2026年9月免费大模型API横评:5家实测额度对比2026 September Free LLM API Benchmark: Quota Comparison Across 5 Providers


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/MsrAzTChwgk5xaComJvaBg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

免费API实测(2026年9月):免费额度难道仅限一次性?其实仍有5家平台可供免费使用,但仅有一家敢于在定价页面明确标注。涵盖智谱、火山引擎、硅基流动、阿里百炼、百度千帆及讯飞。2026年9月免费API实测横向评测与额度对比。
📦 8个部分 + 结论 👉 滑动查看 PART 01 先说结论:没有绝对最好,只有最匹配的 PART 02 智谱 GLM-4.7-Flash:唯一敢承诺永久的 PART 03 火山引擎豆包:唯一每天进行重置的 PART 04 硅基流动:仅需一个 Key 即可调用 100+ 模型 PART 05 阿里云百炼:额度充沛,但存在过期限制 PART 06 百度千帆 · 讯飞:号称永久,但响应速度递减 PART 07 海外三家平台简单概述 PART 08 三点提示建议,推荐阅读后再操作 PART /// 结语:我的个人建议
本次横向评测的缘由:半年前能随口说出的答案,如今已不再适用
上个礼拜一位朋友打听我,眼下哪款免费大模型 API最实用。我刚想回答,却意识到自己根本说不上来。半年前我还能立刻报出一个名字和一连串免费额度,可上个月按照教程去调用时,系统却提示模型不存在——原来那批模型早就下线了。因此这几日我把国内尚可使用的平台重新挨个注册了一番,仔细比对各家的官方定价页面和说明文档。以下列出的数据均为 9 月 1 日当天核实的结果,由于免费额度变动频繁,建议你实际操作前最好再去页面核对一遍。
2026年9月免费大模型API横评:5家实测额度对比

01

PART

先说结论

CONCLUSION · 先看结果

并不存在所谓“最好用”的选项,只有“最契合你当前需求”的。

你要干什么
选哪个
理由
长期免费白嫖、不愿操心续费
智谱推出的 GLM-4.7-Flash
唯独它的官方定价为 0 元,并且没有任何使用期限限制
每天固定跑批量脚本
火山引擎(豆包)
每日额度在 200 至 500 万之间进行自动重置
凭借单个 Key 即可体验各类模型
硅基流动
汇集了超过 100 个模型
跑批量任务、要并发
百度千帆平台的 ERNIE-Speed
QPS 给到 50
长文档、超长上下文
智谱(200K)或者 Gemini(1M)
看你能不能连上
倘若不想阅读全文,建议先前往智谱创建一个 Key 备用,这是唯一一个我敢断言“明年大概率还能用”的
02PART智谱 GLM-4.7-Flash:唯一敢承诺永久免费的 ZHIPU · GLM-4.7-FLASH

经过一番调查,这是我最为推荐的选项,原因非常简单:在官方的价格页面上,glm-4.7-flash 的输入与输出费用均标注为 0 元,既没有“限时”的标签,也不属于“新使用者专享”福利,它本身就是一款免费模型。同时,官方的更新日志中也明确记载,GLM-4.7-Flash 已于 2026 年 1 月 19 日作为免费模型正式发布。

其配置参数为 30B-A3B 的 MoE 架构,支持 200K 的上下文长度。此外,新注册使用者还会额外获赠 2000 万 token 的永久额度,注意这里的期限是永久有效,而非通常的 30 天或者 90 天。

接口对接地址为 https://open.bigmodel.cn/api/paas/v4

近期智谱还推出了 Anthropic 兼容端点,这对于 Claude Code 使用者来说至关重要。只需修改两个环境变量即可无缝切换,完全不需要改动任何代码:

json

{

  "env": {

    "ANTHROPIC_BASE_URL": "https://open.bigmodel.cn/api/anthropic",

    "ANTHROPIC_AUTH_TOKEN": "你的 Key"

  }

}

其劣势同样显著:不仅并发数限制为 1,响应速度也偏慢。根据实测数据,面对“重构这个 repo”的相同指令,Claude Sonnet 4 仅需数秒即可返回结果,而 GLM-4.7-Flash 则需要耗费 15 到 25 秒。若同时运行多个 Agent worker,便会直接触发并发限制。

另外两个坑:

