Google is shutting down gemini-robotics-er-1.6-preview on 31 August 2026

If your code calls gemini-robotics-er-1.6-preview, you have a firm deadline: Google’s official deprecation table lists 31 August 2026 as the shutdown date for that model ID. After that date, the endpoint goes dark. No grace period, no fallback. The named replacement is gemini-robotics-er-2-preview, which launched on 30 July 2026 alongside a separate streaming-optimised sibling endpoint.

What’s being retired and why

gemini-robotics-er-1.6-preview was itself only released on 14 April 2026, so its active life spans roughly four and a half months. That short runway is typical of preview models in the Gemini Robotics ER family: Google ships them to let developers build and provide feedback, then moves the entire ecosystem forward to the next generation rather than maintaining multiple parallel versions.

The Gemini Robotics ER series are vision-language models built for physical AI. They interpret visual data from robot cameras, reason about spatial relationships, plan multi-step tasks, and orchestrate hardware and software tools. The “ER” stands for Embodied Reasoning. These are not general-purpose chat models repurposed for robotics; they are trained specifically on embodied reasoning tasks using Google’s TPU infrastructure, built on top of Gemini 3.5 Flash.

What gemini-robotics-er-2-preview actually adds

The upgrade from 1.6 to 2 is substantial. Here is what the new model brings:

Multi-robot coordination. ER 2 allows different robots to share a semantic understanding of a task and delegate subtasks based on their physical capabilities. In Google’s demonstrations, a wheeled indoor robot and an Apptronik Apollo 2 humanoid worked together, with the humanoid handling terrain the wheeled robot could not navigate. The robots communicate task state through a shared session rather than through separate, siloed model calls.

Agentic code execution and tool orchestration. Developers can declare low-level control interfaces, including Vision-Language-Action (VLA) models and navigation APIs, as function-calling tools. ER 2 can then plan across those tools across multiple steps, executing code and sequencing actions without requiring the developer to manually manage each transition.

Continuous video reasoning. Previous ER models reasoned over static images or short clips. ER 2 reasons over continuous video streams and tracks task state across minutes. It can find specific moments in video with 91.3% accuracy and a mean absolute distance of 0.96 seconds, at four times the execution speed of the previous generation.

Safety improvements. ER 2 can detect nearby people and halt a humanoid robot when someone enters its working area, resuming only once the area is clear. It outperformed ER 1.6 and other evaluated frontier models on safety instruction-following and human-proximity benchmarks. That said, Google’s model card is explicit: do not use Gemini Robotics models for safety-critical applications in healthcare, transportation, or any setting where a malfunction could foreseeably cause death, injury, or property damage.

The streaming endpoint is separate

One detail worth getting right before you migrate: Live API streaming support is not part of gemini-robotics-er-2-preview. It belongs to a distinct model ID, gemini-robotics-er-2-streaming-preview, which launched on the same date.

The streaming endpoint is designed for low-latency, bidirectional sessions where a robot agent needs continuous multimodal input, such as a live video feed, and must react in near real-time. It supports text, image, video, and audio input and uses blocking function calls to serialise the model’s tool-selection loop. Note that blocking function calls serialise the model’s decisions; they do not make the underlying hardware safe.

If your ER 1.6 integration relied on request-response calls for spatial reasoning and task planning, gemini-robotics-er-2-preview is your replacement. If you were building towards a streaming or live-monitoring use case, you want gemini-robotics-er-2-streaming-preview.

What you need to do before 31 August 2026

The migration steps are straightforward in concept, though testing will take the most time:

1. Update your model ID. Replace gemini-robotics-er-1.6-preview with gemini-robotics-er-2-preview in your API calls. If you are building a streaming integration, use gemini-robotics-er-2-streaming-preview instead.

2. Review the current documentation. The Gemini Robotics overview and streaming guide reflect the ER 2 architecture. Both endpoints accept text, image, video, and audio inputs and support function calling, but the agentic tool orchestration capabilities in ER 2 may require changes to how you declare and handle tools.

3. Regression-test before you go live. Spatial reasoning outputs, structured responses, video moment-finding behaviour, and safety control responses may all differ from ER 1.6. Google notes this should be treated as a migration trigger, not an assumption of drop-in compatibility. Test against your real workloads.

4. Check your deployment target. Both ER 2 endpoints are available via the Gemini API and Google AI Studio. Access via the Gemini Enterprise Agent Platform is in private preview.

The deprecations page lists 31 August 2026 as the earliest possible shutdown date, which is Google’s standard language. In practice, treat it as a hard deadline. Once a model is shut down, the endpoint stops responding entirely.

The bigger picture

Gemini Robotics ER 2 represents a meaningful generational step: from a model that reasons about snapshots of the physical world to one that reasons continuously over it, coordinates across multiple robots, and executes multi-step agentic plans. For developers building in this space, the deprecation of ER 1.6 is the practical trigger to engage with those capabilities properly, rather than treating the new model as a drop-in swap and moving on.

If you have not started testing against ER 2 yet, now is the time to do it with enough runway to catch surprises before August.