Free LLM API Tiers Ranked 2026: Gemini, xAI, DeepSeek, AWS — Which Free Credits Are Actually Usable for Coding?

For a no-cost coding evaluation, start by checking a Gemini model that currently lists a free API tier. Treat DeepSeek as a paid option unless your account shows a grant, xAI credits as promotion-dependent, and AWS credits as account-dependent. There is no defensible universal winner across these four offers.

This September 7 update corrects the earlier article’s universal signup-credit claims, fixed Gemini quotas and outdated DeepSeek prices. This is a comparison of published terms, not a new coding benchmark. It does not claim that we ran four equivalent agent workloads.

The ranking, with eligibility made explicit

The order below is the order to investigate for a small coding evaluation, not a model-quality ranking.

If you already have eligible AWS credits, Bedrock may be your first option. If you need a particular model or tool feature, start there and then check the funding route.

Gemini: check the model and your project’s actual quota

The official Gemini API pricing page lists free input and output for selected models, including Gemini 3.1 Flash-Lite at the time of this check. Free availability differs by model and operation: a free text request does not imply free grounding, caching or every other feature.

Do not budget from the previous version of this article’s 1,500 requests-per-day and 15/30 requests-per-minute table. Check the rate-limit documentation and your AI Studio rate-limit panel. Limits apply to projects rather than multiplying with each key. Model access, request limits and token limits can each constrain an agent.

The pricing page also distinguishes data use: free-tier data may be used to improve Google’s products, while the paid-tier columns have different treatment. Read the terms that apply to your account before submitting private source code.

A useful trial is a small task from your own codebase: generate a function, read a file, complete one tool interaction, and run its tests. Free-tier eligibility does not establish whether the model can handle your agent’s protocol or long context.

DeepSeek: verify grants separately from paid prices

The DeepSeek pricing page describes deduction from granted and topped-up balances, with granted balance used first. It does not establish the earlier article’s promise that every new account receives five million tokens for 30 days. Check your account’s actual balance and expiry.

Current published prices per million tokens:

The page defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday; other hours are off-peak. These are direct-provider prices, not an Ofox price quote.

For example, 50,000 uncached input tokens plus 15,000 output tokens on V4 Flash cost $0.0209 off-peak or $0.0418 at peak, before any additional requests. That is arithmetic from published prices, not an observed coding-session cost. Agent loops, retries and accumulated context can change the total substantially.

For more detail, see the DeepSeek price-change guide and V4 Pro versus Flash. A low paid rate should not be described as free access.

xAI: budget from the promotion shown in your team

The current xAI billing documentation describes prepaid credits, promo codes and monthly invoiced billing. It does not substantiate a universal $25 signup grant or a guaranteed $150 monthly grant after spending $5.

The earlier article also said that enabling data sharing was permanent and that every prompt would become training data forever. Those blanket claims are not established by the billing source. If your team is offered a data-sharing promotion, read its current eligibility, scope, retention and withdrawal terms before enabling it. Absence from the billing page does not prove that no targeted promotion exists.

Check auto top-up and invoice settings as well as promotional balance. A promotion running out and a paid balance automatically renewing are different events. Estimate costs using the exact model and current rates in the console; do not convert a remembered promotion into a recurring budget.

AWS Bedrock: credits, account plan and model access are separate

AWS Free Tier advertises $100 at signup and up to $100 more from eligible activities. The Free plan lasts up to six months or until credits run out, whichever happens first. That plan duration is not a universal expiration rule for every AWS promotional credit.

Before treating credits as a Bedrock budget, verify:

• Your account plan allows the service you intend to use.
• Your particular credit applies to the usage and has sufficient remaining validity.
• The selected model is accessible in your AWS region and account.
• Associated storage, logging and compute costs are included in your budget.

AWS Activate is a separate startup program with eligibility and offer-specific terms. Do not assume approval or a particular grant amount. See Bedrock pricing for model and inference-mode charges.

We removed the earlier “one weekend” estimate: a dollar credit cannot be converted into weekends of agent work without a workload and measured token use.

How to compare a coding workload fairly

Use the same repository snapshot, task, allowed tools and success criteria for each candidate. Record model ID, date, request count, input/output tokens, cache use, errors, elapsed time and whether the tests passed. Start with a scratch checkout.

For uncached requests, the basic estimate is:

cost = input_tokens / 1,000,000 × input_rate
     + output_tokens / 1,000,000 × output_rate

Add cache charges, tools, storage and retries where applicable. A daily request limit is not a session count: one agent task can make many requests. Compare cost per successful task, rather than treating a cheaper failed run as a win.

When a gateway becomes useful

If you need several providers in one application, a gateway can centralize keys and billing. Check Ofox’s model catalog for the exact model, price and protocol before configuring it. Provider-native free credits do not automatically transfer to a gateway, and a shared account does not guarantee every model works through every protocol.

For Codex, follow the Responses-based custom provider configuration. For Claude Code, use the backend switching guide. The pricing index links to the vendor-specific checks when you move from evaluation to paid usage.