Google I/O 2026: Gemini 3.5 Flash, Gemini Omni, and what teams should test

At Google I/O 2026, Google showed two different stories: Gemini 3.5 Flash for fast agentic and coding workflows, and Gemini Omni Flash for creating and editing video from different inputs.

For a software team, the useful question is simple: what is worth testing in real work now, and what should stay in demo territory for the moment?

Problem / Context

Google presents Gemini 3.5 Flash as the first release in the new Gemini 3.5 family. The emphasis is not only “a smarter model” but action: coding tasks, longer agentic work, subagents, Gemini API, Google AI Studio, Android Studio, Antigravity, and Gemini Enterprise.

In parallel, Gemini Omni Flash starts as a multimodal model for video. It can use text, images, video, or audio as references and generate or edit video through conversational instructions. Google also says Omni Flash is available in the Gemini app, Google Flow, and YouTube Shorts/Create, with developer and enterprise APIs coming later.

These are not one model for everything. Gemini 3.5 Flash should be evaluated as a working engine for code and tool-heavy tasks. Gemini Omni Flash should be evaluated as a media workflow where consistency, rights, watermarking, and review control matter.

Why it matters

For developers, the most interesting signal in Gemini 3.5 Flash is not the phrase “frontier intelligence”; it is the combination of speed, tool use, and agentic scenarios. Google says 3.5 Flash outperforms Gemini 3.1 Pro on several coding and agentic benchmarks, including Terminal-Bench 2.1, GDPval-AA, and MCP Atlas, while running faster than other frontier models by output tokens per second.

But a benchmark is not production proof. A team still needs to test its own tasks: its monorepo, tests, PR review rules, and data constraints.

Gemini Omni Flash matters from another angle. If a team creates learning material, product demos, explainers, or internal video, conversational video editing can save hours. The risks are different: rights to reference material, incorrect process visuals, synthetic media disclosure, and brand control.

What the official tests show

The most useful part of the official Gemini 3.5 Flash model card is not the broad claim that the model is better than the previous generation. It is the comparison by task type. It does not prove that a team should replace Codex or Claude tomorrow, but it does show where Google deserves a serious pilot.

Short version from Google’s May 2026 numbers:

The honest conclusion: Gemini 3.5 Flash does not look like a universal replacement for Codex/GPT-5.5 or Claude. It looks like a strong candidate for workflows where speed, tool use, MCP, multimodal input, and controlled agency matter. For very hard coding in a large repository, where mistakes are expensive, compare it against your current Codex or Claude setup on the same PR tasks instead of migrating because of one benchmark.

Gemini Omni Flash needs an even more careful reading. Google describes the capabilities, distribution channels, and SynthID watermarking, but the model card explicitly says evaluations for T2VA, I2VA, R2VA, video editing, and image generation will be shared later when the model rolls out to developers and enterprise customers via APIs. So the right mode for Omni today is creative/media pilot, not production replacement.

Should you move from Codex or Claude?

If a team already uses Codex/GPT-5.5 or Claude for code, do not make a big switch. Split the work instead.

Add Gemini 3.5 Flash to a pilot if:

• you have many tool-heavy tasks: MCP servers, browser/UI control, agents, internal tool integrations;
• response speed and cheaper iteration loops matter;
• tasks include charts, screenshots, video, audio, or mixed inputs;
• you need drafts, prototypes, first bugfix attempts, code explanation, or agent planning in a sandbox.

Keep Codex/GPT-5.5 as the primary option for:

• large refactors and tasks with very long context;
• critical PRs where reasoning quality and low random change rate matter most;
• difficult debugging where the model must hold many files, logs, and constraints at once.

Keep Claude Opus in the comparison for:

• hard codebase tasks, especially if your local evaluation resembles SWE-Bench-style repair;
• architecture review, trade-off explanation, and long-form reasoning tasks;
• final second opinion before important changes.

Gemini Omni Flash does not replace Codex or Claude. It is a separate track for video, explainers, storyboards, demo clips, and multimodal content workflows. It should not be part of the coding replacement discussion.

