Google Colab Free Tier Limits (2026): GPU, Runtime & Pricing

Google Colab is free to use, but Google does not publish a fixed number of free GPU hours. Free-tier GPU access, idle timeouts, maximum VM lifetime and available accelerator types change with demand and usage history. A free notebook can run for at most 12 hours and may end sooner; neither a GPU nor a particular GPU model is guaranteed.

Pricing check: The UK prices and plan allowances below were checked on 6 August 2026. Colab pricing, available hardware and usage policies can change, so confirm the current offer on the official Google Colab plans page before paying.

Google Colab free tier limits at a glance

What Google officially guarantees, and what it does not

Colab is a hosted Jupyter Notebook service that requires no local setup. It is designed primarily for interactive programming, machine learning, data science and education. A notebook runs inside a virtual machine assigned to your account, while the notebook itself can be stored in Google Drive.

The important distinction is between access and entitlement. Free Colab may give you a CPU, GPU or TPU runtime, but it does not promise that a GPU will be available when you need one. It does not promise a particular model, a fixed idle timeout or a set number of sessions each week. Google says overall limits, maximum VM lifetime, idle periods and available hardware can fluctuate.

The official Google Colab FAQ also says free notebooks can run for at most 12 hours. “At most” matters: it is a ceiling rather than a minimum service level. A session can end earlier because it has been idle, demand is high, the workload conflicts with Colab’s priorities or the account has recently consumed substantial resources.

Paid plans improve access rather than creating unlimited compute. If a paid account exhausts its compute units, it returns to free-tier policies and restrictions until the balance is replenished. Premium accelerators also remain subject to availability.

Google Colab UK pricing in 2026

A compute unit is not the same as one GPU hour. The rate at which units are consumed depends on the virtual machine, accelerator and other resources selected. A faster GPU can use the balance more quickly than a standard CPU runtime. It is therefore misleading to turn 100 units into a universal number of hours without knowing the runtime configuration.

This is also why the cheapest plan is not automatically the lowest-cost option. A notebook that repeatedly disconnects, reloads a large model and reprocesses the same files can waste both time and compute. For a broader view of that trade-off, see my analysis of AI compute costs for UK businesses.

Which Colab plan should you choose?

Start with the free tier when the work is interruptible and you are still proving that the notebook works. CPU-based data cleaning, spreadsheet transformation, small model tests and teaching exercises often do not justify a subscription.

Pay As You Go makes sense when you need a temporary increase but cannot predict regular monthly use. Pro is the more natural choice for someone opening Colab throughout the month and wanting a recurring allocation. Pro+ is aimed at heavier users who can benefit from a larger balance and longer background execution.

None of these plans should be treated as a guaranteed production server. If a customer-facing process must start at a precise time, run continuously or meet a service-level agreement, use infrastructure with explicitly reserved capacity. Colab is strongest as an interactive development and batch-processing environment. My guide to UK small-business automation explains how to separate useful prototypes from dependable operational systems.

Case study: the 3,500-image Reddit anecdote

One Reddit user reported using a Python script in free Google Colab to remove the backgrounds from approximately 3,500 images. They said the job took around three hours, was delivered within 24 hours and earned $200.

It is a useful example of Colab’s appeal, but it is an attributed anecdote rather than an official benchmark. The report does not establish which accelerator was assigned, whether every image required the same processing, or whether another account would receive the same runtime. It certainly does not prove that free users receive three guaranteed GPU hours.

The transferable lesson is the workflow: a browser notebook can make a one-off batch task economical because there is little environment setup. The quota conclusion is not transferable. Anyone repeating the job should design for an interrupted session and test a representative sample before promising a delivery time.

How to make Colab batch jobs survive interruptions

1. Use small batches. Process a defined group of files, save the results and then start the next group. Do not leave all progress in runtime memory.
2. Make the script resumable. Keep a manifest of completed file names or record status in a CSV so a restarted notebook skips finished work.
3. Copy working data to the runtime. Repeatedly reading thousands of small files from a mounted Drive folder can be slow and may hit Drive limits. Transfer an archive, unpack it locally and upload completed batches.
4. Choose the cheapest adequate runtime. A GPU runtime consumes scarce resources even when the code is not using the GPU. Use a standard runtime for ordinary Python, file handling or spreadsheet work.
5. Validate output continuously. Check dimensions, file counts, empty outputs and a visual sample after each batch. AI-generated code still needs review and testing.

Colab also works well for temporary spreadsheet analysis and data preparation. If the final process needs to exchange structured data with an AI assistant, the pattern in my ChatGPT and Google Sheets guide provides a more controlled route than manually moving results between tools.

Colab and the Model Context Protocol

The Model Context Protocol is an open standard for connecting AI applications to tools, data sources and workflows. The official MCP introduction describes it as a standard interface, rather than a Colab-specific feature.

A developer can build tooling that lets an MCP-capable assistant interact with a notebook workflow, but that does not mean Google guarantees remote control of free Colab sessions. Free managed runtimes restrict some remote-control and distributed-worker behaviour, and sessions remain temporary. Treat any Colab-to-MCP bridge as a custom integration that needs authentication, restricted permissions and confirmation before writing or deleting data.

Data, secrets and client work

Do not paste API keys into notebook cells or commit them with a shared notebook. Use Colab’s secrets features or inject credentials at runtime, grant only the permissions required and revoke temporary credentials after the job.

Be particularly cautious with confidential or personal data. Google states that its generative AI features may collect prompts, related code, generated output, usage information and feedback, and advises users not to submit sensitive or personally identifying information. For client work, confirm that uploading the data is permitted, minimise what is transferred and use a controlled alternative when confidentiality requirements are incompatible with the service.

Google Colab limits: common questions

How many free GPU hours does Google Colab provide?

There is no published fixed allowance. Free GPU availability and usage limits vary with demand, account usage and available hardware.

Does Colab Free guarantee an Nvidia T4 or another GPU?

No. A GPU is not guaranteed, and Google says available GPU and TPU types change over time.

Is Google Colab Pro unlimited?

No. Pro provides 100 compute units at the UK price checked above. Resource availability remains dynamic, and exhausted balances revert to free-tier policies.

How long can a Colab notebook run?

Free notebooks can run for at most 12 hours and may stop sooner. In general, Colab sessions are temporary. Pro+ can support continuous execution for up to 24 hours when enough compute units remain.

Do unused Pay As You Go units expire?

Yes. The UK purchase options checked on 6 August 2026 stated that purchased compute units expire after 90 days.

Is Colab suitable for business automation?

It is useful for experiments, analysis and interruptible batch work. It is a poor foundation for a process that requires guaranteed hardware, permanent storage, continuous availability or a contractual service level.