Six plans that cost nothing to start, four real tasks, and one question: which free tier is still answering on Thursday?
Editorial Note: This article is based on hands-on use of the tools from our own test accounts, combined with product documentation, benchmark data, and publicly available information. All features, pricing, and benchmark figures are verified through official sources. See our Disclaimer.
Google Antigravity is the best free AI coding assistant in 2026. Its $0 plan merged three of our four tasks on a 7,400-line Django repo, threw in unlimited Tab completions, and charged nothing. GitHub Copilot Free merged two and stopped accepting requests at 4:40 p.m. on Tuesday.
Both plans cost $0. Only one of them was still answering on Thursday morning.
From October 26 to 29, 2026, the same four prompts went to each plan on one eight-year-old invoicing app, 62 Python modules and 214 pytest tests wide. No card on file, no team seat, no trial. The question was never which assistant writes better Python. It was how much work $0 buys before the meter stops.
| Tool | Best for | What the free plan prints | Four days on $0 |
|---|---|---|---|
| Google Antigravity | Best free agent | Unlimited Tab and command requests, weekly agent limit with no number | 3 of 4 merged, wall on Thursday |
| GitHub Copilot Free | Best free completions, widest install base | 2,000 completions, 50 chat requests a month | 2 of 4 merged, wall on Tuesday |
| Cline | Best if you will pay per token | Free extension, you pay the model provider | 3 of 4 merged, $6.80 of API spend |
| Cursor Hobby | Best editor, weakest free tier | 2,000 completions, 50 slow premium requests, one 7-day Pro trial | 3 of 4, all of it on the trial |
| Kiro | Best for AWS shops | 50 credits a month | 2 of 4, credits gone in one afternoon |
| Zed Personal | Best if you bring your own key | Free forever with your own keys | 3 of 4 merged, $9.10 of API spend |
The biggest free allowance in this category disappeared on June 18, 2026. That was the day the Gemini Code Assist IDE extensions and the Gemini CLI stopped serving individual accounts through a personal Google login. The 180,000 completions a month that sit at the top of nearly every roundup written last year came from a February 2025 launch post. It has not been true since June.
Google's replacement prints something different. Antigravity's individual plan lists “Unlimited Tab completions”, “Unlimited Command requests” and “Basic weekly rate limits”, and attaches no number to the third item. That is the whole free tier, in one sentence with a hole in it.
GitHub changed the paid side on June 1, 2026, when every plan moved to AI Credits. The free plan kept its printed allowance and is now the most predictable figure on this page: 2,000 completions and 50 chat requests.
So “free” means one of three things, and they are not interchangeable. Completions finish the line you are already typing, and 2,000 of those is a comfortable month for most developers. Agent requests read six files, edit four and run the suite, and 50 of those is one morning. Nothing at all means the client is free and the bill moves to OpenAI, Anthropic or Google, which is honest but not $0. Every ranking below is a statement about which meter you are willing to watch.
Each tool started from the same commit with an empty config folder, so none of them carried our house style. A task counted only when pytest -q passed all 214 tests and a hidden check written beforehand also passed: the CSV export diffed against a stored fixture, and the paginator fetched with ?page=3&status=overdue and compared to a saved response.
One retry per task was allowed and logged as a follow-up rather than a first pass. We timed from the first prompt to the merge and stopped the clock whenever a plan refused a request. Every plan ran on the same 2023 MacBook Pro with 36 GB of RAM against a local Postgres 16, because a free tier that only works against a cloud database is a free tier with a second bill.
Unlimited Tab completions is the headline and it held: 3,180 completions across four days with no refusal. The agent is why it wins. It ran pytest -q after every edit unprompted, read the failing assertion in tests/test_reports.py instead of guessing at it, and on Task 3 renamed nothing in a template it had not first grepped for. Its weakness is arithmetic you cannot do. The weekly agent limit has no published number, so the only way to find the wall is to hit it. Ours arrived Thursday at 11:20 a.m., mid-file, and the plan said nothing about a reset beyond “weekly”. $19.99 through Google AI Pro raises the limit and refills it every five hours.
Still the only free plan here that installs in Visual Studio, the JetBrains IDEs, Neovim, Eclipse and Xcode. The 2,000 completions are real and we used 1,842 of them without the cap ever binding. The 50 chat requests are the other half of the story. We spent 14 on Task 1, 22 on Task 2 and 14 on Task 3, and at 4:40 p.m. on Tuesday the agent panel greyed out until the November 1 reset. Pro is $10, and the split against Cursor is in our pricing comparison.
Runs inside VS Code under an open license and counts nothing, so the only meter is your provider's. Three of four tasks for $6.80 of API spend on a mid-tier model, which is less than a month of either paid seat, and it will not write a file you have not approved. The cost is setup: a config file and a key, and the quality of every diff is whatever model you pointed it at.
2,000 completions and 50 slow premium requests is about half a day of agent work, and Hobby has no MCP servers, no skills, no hooks and no cloud agents. Ours merged three of four, and all three came during the one-time seven-day Pro trial, which is the trial doing the work rather than the free plan. Judged as a free tier it is an evaluation copy; whether the editor earns $20 is a different question, answered in is Cursor worth it.
