On September 23, 2026, how many free AI APIs are still available?

Info: Machine translation This post was machine-translated from the Chinese original. Wording may be rough in places — the Chinese version is authoritative.

Summary Special Topic Date: 2026-09-23 · Second Review: 2026-09-23 HyphenTech

Tip This issue, strictly speaking, three platforms passed at least three valid calls and one longer real-world task from my Mac network perspective; This does not mean long-term stability or universal availability.

Warning Official pages only prove the existence of policies or models; whether the account can be accessed depends on the current request; No credentials, different protocols, or insufficient page crawling are listed separately and not marked as invalid.

How do I retest, and where exactly is the free one stuck with?

I first list the model locally, then do short Q&A and slightly longer tasks on the same interface. Being able to list is just the first gate; Only when you actually get the words out can you cross the second gate. Gemini’s dialogue and vector interfaces pass separately, and Groq’s specified model also continuously returns valid body text. This is a record from my account and my network perspective, not a valid guarantee for everyone.

The most common misjudgment is the word “free.” Who profits? The platform gets developers’ attention and trial traffic; Who bears the cost? The caller spends time applying, migrating, and handling traffic restrictions, while providers bear the inference resources. My judgment is: the free tier is more like a boundary entry point than infrastructure that can be relied upon infinitely. This commercial motivation judgment is not an official platform promise, so I separate it from the original text of price and quota.

The failures encountered today are not the same problem: silicon-based flows return insufficient balances in this account; Pollinations public model directories are readable, but actual anonymous requests repeatedly report server errors; OpenRouter and Agnes fail first and recover later. Writing all these issues as “model lost” will bias the selection. Especially note: although the Agnes tested model did return text, the official website’s free display did not fully correspond to the tested model, so I did not count it in the platform count in the title.

For regular users, I first try using short tasks without privacy, then check slightly longer tasks, and finally observe the rate limit recovery. Don’t just stuff customer data, private warehouses, or keys into a free endpoint that hasn’t read the privacy policy yet. The itemized table at the end of the main text keeps the missing field as “pending verification”; It may look better than filling it completely, but it’s more practical than copying old quotas into today’s policies.

One-sentence recommendation

• This issue’s real requests passed: OpenRouter, Groq, Google Gemini API; Test based on your account and region first.
• Gemini’s Chat and Embedding are verified separately, without using Chat error requests to misjudge Embedding.
• Free entry points carry different risks such as 402, 429, 502, and 503; Long tasks and continuous call records are more important than zero-price web pages.

Verify and investigate promptly

Special issue date: 2026-09-23; status reviewed on 2026-09-23

Warning Only tests the legally held credentials and the public keyless interface; No sharing or trading keys, and no restrictions on verification codes, real-name registration, payment, region, or account limits.

OpenRouter

Retest can be called; First test 503, non-stable commitment

Warning Different free models are not identical in availability; 429, 503, and account balance constraints must be viewed separately.

• Official source: https://openrouter.ai/docs/api-reference/limits

• Official source: https://openrouter.ai/nvidia/nemotron-3-super-120b-a12b%3Afree

• This issue’s detection: openrouter/liquid/lfm-2.5-2.6b: free · chat · Effective 0/4 · probe-report.local.json

• This issue’s detection: OpenRouter/nvidia/Nemotron-3-super-120b-a12b:free · chat · Effective 3/4 · probe-report.recheck.json

SiliconFlow

Insufficient account balance; Free platform models cannot be declared invalid based on this

Warning The official announcement states that account usage requires real-name verification; Retiring the old PaddleOCR-VL does not mean retiring the PaddleOCR-VL-1.5.

• Official source: https://siliconflow.cn/pricing

• This issue’s detection: siliconflow/Qwen/Qwen2.5-7B-Instruct · chat · Effective 0/4 · probe-report.local.json

Zhipu BigModel

Visible in the official directory; Account-level calls pending verification

Warning Old models and quotas must not be directly reused.

• Official source: https://docs.bigmodel.cn/cn/guide/start/model-overview

Magic ModelScope

Account-Level Calls Pending Verification

Warning You can’t treat the August list as the available model for September.

• Official source: https://www.modelscope.cn/docs/model-service/API-Inference/intro

Groq

Usable for this issue; Stability varies among different models

Warning Different models on the same platform cannot share stability conclusions.

• Official source: https://console.groq.com/docs/rate-limits

• Official source: https://console.groq.com/docs/your-data

• Official source: https://console.groq.com/docs/deprecations

• This issue’s detection: Groq/OpenAI/GPT-OSS-20B · chat · Effective 2/4 · probe-report.local.json

• This issue’s detection: groq/qwen/qwen 3.8-27b · chat · Effective 4/4 · probe-report.recheck.json

Cloudflare Workers AI

Official free file confirmation; Account-level calls pending verification

Warning If you exceed the limit, the free plan will fail; Neurons cannot be converted into a fixed number of chats.

• Official source: https://developers.cloudflare.com/workers-ai/platform/pricing/

Agnes AI

Interface retesting can be called; Zero-price qualification of tested model pending verification

Warning The official Free API promotion cannot automatically apply to 2.5 Flash; You cannot equate three consecutive successes with long-term stability.

• Official source: https://agnes-ai.com/

• This issue’s detection: agnes/agnes-2.0-flash · chat · Effective 1/4 · probe-report.local.json

• This issue’s detection: agnes/agnes-2.5-flash · chat · Effective 3/4 · probe-report.recheck.json

Hugging Face Inference Providers

Official free quota confirmation; Account-level calls pending verification

Warning BYOK routing charges differ and cannot be understood as unlimited calls from the platform.

