Gemini's June 8 Hard Cutoff: Everything That Breaks and How to Fix It

Eight days.

That is how long Google is giving builders to migrate off the legacy Gemini Interactions API schema before it stops working entirely. On June 8, 2026, Google removes the outputs array from Gemini API responses — permanently. No opt-out. No pinned version escape hatch. If your code touches interaction.outputs, it breaks.

This is not a soft deprecation with a long tail. The timeline moved fast: the new schema became the default on May 26 (already live), and the legacy format disappears June 8. That is a 13-day window from default flip to hard removal. Most teams have not acted yet.

Here is what breaks, why, and exactly how to fix it.

The Timeline

If you have not upgraded your SDK or migrated your response-parsing code, the May 26 switch may already be causing silent failures in staging. The June 8 date is when production breaks.

The Core Change: outputs → steps

The fundamental restructuring is this: Gemini API responses used to return a flat outputs array containing only the model’s generated content. Now they return a steps array — a complete execution timeline of everything that happened during the request.

Legacy response (breaks June 8):

{
  "outputs": [
    { "type": "text", "text": "Here is the answer..." }
  ]
}

New response (required after June 8):

{
  "steps": [
    {
      "type": "user_input",
      "content": [{ "type": "text", "text": "Tell me a joke." }]
    },
    {
      "type": "model_output",
      "content": [{ "type": "text", "text": "Here is the answer..." }]
    }
  ]
}

The steps array can also include thought steps, function_call steps, function_result steps, google_search_call steps, and google_search_result steps — making the full execution chain visible and inspectable.

Why Google did this: The old outputs array only gave you the end result. The new steps timeline gives you the full trace: what the model thought, what tools it called, what results came back, and what it generated. For agentic workloads, this is table stakes. For simple text generation, it’s a structural change you have to handle.

SDK Version Requirement

This is the fastest fix for most builders:

Upgrade command:

pip install --upgrade google-genai  # Python
npm install @google/genai@latest    # Node.js

After upgrading, the SDK exposes convenience properties that handle the steps traversal for you:

# After SDK 2.0.0 upgrade — works immediately
interaction = client.interactions.create(
    model="gemini-3.5-flash",
    input="Summarize this document."
)
print(interaction.output_text)   # convenience property — no steps traversal needed
print(interaction.output_image)  # if image was generated
print(interaction.output_audio)  # if audio was generated

The convenience properties are the fast path. If you only need the final output, upgrading the SDK and switching to interaction.output_text instead of interaction.outputs[-1].text is the entire migration.

If you do need to inspect the steps themselves (for tracing, debugging, or agentic workflows), read on.

Five Areas That Break

1. Basic Text Generation

Before (breaks June 8):

interaction = client.interactions.create(
    model="gemini-3.5-flash",
    input="Tell me about agentic AI."
)
text = interaction.outputs[-1].text

After:

interaction = client.interactions.create(
    model="gemini-3.5-flash",
    input="Tell me about agentic AI."
)
text = interaction.output_text  # SDK convenience property

If you need to walk the steps manually:

for step in interaction.steps:
    if step.type == "model_output":
        for content_block in step.content:
            if content_block.type == "text":
                print(content_block.text)

2. Streaming

Streaming has the most invasive change. The SSE event structure is completely different.

Before (breaks June 8):

with client.interactions.stream(...) as stream:
    for event in stream:
        if event.type == "content.delta":
            print(event.delta, end="")

After:

with client.interactions.stream(...) as stream:
    for event in stream:
        if event.type == "step.delta":
            if hasattr(event, "text_delta"):
                print(event.text_delta, end="")

New streaming events to handle:

Builders with custom streaming parsers need to rewrite their event handlers. The content.delta event type does not exist in the new schema.

3. Function Calling

Function calling has changed from embedded outputs to discrete step types.

Before (breaks June 8):

response = client.interactions.create(
    model="gemini-3.5-flash",
    input="What's the weather in Tokyo?",
    tools=[weather_tool]
)
# Tool calls were embedded in outputs
tool_calls = [o for o in response.outputs if o.type == "tool_call"]

After:

response = client.interactions.create(
    model="gemini-3.5-flash",
    input="What's the weather in Tokyo?",
    tools=[weather_tool]
)
# Check if action is required
if response.status == "requires_action":
    for step in response.steps:
        if step.type == "function_call":
            # Execute function_call.name with function_call.arguments
            result = execute_function(step.function_call)
            # Submit result back
            response = client.interactions.submit_tool_outputs(
                interaction_id=response.id,
                tool_outputs=[{"step_id": step.id, "output": result}]
            )

The key difference: the interaction status is now "requires_action" when tool execution is needed, matching the OpenAI Assistants API pattern. Tool calls and results appear as separate, typed steps in the timeline.

4. Stateful Conversations (Multi-turn)

Multi-turn conversation handling changes significantly — actually for the better.

