Free stopped meaning "trial"
Two years ago, free meant fourteen days and then a card form. What GitHub, Cursor, Google, AWS and Cognition ship now is different: a free tier that keeps running, with no expiry date, because inference got cheap enough to hand out and because a free user is a seat that might convert. That sounds like generosity. It is a product decision, and it shapes what you can build with it.
The important detail is that the free tier is not the paid product with the meter switched off. It is a deliberately different product. Copilot Free runs auto model selection only, so you never choose the model. Cursor Free has no third-party models and no bring-your-own-key. Claude Free does not include Claude Code at all. ChatGPT Free keeps the chat but throttles Codex. Vendors are not hiding this; it is in the plan tables. It just does not sink in until the first afternoon when the tool you were relying on stops halfway through a refactor.
There is a second kind of free that gets mixed into the same conversation: open-source software with no inference attached. Cline, Continue, OpenCode, Pi and Zed install for nothing and run on your key. That is a real way to work, but the software being free tells you nothing about the tokens. You have moved the cost, not removed it.
So "can I vibe code for free" has a boring answer: yes, and the interesting question is what kind of work survives the allowance.
What $0 actually buys
Copilot Free is the most generous-feeling offer and the most limited one in practice. It gives you 2,000 code completions a month, which is plenty for autocomplete and almost nothing for an agent that reads files and proposes edits. There is chat and agent usage on top, capped and metered in GitHub's AI credits, and the model is chosen for you. If your work is "finish this function", it holds up. If it is "move this module onto the new API", it will stall in the middle.
Antigravity Free is the opposite shape: unlimited tab completions and command requests, with what Google calls basic weekly rate limits on the agent. Kiro Free is 50 credits a month, which sounds small because it is. Warp Free gives you the terminal, three indexed codebases and thirty cloud conversations, with agent credits sold separately. Replit Starter limits agent runs and keeps your apps public. Devin Free gives you a daily quota in the desktop app and CLI and no cloud agent capacity.
The chat subscriptions are their own category. ChatGPT Free has the most generous chat of anything here and the weakest agent: Codex works, at a low ceiling. Claude Free is a good chat product with no Claude Code, which is the part developers actually pay for.
Then there is the route that most budget guides miss: an open-source agent pointed at a free model endpoint. Two of those endpoints matter. Google's AI Studio free tier gives you Gemini models with per-model limits on requests per minute, tokens per minute and requests per day, and preview models get the tightest caps. OpenRouter's free variants are capped at 20 requests a minute, with a daily ceiling of 50 requests until you have bought $10 of credit in your lifetime and 1,000 after that. The $10 does not buy the requests. It buys you out of the small daily bucket, permanently, and it stays in the account as credit for paid models.
That last sentence is the single most useful thing in this brief. If you are going to spend anything at all on a free setup, spend $10 there before you spend it on a subscription.
The week the limits arrive
Free tiers fail on a schedule, and the schedule is roughly "day three". Day one feels unlimited because you are writing new code, which is what these tools are best at. Day three is when you start changing code that already exists, and an agent that has to read four files, edit two, run the tests and repair one failure will spend somewhere between five and twenty model calls on a single task. Call it ten as a working average. At 50 requests a day that is a handful of tasks. At 20 requests a minute you can forget about running two agents at once.
This is the part the marketing pages leave out, and it is not dishonesty so much as a different unit. Vendors count completions, credits, messages or requests. You experience tasks. The conversion between the two depends entirely on the shape of your work, which is why two people on the same plan can have completely different opinions about whether it is usable.
Free tiers also get re-cut. Continue removed its free tier in 2026 and now charges $3 per million tokens through Continue Starter. Amp closed its free tier to new signups and sells pay-as-you-go at list price. Neither vendor broke a contract. The free tier was the marketing, and the marketing changed. Build your workflow so that losing one free tier costs you an afternoon, not a project.
Are the free models good enough?
Mostly, yes, and the numbers are closer than the pricing suggests. The models free tiers tend to hand you sit between 34 and 45 on the Artificial Analysis intelligence index: GPT-6 Luna at 38, Gemini 3.8 Flash at 41, GLM-5.3 Flash at 42, Claude Haiku 5.5 at 43, DeepSeek V4.1 Flash at 45. The paid flagships sit at 53 to 58: GPT-6 Astra 53, Gemini 4 Argon 53, Claude Sonnet 5.5 56, Claude Opus 5.5 58.
A ten to fifteen point gap does not show up on a landing page or a small script. It shows up when a task spans a dozen files and the model has to keep a plan straight across all of them. That is the thing you are buying at the top end, and it is why the cost per task numbers look so lopsided: $0.07 for GPT-6 Luna against $5.98 for Claude Opus 5.5 on the same task set. You are not paying for tokens. You are paying for the percentage of tasks that come back finished.
One row breaks the pattern in a useful way. DeepSeek V4.1 Flash scores 45, runs at 216 tokens a second, has open weights, and costs $0.26 per task: eight points behind GPT-6 Astra at roughly a twelfth of the cost. That is why so many budget plans resell it, and why a free variant of it is the best thing you can put behind a free agent.
My read: free models are worth it for learning, prototypes and side projects you would finish anyway. They are not worth it for client work with a deadline, where one lost afternoon costs more than a month of Cursor Plus.
The privacy line moved in 2026
This is the part people skip, and it is more consequential than the rate limits. Free tiers and training data are separate questions, and the three big vendors answer them differently. GitHub changed its policy on 24 April 2026: interaction data from Copilot Free, Pro and Pro+ users, including inputs, outputs and code snippets, is used to train its models unless you opt out. Business and Enterprise are excluded. OpenAI keeps "Improve the model for everyone" switched on by default for personal Free, Plus and Pro workspaces, and that setting covers Codex tasks as well as chat. Anthropic goes the other way: consumer chats and coding sessions are used for training only if you opt in, or if a conversation gets flagged for a safety review.
Two out of three defaults are opt-out, so the default is the thing worth checking. If you are pasting client code, do it before the first prompt rather than after the first surprise.
That is the argument. What follows is the reference version: every zero-price plan in the dataset, the model numbers behind the claims above, the training defaults in one table, and the arithmetic for when paying is cheaper than working around a cap.