How to test Gemini 3.5 Flash

Start with a small evaluation set. Do not replace every AI tool in the team at once.

Choose 3-5 tasks:

• fix a small bug with a failing test;
• perform a behavior-preserving refactor;
• write an integration test for an existing endpoint;
• explain a legacy module and propose a split plan;
• build a small UI prototype from a description.

For each task, capture a baseline: how long a human or current model takes, what the diff looks like, how many review edits are needed, and whether tests pass.

Minimal matrix:

Scenario	What to measure	Good result	No-go
Bugfix	test pass, diff size	minimal diff, green test	changes unrelated files
Refactor	behavior, review effort	same tests, simpler code	breaks public API
Test generation	coverage, useful asserts	test catches a real case	test checks implementation detail
Agent task	steps, tool use, rollback	plans, executes, explains diff	performs irreversible actions
UI prototype	accessibility, responsive layout	usable preview	pretty but non-functional mockup

Run the first pass in a sandbox: test branch, test repo, or isolated workspace. Do not provide production secrets, customer data, or write access to critical systems.

How to test Gemini Omni Flash

Omni Flash should not be judged as a “model for code.” Test it where the output is media or an explanatory visual.

Start with three tasks:

1. A short product demo from an existing screen recording.
2. A learning explainer for a complex topic.
3. A video variation with a consistent character or style.

Measure more than the wow factor. Look at consistency across iterations, control over details, time to a usable result, number of manual edits, and whether viewers can understand that the content is AI-generated. Google says Omni videos include SynthID watermarking and can be verified through the Gemini app, Chrome, and Search, but that does not replace your own publishing policy.

A sane team rule: pilot Omni for internal demos, drafts, storyboards, and learning illustrations. For public branding, ads, or material involving people, require separate review.

What NOT to do

Do not compare models with one beautiful prompt. A single demo prompt says little about reliability in real work.

Do not mix coding and media evaluation. Gemini 3.5 Flash and Omni Flash have different strengths, risks, and metrics.

Do not give an agent production access without rollback. Even if the model plans well, the first pilot belongs in an isolated environment.

Do not trust benchmarks without local validation. Terminal-Bench or MCP Atlas are useful signals, but the decision must be made on your tasks.

Do not publish AI video without a policy. You need rules for reference material, consent, synthetic media disclosure, and fact-checking.

A practical 60-minute plan

1. Choose 3 coding tasks and 2 media tasks.
2. Create a separate test branch or sandbox.
3. Run Gemini 3.5 Flash on the coding tasks.
4. Run Omni Flash on one explainer or demo draft.
5. Record latency, quality, manual edits, and no-go signals.
6. Decide adopt, pilot, wait, or reject for each scenario.

If the result is good, do not “migrate to Google” in one move. Add one specific workflow: for example, Gemini 3.5 Flash for bugfix drafts in a sandbox, or Omni Flash for internal training videos.

Conclusion

Google’s new models after I/O 2026 look strong, but they should not be evaluated through one general impression.

Gemini 3.5 Flash is a candidate for agentic coding, longer tasks, and tool-heavy workflows. Gemini Omni Flash is a candidate for multimodal video creation and editing. Both can be useful if you test them separately, on real tasks, with clear no-go rules.

The best next step is a small one-day evaluation set. Not hype, not faith in benchmarks, but a controlled pilot with metrics.

Sources

• Google Blog: Gemini 3.5: frontier intelligence with action — https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/
• Google Blog: Introducing Gemini Omni — https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/
• Google DeepMind Model Card: Gemini 3.5 Flash — https://deepmind.google/models/model-cards/gemini-3-5-flash/
• Google DeepMind Model Card: Gemini Omni Flash — https://deepmind.google/models/model-cards/gemini-omni-flash/
• Google Blog: Building the agentic future: Developer highlights from I/O 2026 — https://blog.google/innovation-and-ai/technology/developers-tools/google-io-2026-developer-highlights/
• Google Blog: 100 things we announced at I/O 2026 — https://blog.google/innovation-and-ai/technology/ai/google-io-2026-all-our-announcements/