Introduction

Installing TensorFlow into an Anaconda environment built on Python 3.6 is now a legacy setup rather than a modern default. That distinction matters, because current TensorFlow installation guidance targets newer Python versions and recommends pip inside an isolated environment, so if you must stay on Python 3.6 you need to choose an older TensorFlow release that still published compatible wheels.

Understand the Compatibility Constraint First

The most important detail is compatibility, not the exact command. Current TensorFlow releases no longer target Python 3.6, so a fresh pip install tensorflow in a Python 3.6 environment will usually fail or select nothing usable.

The safe approach is:

• Create a dedicated conda environment with Python 3.6
• Activate it
• Install a TensorFlow version that still supported Python 3.6
• Verify the import immediately

In practice, this means treating the environment as a legacy project environment rather than a long-term default for new work.

Create the Conda Environment

Start by creating and activating a separate environment.

conda create -n tf36 python=3.6
conda activate tf36

Keeping TensorFlow in its own environment avoids conflicts with newer projects and prevents accidental upgrades that break the setup.

Install TensorFlow with pip Inside the Conda Environment

Even when you use Anaconda for environment management, TensorFlow is commonly installed with pip inside that environment.

python -m pip install --upgrade pip
python -m pip install "tensorflow==2.6.*"

If your platform or project requires an earlier release, use the version your codebase expects instead. The key point is that the TensorFlow version must still have wheels built for Python 3.6.

After installation, verify it:

python -c "import tensorflow as tf; print(tf.__version__)"

If TensorFlow imports and prints a version number, the basic installation succeeded.

Why Not conda install tensorflow?

Many developers expect conda to provide the cleanest installation path, but TensorFlow’s own guidance has favored pip packages for official releases. In a conda-managed environment, that usually means:

• Use conda to manage Python and system isolation
• Use pip to install the TensorFlow wheel

That hybrid approach is normal and avoids confusion when conda channels lag behind the official package releases.

GPU Installs Need Extra Compatibility Checks

If you need GPU support in an old Python 3.6 environment, the setup becomes stricter. You must match:

• TensorFlow version
• CUDA version
• cuDNN version
• Operating system support
• Driver compatibility

For many legacy projects, CPU-only installation is simpler and more reproducible. If you do need GPU support, pin the exact versions used by the project rather than trying random combinations.

Verify the Installation with a Small Test

Do not stop after the package installs. Run a minimal TensorFlow operation to confirm the runtime is working.

1import tensorflow as tf
2
3x = tf.constant([[1.0, 2.0], [3.0, 4.0]])
4y = tf.reduce_sum(x)
5
6print(tf.__version__)
7print(y.numpy())

If this prints the TensorFlow version and a numeric result, the environment is at least functional for basic use.

When You Should Upgrade Instead

If you are starting a new project, Python 3.6 is the wrong place to begin. Use a modern Python version and the current TensorFlow install path instead. Staying on Python 3.6 only makes sense when:

• You maintain an older production model
• A library dependency hard-requires the old stack
• You need reproducibility for historical experiments

In those cases, write the environment details down in a reproducible form such as environment.yml or requirements.txt.

Common Pitfalls

The most common mistake is trying to install the newest TensorFlow into Python 3.6. Current packages are built for newer Python versions, so this usually fails before you even get to runtime.

Another mistake is mixing too many package managers and channels at once. Create the conda environment, then install the TensorFlow wheel cleanly instead of layering multiple conflicting installations.

A third mistake is ignoring GPU compatibility tables. A package version may install successfully but still fail at runtime if its CUDA and cuDNN expectations are wrong.

Summary

• Python 3.6 plus TensorFlow is a legacy environment and must use an older compatible TensorFlow release.
• Use conda to create the environment, then install TensorFlow with pip inside it.
• Verify the installation with both import tensorflow and a small tensor operation.
• Prefer CPU-only installs unless you know the exact GPU dependency matrix.
• For new work, upgrade Python and use the current TensorFlow installation guidance instead.