named enterprise pilots stake out the target workloads

The partner list is unusually concrete about the kind of work Google is chasing: Shopify running parallel subagents for merchant growth forecasts, Macquarie Bank reasoning over 100-plus-page onboarding documents, Salesforce embedding Flash into Agentforce for multi-turn tool calling, Ramp doing invoice OCR combined with historical reasoning, Xero autonomously gathering 1099 tax-form data over multi-week workflows, and Databricks diagnosing data issues across large datasets.

The common thread is document-heavy, multi-step back-office automation in finance and operations — the "toil and complexity" Google says it profiled with partners while building the series. These are workloads defined by long horizons and cost sensitivity, which is precisely why leading with a fast, cheap model makes strategic sense. The consumer side — Gemini Spark as a 24/7 personal agent, defaulting Flash across the Gemini app and Search AI Mode — spreads the same model to billions of users at once.

What Google is actually selling here

The specific implication of this launch is that Google is repositioning its cheapest tier as the default substrate for agentic work, and betting that most economic value lives in high-volume, multi-step tasks where throughput and price beat marginal answer quality. The Pro model still exists and arrives next month, but the release is structured so that the fast model is the one enterprises and consumers meet first.

For teams evaluating this, the practical questions are: how much of the demonstrated capability survives outside the Antigravity harness and through the raw Gemini API, and whether the tool-calling and long-horizon reliability hold on your own workflows rather than curated demos. The benchmark numbers and named pilots are a starting point for that testing, not a substitute for it — especially for the finance and document-processing use cases Google is explicitly courting, where an incorrect autonomous action is expensive to unwind.