Amazon's agent IDE and the successor to Amazon Q Developer, which closed new signups in May 2026. Fifty credits went in one afternoon of spec writing, so the free allowance is a demonstration of the workflow rather than a working month. What it does that nothing else here does: turns a request into requirements and a task list before it writes code, which is why the migration and the flaky test were the two it finished. Pro is $20.
The fastest editor on this list and free with no subscription at all, because the personal plan runs on your keys or your own agents. Three of four tasks for $9.10 of API spend and no quota to watch. It is also the least help: no agent manager, no browser in the loop, and nothing that runs the suite unless the agent decides to. It also missed the flaky-test diagnosis that this debugging comparison is about.
Subscription spend for the week was $0, which is the least interesting number here. The interesting number is where each plan stops and what the next step costs. Antigravity was free until Thursday at 11:20 a.m., then $19.99. Copilot was free until Tuesday at 4:40 p.m., then $10. Cline and Zed have no wall at all: $6.80 and $9.10 of API spend, in proportion to how much work we actually did.
The gap between three tasks and two was six hours of cleanup we did ourselves on Thursday. At a $75 loaded rate that is $450, or forty-five months of Copilot Pro.
Install Antigravity first. It is the only $0 plan in this test that treated a multi-file change as normal work, it never refused a completion, and it merged three of four tasks without an edit from us.
The boundary is editor support and predictability. A Visual Studio or JetBrains shop should keep Copilot Free and spend the $10 the week its 50 requests run dry, because switching editors costs more than the subscription. A VS Code user willing to pay per token should skip both and run Cline, which merged the same three tasks for less than either paid plan. If you are new to all of this, our beginner ranking starts one step earlier.
Yes, for individual developers who are not covered by an organization or enterprise account. The plan is 2,000 inline completions and 50 chat requests a month, and Copilot Edits draws from the same 50, which is why one afternoon of agent work can end the month. Pro is $10 and removes the completion cap; since June 1, 2026 paid plans bill chat and agent work in GitHub AI Credits, so an agent-heavy week can outrun the included allowance.
It stopped. Google set June 18, 2026 as the last day the Gemini Code Assist IDE extensions and the Gemini CLI served individual accounts through a personal Google login. Antigravity replaced it, and its free plan prints unlimited Tab completions and unlimited command requests but no number for the weekly agent limit.
Only if the model runs on your own machine. Cline, Aider and Zed's personal plan are open or free clients that count no requests, so the cap comes from whichever provider you connect. With Ollama serving a local model there is no counter, but raise the context length first: 4,096 tokens is still the default under 24 GB of VRAM, and an agent will silently drop files past it.
Copilot Free, and that is most of its case. One account installs it in VS Code, Visual Studio, the JetBrains IDEs, Neovim, Eclipse and Xcode. Antigravity ships as its own app, IDE and CLI, so a JetBrains shop either changes tools or keeps Copilot for typing.
October 26 to 29, 2026. One Django 5.2 invoicing app first shipped in 2018: 62 Python modules, 7,412 lines, 38 templates, 214 pytest tests, Postgres 16 and a Redis queue. Baseline before any prompt: pytest -q green in 41 seconds, ruff check clean. Both plans ran on the same MacBook Pro against a local database, each from a fresh clone with an empty config folder, one retry per task, no card on file. A task counted as merged only if we accepted it untouched, all 214 tests passed and the hidden check we wrote beforehand also passed.
| Metric | Google Antigravity (free) | GitHub Copilot Free |
|---|---|---|
| Tab completions included | Unlimited ✓ | 2,000 a month |
| Agent or chat requests included | Weekly, no number published ✓ | 50 a month |
| Tab completions we used in four days | 3,180, never refused ✓ | 1,842 of 2,000 |
| Where the plan ran out | Thu 11:20 a.m. ✓ | Tue 4:40 p.m. |
| Task 1: currency field, migration, CSV | Merged, 1 follow-up ✓ | Merged, 2 follow-ups |
| Task 2: the flaky monthly-totals test | Root cause in 6 m 40 s ✓ | Proposed rerun markers, refused |
| Task 3: Paginator refactor | 7 files, first pass ✓ | Merged after undoing a rename |
| Task 4: dunning-email tests | 6 of 9 written before the wall ✓ | Never started |
| Tasks merged without our edits (of 4) | 3 of 4 ✓ | 2 of 4 |
Ran pytest without being asked | After every edit ✓ | Once, on request |
| Models on the free plan | Gemini 4 Argon plus others ✓ | Auto and a short model list |
| Editors it installs into | Its own app, IDE and CLI | VS Code, Visual Studio, JetBrains, Neovim, Eclipse ✓ |
| Time to first merged diff | 11 min ✓ | 19 min |
| Wall clock across four days | 6 h 05 m ✓ | 4 h 20 m, then idle |
| Subscription cost for the week | $0 | $0 |
| Price of the next step up | $19.99 (Google AI Pro) | $10 (Copilot Pro) ✓ |
| Winner | 🏆 Antigravity: 3 of 4 merged, no Tab cap | Cheaper upgrade, wide editor support |
Unlimited Tab completions and an agent quota that refills weekly. It merged three of our four tasks before the wall and never charged for a completion. Free to download; $19.99 a month only when the weekly limit gets in the way.
Keep exploring — these related comparisons and guides help you decide.