Free can mean a client, a small inference allowance, or a trial with restricted output rights. Choose the task first, then read the catch in the same row. Use the cost-saving playbook to count setup, retries and review time; the October buying guide owns dated paid choices.
Checked October 3, 2026 against provider documentation. Task fits are editorial suggestions. Account access, inference, latency, quotas and accepted-result outcomes were not tested.
Start Here
| Task | Resource | What is free | Catch | Details |
|---|---|---|---|---|
| Coding agent | OpenCode / Zen | Client and selected Zen offers | Data-use exceptions. Limited-time offers; billing details required. Check training, retention and auto-reload. | Zen roster |
| Internal API evaluation | NVIDIA trial API | Trial; no universal free quota verified | No production API or output. No confidential input; personal data needs express service permission. Content can support model improvement. | Trial rules |
| Text/audio prototypes | Groq | Limited API allowance | Model-specific limits. Retention exceptions for stateful features and reliability/abuse logs; check controls. | Limits · Data controls |
| Multimodal experiments | AI Studio / Gemini API | Studio and eligible API quota | Improvement use and human review. Avoid sensitive inputs; regional/billing exceptions below. Model-specific quotas. | Data exceptions · Terms |
| Small deployed features | Workers AI | 10,000 Neurons/day | Metered compute. Separate Workers limits; some models require paid billing. Free-plan requests stop at allowance. | AI pricing · Workers limits |
| Model comparison | OpenRouter | Limited requests to free variants | Limited requests. Capacity varies; upstream training/retention policies differ. | Limits · Provider policies |
| Agent IDE | Antigravity | $0 Individual plan | Weekly limits. Review interaction-data controls before sending project context. | Current access |
| BYO agent | AMP | Agent harness | Supply compute and model access. Old daily grants are historical. | Current access |
| Coding client | Kilo | Client; changing model offers | Inference/compute separate. BYOK, local and promotional routes have different terms. | Kilo guide |
| Local/offline | Local inference | No hosted API charge | Hardware and license constraints. Memory, electricity and setup still cost; check runtime networking. | Local guide · Gemma |
Check the data path
Gemini data exceptions
Google’s terms apply paid-service data handling throughout the EEA, Switzerland and UK, including unpaid quota. Elsewhere, Studio qualifies when your account has access to an actively billed Cloud project or is a Workspace enterprise account; API access qualifies through an actively billed project. Paid-service prompts/responses are not used for product improvement, but safety/legal logging remains. Serving API clients to users in those European regions requires Paid Services.
For Groq, use Data Controls to check zero-retention settings; persistence features and abuse/reliability exceptions make an unconditional privacy promise inaccurate. For OpenRouter, review both its own settings and the selected provider’s policy. Training opt-out is not a blanket no-retention guarantee.
For proprietary work, a free price alone is insufficient permission. In particular, NVIDIA’s trial terms exclude confidential inputs and production use of the service and generated content. The Kimi access guide separates model listings from tested entitlements.
FAQ
What is the best free AI coding tool right now?
Choose by task and restriction: a free client may still need paid inference, and a free hosted allowance may have data-use or trial limits. This directory does not rank an untested winner.
Where can I trial a coding API before paying?
Groq, Gemini API and OpenRouter document limited free access. NVIDIA is an internal-evaluation trial: neither the service nor its generated content may be used in production under the trial terms. Account eligibility remains untested.
Is Qwen still free on OpenCode?
Check the dated free-priced Zen roster in the OpenCode guide. A model name or an old promotion does not establish a current free entitlement.
Is Kimi still free anywhere?
NVIDIA lists Kimi K3, but its trial listing does not establish a universal no-payment quota. The trial service and generated content are excluded from production; use the Kimi access guide for current route boundaries.
What should I use if I want local or on-device AI?
Use the local-inference guide to choose a runtime and model that fit your hardware. Local inference avoids hosted API charges, but hardware, electricity, setup and model-license constraints remain.
Are ChatGPT Free and Claude Free good enough for coding?
They can help with occasional questions. For repeatable coding work, compare the workflow, allowance and data rules in this directory, then test a bounded task. No comparative outcome was measured here.
Related links
For the live model list, use the OpenCode roster. For a paid upgrade, return to the buying guide and evaluate cost per accepted result.
- Toybox Clock protocol — Protocol preview—results pending; historical seven-route snapshot, with current availability documented separately.