• Official source: https://huggingface.co/docs/inference-providers/pricing

NVIDIA NIM (build.nvidia.com)

Official portal confirmation; Account-level calls pending verification

Warning A successful NVIDIA model forwarded by OpenRouter does not necessarily mean the NVIDIA direct connection port is tested successfully.

• Official source: https://build.nvidia.com/models

Google Gemini API

This issue’s Chat intersection is free to pass; Embedding calls are valid, zero-price pending approval

Warning Free tier requests may be used to improve Google products; Sensitive content should not be included. Model, region, and account eligibility affect the free tier.

• Official source: https://ai.google.dev/gemini-api/docs/rate-limits

• Official source: https://ai.google.dev/gemini-api/docs/pricing

• This issue’s detection: gemini/gemini-3.1-flash-lite · chat · Effective 4/4 · probe-report.local.json

• This issue’s detection: gemini/gemini-embedding-001 · embedding · Effective 4/4 · probe-report.local.json

Alibaba Cloud Bailian Model Studio

Policy update confirmation; Account-level calls pending verification

Warning This is not a long-term free period; The terms for historical account holders do not change automatically due to announcements.

• Official source: https://www.alibabacloud.com/help/zh/model-studio/new-free-quota-validity-adjustment

iFlytek Spark

Official agreement confirmation; Account-level calls pending verification

Warning Some endpoints are not Chat; The key trio cannot be written to the probe report.

• Official source: https://www.xfyun.cn/doc/spark/Web.html

Cerebras

Account-Level Calls Pending Verification

Warning If you need to bind a payment method, it cannot be included in the no-threshold free tier.

• Official source: https://inference-docs.cerebras.ai/support/pricing

Together AI

Account-level free qualification and call verification pending

Warning Dynamic speed limits should not be mistakenly written as fixed free quotas.

• Official source: https://docs.together.ai/docs/rate-limits

Pollinations

This period, all four calls were 500; failed to use the real energy gate

Warning Model discovery does not mean Chat is available; You cannot use the August test to replace this issue.

• Official source: https://text.pollinations.ai/models

GitHub Models

Retired and not considered as a candidate

Warning Old links and old quotas must not continue to be promoted.

• Official source: https://docs.github.com/en/github-models

Open source hosting (Ollama / llama.cpp / whisper.cpp / Piper)

Self-hosted but not a public free cloud API

Warning Requires self-deployment, maintenance, and bearing hardware and energy costs.

• Official source: https://github.com/ggml-org/llama.cpp

OpenCode Zen

Official Candidate; Account-level calls pending verification

Warning Do not use document names as interfaces; Do not bypass registration, real-name verification, or payment verification.

• Official source: https://opencode.ai/docs/zen

FreeInference

Official Candidate; Embedding Call to Be Verified

Warning Do not use document names as interfaces; Do not bypass registration, real-name verification, or payment verification.

• Official source: https://doc.freeinference.org/models

USTC Public Model Service

Not part of the public free API, not included in recommendations

Warning Do not use document names as interfaces; Do not bypass registration, real-name verification, or payment verification.

• Official source: https://llm.ustc.edu.cn/guide/technical-updates/

Ant larks

Insufficient page crawling, not included in available statistics for now

Warning Do not use document names as interfaces; Do not bypass registration, real-name verification, or payment verification.

• Official source: https://www.ant-ling.com/

Key reminder

• Groq’s official retirement records show that the qwen/qwen3.6-27b was discontinued as of September 14 for free/developer level calls; The exact model routing for this machine is disabled, but the successor qwen/qwen3.8-27b remains available.
• The silicon-based mobile account 402 only indicates insufficient account balance and cannot declare the platform’s free services invalid.
• Pollinations’ public directory is readable this issue, but anonymous chat returned 500 four times.
• Both OpenRouter and Agnes experienced failures followed by recovery, making them unsuitable for promising uninterrupted production.
• Institution-limited, one-time grants, and free cloud API counts that do not participate in public cycles of free cloud API counting from open-source custody.

My access priority

• First, non-sensitive testing was conducted using Gemini Chat/Embedding and Groq-specified models, which were valid on all four real-world calls in this issue.
• OpenRouter and Agnes can be alternatives, but clients should have explicit error messages and switching policies.
• For platforms without account-level verification, first check the official backend rights and privacy terms, then make the minimum request of the same type.

Summary: Final conclusion According to this issue’s legal and genuine requests, three platforms reached the minimum “callable” threshold; This is not a long-term stable leaderboard. What was not tested is not invalid, and the zero price on the ad page is not a true return all at once.

🧰 Tools I build

I maintain all of these tools myself. Preview builds are clearly labeled; the release pages are the source of truth for downloads, updates and known limits.

Info: HyphenBox Status: Official releases

A radar for free LLM APIs: availability is re-tested continuously, one local interface for all of them, and keys stay on your machine

Downloads & updates

Info: LocalBrain Status: Official releases

A multimodal MCP toolbox for local models: TTS, Whisper and video generation in one place

Downloads & updates

Info: ScreenLex Status: Official releases

Learn new words while you watch shows. Free, for Mac and Windows

Downloads & updates

Info: HyphenScreen Status: Official releases

Screen recording and smart editing in one: a DaVinci-style timeline, automatic redaction and a check of the finished video before export. Free

Downloads & updates

Quote: HyphenTech Make AI your superpower Local deployment · Free resources · Self-made software https://hyphentech.top