Before (breaks June 8):

# Manually maintain history array
history = []
for user_message in conversation:
    history.append({"role": "user", "parts": [user_message]})
    response = client.generate_content(
        model="gemini-3.5-flash",
        contents=history
    )
    history.append({"role": "model", "parts": [response.outputs[-1].text]})

After:

# Server-side state management via previous_interaction_id
first_response = client.interactions.create(
    model="gemini-3.5-flash",
    input="My name is Alex. Remember that."
)

second_response = client.interactions.create(
    model="gemini-3.5-flash",
    input="What's my name?",
    previous_interaction_id=first_response.id  # server manages history
)
print(second_response.output_text)  # "Your name is Alex."

Google manages conversation state server-side. Interactions are retained for 55 days on paid tier, 1 day on free tier. This eliminates the manual history-array pattern entirely — a real simplification if you adopt it.

Note: If you were passing the steps array manually in the input field of subsequent requests (as described in early migration docs), the cleaner path is previous_interaction_id.

5. Multimodal and Structured Output

Two additional schema changes affect multimodal and structured output builders:

response_format is now an array (multimodal):

# Before
response = client.interactions.create(
    model="gemini-3.5-flash",
    input="Describe this image and generate a caption image.",
    response_format={"type": "text"}  # single object
)

# After
response = client.interactions.create(
    model="gemini-3.5-flash",
    input="Describe this image and generate a caption image.",
    response_format=[{"type": "text"}, {"type": "image"}]  # array
)

response_mime_type removed; response_format used for JSON:

# Before
response = client.interactions.create(
    model="gemini-3.5-flash",
    input="Extract the names from this text.",
    generation_config={"response_mime_type": "application/json"}
)

# After
response = client.interactions.create(
    model="gemini-3.5-flash",
    input="Extract the names from this text.",
    response_format={"type": "json_object"}  # or json_schema with schema
)

image_config moves from generation_config to response_format:

# Before
response = client.interactions.create(
    model="gemini-3.5-flash",
    input="Generate an image of a forest.",
    generation_config={"image_config": {"aspect_ratio": "16:9"}}
)

# After
response = client.interactions.create(
    model="gemini-3.5-flash",
    input="Generate an image of a forest.",
    response_format={"type": "image", "image_config": {"aspect_ratio": "16:9"}}
)

Server-Side Tools: Search and Code Execution

If you use Google Search or Code Execution as built-in tools, their results now appear as typed steps:

response = client.interactions.create(
    model="gemini-3.5-flash",
    input="What are today's top AI headlines?",
    tools=[{"google_search": {}}]
)

for step in response.steps:
    if step.type == "google_search_call":
        print(f"Searched: {step.query}")
    elif step.type == "google_search_result":
        print(f"Found: {step.results}")
    elif step.type == "model_output":
        print(f"Answer: {step.content[0].text}")

Citations now appear inline as content annotations rather than as a separate citations field. Builders who parse response.citations will need to update to step.content[n].annotations for the model output step.

How to Know If You’re Affected

Run this grep against your codebase:

# Python
grep -rn "\.outputs\[" . --include="*.py"
grep -rn "\.outputs\[-1\]" . --include="*.py"
grep -rn "response_mime_type" . --include="*.py"
grep -rn "content\.delta" . --include="*.py"
grep -rn "generation_config.*image_config" . --include="*.py"

# JavaScript
grep -rn "\.outputs\[" . --include="*.ts" --include="*.js"
grep -rn "responseMimeType" . --include="*.ts" --include="*.js"
grep -rn "content\.delta" . --include="*.ts" --include="*.js"

Each match is a migration point. If you find zero matches and you’re on SDK 2.0.0+, you are already compliant. If you are on SDK 1.x and have zero matches, you may still have silent failures after May 26 that become hard failures on June 8.

The One-Day Migration Path

For most builders, this is the sequence:

1. Upgrade SDK — pip install --upgrade google-genai / npm install @google/genai@latest
2. Replace outputs[-1].text with output_text across your codebase
3. Rewrite streaming handlers — swap content.delta for step.delta
4. Update function calling — check for requires_action status, handle function_call steps
5. Fix response_mime_type — replace with response_format object
6. Fix image_config — move from generation_config to response_format
7. Test multi-turn flows — consider switching to previous_interaction_id server-side state

Steps 1 and 2 cover 80% of affected codebases. Steps 3–7 apply to builders using advanced features.

The new schema is genuinely better — the execution timeline is more useful than a flat outputs array, and server-side conversation state is simpler than manual history management. The migration is worth doing even if the deadline were further out.

It is not. Eight days.

Google’s official migration guide is at ai.google.dev/gemini-api/docs/interactions-breaking-changes-may-2026. SDK upgrade docs at ai.google.dev/gemini-api/docs/migrate-to-interactions.