Free AI coding assistants in 2026: what each free tier really caps

Every one of the free AI coding assistants we checked on October 23, 2026 still costs $0, but only one kind of them tells you where the ceiling is. GitHub publishes a hard number for completions. Cursor, Devin and Google Antigravity publish no numbers at all for their free quotas. Open-source clients publish nothing because the limit is your own model bill.

In our view that difference is the decision. A free tier with a printed cap tells you when to budget for a paid plan; a free tier described as "limited" tells you only that it will stop, at a moment you cannot plan around.

What free ai coding assistants cap in October 2026

The caps have moved units. Across the pricing pages we read on October 23, 2026, the free tiers that state a limit at all state it in something the vendor meters rather than something you can count ahead: an allowance of credits, a "light quota", "basic weekly rate limits".

GitHub is the clearest of the group, and even it is only half-clear. Its plans page says Copilot Free is limited to 2,000 completions per month, with auto model selection only and no access to the Copilot cloud agent or third-party agents. Chat and agent work draw on "an allowance of GitHub AI Credits" - a phrase the page never puts a number against for the free plan, though it does for the paid ones: 1,900 credits per user per month on Copilot Business at $19 per seat, 3,900 on Copilot Enterprise at $39, with usage beyond the pool charged at $0.01 per AI credit.

Those paid numbers are the useful part. They tell you what a credit is worth in dollars, which lets you price an upgrade before you need it. In our view that is the most honest thing any of these pricing pages does, and the free tier would be easier to recommend if it carried the same arithmetic.

This shift in units is not unique to coding tools, but it bites hardest here, because one agent run on a large repository can consume what a week of autocomplete does not.

Is GitHub Copilot free, and what runs out first

Copilot Free is genuinely free, and in our view it is the easiest of these to live inside if your work is autocomplete-first. The 2,000-completion cap is the number to watch: GitHub's plans page notes that code completions and next edit suggestions are not billed in AI credits and stay unlimited on all paid plans, which is a plain statement that the free plan's pinch point is completions, not chat.

Two thousand completions is not a lot if you accept suggestions all day in an editor that offers one per keystroke pause. It is plenty if you write most code yourself and want a suggestion at the end of a line.

Skip Copilot Free if you want to choose a model. Auto model selection only means you cannot pin work to a specific model, which matters if you have decided a particular model is the one that gets your codebase right.

Does Cursor have a free plan, and what the free tier leaves out

Cursor's pricing page lists a Free plan at $0 with limited agent requests and limited chat requests, and no numbers beside either. Paid tiers start at $8 for Plus, with a footnote that on that plan only Grok models are available, then $30 for Premium, $100 for Super and $200 for Ultra.

The sentence that matters most is further down the same page: "Third Party model usage in Cursor is not included in the monthly prices shown and must be purchased separately." Cursor's FAQ repeats that Premium and above do not include an allowance for third-party models. So if your habit is Claude or GPT inside Cursor, the subscription is a platform fee and the models are a separate, usage-priced bill - we worked through what that costs in our post on the Cursor X subscription.

For a free-tier shopper that makes Cursor a trial rather than a home. You can see whether the editor suits you. You cannot predict what a month of real work on it will cost, because the published price does not contain the models most people want.

Devin and Google Antigravity: $0 tiers with unpublished limits

Two of the more generous-sounding free tiers belong to agent-first tools, and both buy that generosity with vagueness.

Devin's pricing page lists a Free plan at $0 with a "light quota to code with agents", limited model availability, unlimited inline edits and unlimited Tab completions, before Pro at $20 and Max at $200. Unlimited tab completion on a free tier is a real offer - it is the thing Copilot caps. The light agent quota is the unknown.

Google Antigravity's pricing page is structured the same way: Individuals at $0 with no subscription, access to multiple agent models, unlimited Tab completions, unlimited Command requests, and "basic weekly rate limits" whose numbers live behind an FAQ link rather than on the page. Business access runs through Google Cloud from $30 per seat per month, with a pay-as-you-go option at a $0 seat fee and consumption pricing.

Weekly is the word to notice. A quota that resets weekly rather than daily means a bad afternoon can cost you the rest of the week, which is a different risk profile from a monthly completion count.

What happened to Gemini CLI's free tier

If you are searching for Gemini CLI free limits, the answer is on the project's own repository: "Gemini CLI was replaced by Antigravity CLI on June 18th, 2026", applied to unpaid tier and Google One users. The repo also says Gemini CLI has not been optimized for Gemini 4 Argon and points individuals at Antigravity instead. Any 2026 listicle still quoting Gemini CLI's old free request-per-minute and request-per-day figures is quoting a tier that no longer serves those accounts - and several of the 2026 roundups still ranking for this query do exactly that.

Open-source coding assistants are free software, not free inference

The honest answer to "are open source AI coding assistants really free" is that the client is free and the thinking is not.

Cline's pricing page says the open-source extension is "completely free for individual developers", with usage billed separately - you pay only for AI inference, using your own keys from providers like OpenAI, Anthropic and Google, or Cline's own provider. There is no seat fee and no subscription, and the page mentions no free credits.

Kilo Code publishes the widest set of options we found. Individuals are $0 a month, with three ways to get models: Auto Free, which routes to free models through its gateway "where available"; your own key, including Anthropic, OpenAI, Google, Azure and AWS Bedrock; or a local model. Paid layers sit beside it rather than above it - Teams at $15 per user per month, and a Kilo Pass subscription from $19 a month for up to $26.60 in credits, with a 5% processing fee on credit purchases.

This is the category we would start a beginner in, with one caveat that applies to all of it: a bring-your-own-key setup can run to real money in an afternoon of agent work, and nothing stops it. Set a spend limit at your model provider before you set up the extension, not after.

One more caution before you commit to an open-source client. Continue, long a standard free pick, now carries a notice on its own site saying it was acquired by Cursor, with the open-source codebase remaining "freely available as a foundation for others". The site's own FAQ headings about data and subscriptions had no visible answers when we read the page on October 23, 2026. We would not start a new team on it this month without asking Cursor what happens to the hosted product, and we would keep the local, self-configured path in mind as the fallback.

As of October 2026, Upfynd lists 121 products in AI coding tools and 211 in open source, and the overlap between those two is where most of the genuinely free options live.

How to choose, in three rules

1. If you want a cap you can plan around, take the one that prints a number. Only Copilot's free tier gives you a figure - 2,000 completions - and only its paid tiers tell you what a credit costs in dollars.
2. If you want agent work for nothing, expect the limit to arrive without warning. The $0 agent tiers at Devin and Antigravity describe their quotas in adjectives. Try them on a real task and find out where they stop before you depend on either.
3. If you want control more than convenience, pay your model provider directly. An open-source client plus your own key removes the vendor's quota from the equation and replaces it with a bill you can read line by line - and a spend limit you must set yourself.

The one combination we would avoid is the opposite of all three: committing a team to a free tier whose limits are unpublished, then discovering the ceiling in the middle of a release.

The next step

Pick two of the free tiers above - in our view one capped and one open-source - and run the same real task through both this week: a bug fix in your own repository, not a toy script. The one that survives contact with your codebase is the one to pay for later.

Then check the pricing pages again before you upgrade. Every number in this post was read on October 23, 2026, and the units these limits are stated in keep changing. If you make a coding tool yourself, you can list it on Upfynd for free and say exactly what your free tier includes, which is a lower bar than most of these pages currently clear.