配置 model id 时请避开旧版本。glm-4.5-flash 早已下线,glm-4.5-air 则属于付费套餐,目前网络上大量教程仍在推荐这二者,盲目照搬必然导致报错。当前有效的版本应为 glm-4.7-flash。
建议不要通过 OpenRouter 调用智谱。OpenRouter 平台上 z-ai/glm-4.7-flash 的定价为每百万 token 输入 0.06 美元、输出 0.40 美元,这里面包含了中间商差价。若想免费使用,唯有直接连接 bigmodel.cn。
03PART火山引擎(豆包):唯一每天重置的VOLCANO · DOUBAO每日提供 200 万至 500 万 token,次日自动重置额度,完成实名认证再额外赠送 50 万。这种模式目前在业内独一无二。其他平台大多采用“一次性发放一笔额度,用尽或到期即失效”的策略,唯有豆包坚持每日为你补充。倘若你的需求是每日执行固定批次的任务——比如定时采集数据、批量内容改写、生成每日报表——那么豆包绝对比任何一次性大额套餐更为契合。至于模型性能就不多夸了,应对日常需求绰绰有余,但别期待它处理繁重工作。
04PART硅基流动:仅凭一个 Key 可调用 100 多个模型SILICONFLOW新用户注册可获 2000 万 token,9B 以下的小模型永久免费。API接入地址https://api.siliconflow.cn/v1/chat/completions它其实是一个模型聚合平台,Qwen、DeepSeek、GLM 等都能通过同一个 Key 进行调用。当你需要横向对比不同模型的表现时,用这个平台最省心,无需逐一去各家注册账号。需要注意的是,额度的扣减速度与所选模型相关。有实测反馈称默认额度绑定于 Qwen3-30B-A3B,若切换至 DeepSeek-V3,额度消耗会明显加快,这是因为两者的计费系数不同。不过该说法目前仅有单一信息源,未能进行交叉验证,建议大家在自身使用时多留意用量页面的数据。
05PART阿里云百炼:赠送额度充沛,但存在过期限制ALIYUN · DASHSCOPE

各个模型均赠送 100 万 token,Qwen 全系列与 DeepSeek 的额度可叠加使用,累计总量可达千万级别。

接口地址https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions

痛点在于其仅有 90 天的有效期限,一旦逾期便全部作废。

适合近期有明确且大批量任务需求的用户使用,不宜长期囤积。
06PART百度千帆和讯飞:永久免费,但速度方面都略显迟缓QIANFAN · SPARK百度千帆旗下的 ERNIE-Speed 提供永久免费服务,QPS 高达 50,这种并发量在免费层级中可以说是相当大方了。讯飞星火推出的 Spark Lite 同样支持永久免费,官方宣称不限制 token 数量。然而其 QPS 仅仅只有 2——每两秒才能处理一个请求,用来进行日常对话还算凑合,要是执行批量任务真能把人急坏。因此,这两家更适合作为备用选项,核心的主力工作可不能指望它们。
07PART海外三款服务,简要概述OVERSEAS国内直连的稳定性欠佳,这里仅列出相关数据,不作赘述。Google AI StudioGemini Flash 系列提供免费服务,大约支持 1500 请求/天、15 RPM、1M token/分钟,上下文长度最高达 1M。配额给得相当充裕,至于能否连通则取决于你自身的网络状况。Groq1000 请求/天、30 RPM,Llama 4 Scout 提供 500K token/天。得益于 LPU 硬件,其速度在所有免费层级中拔得头筹。OpenRouter若不进行充值,每日限 50 请求,若一次性购买 10 美元额度,则可提升至 1000 请求/天。平台上挂载了十余个带 free 后缀的模型,适合进行对比体验,并不适合用于实际生产。顺带提个坏消息:Cerebras 自 2026 年 8 月起废止了免卡免费档,新注册账号改为发放 5 美元试用金,且有效期为 30 天,同时必须绑定支付方式。它曾是免费额度最为慷慨的一家,如今这福利已不复存在。
08PART三个提醒PITFALLS 优先使用有期限的额度,然后再用永久的。像百炼那种有效期为 90 天的应当先消耗掉,智谱的永久额度不妨留作底牌。如果顺序弄反了就会造成无谓的浪费。切勿使用网络上的共享 Key 论坛和社群中散布的免费共享 Key,不是随时可能失效,就是存在数据安全隐患。自己花几分钟注册一下就能搞定。免费 API 或许会用你的数据来训练模型 Google AI Studio 在某些地区的条款中已明确指出数据将用于训练。凡是涉及个人隐私、账号密码及业务数据的,千万别放进免费档里,在本地运行开源模型更为稳妥。
///LAST写在最后ADVICE如果只能挑一个长久使用,我会选择智谱 GLM-4.7-Flash。这并非由于它的配额最高,而是因为它不设使用期限——过去两年我目睹了太多“限时免费”最终演变成“免费用完即收费”的案例,智谱这样将 0 元清清楚楚标注在价格页面的做法,反倒令人觉得最踏实。豆包每天重置的额度适合作为备选方案搭配使用。若需短期内大量执行 Agent 任务或进行代码重构,建议先领取百炼那批千万级别的额度并尽快消耗,以免放置 90 天后失效。最后讲点掏心窝子的:免费配额的初衷是供你试错与练习,并非让你将生产环境搭建于此。一旦流量上去了,选个价格实惠的付费方案反而更省事——毕竟凌晨三点服务崩溃时,你绝不会庆幸当初省下了那区区几十块钱。合规方面照旧提醒一句,使用各厂 API 前务必查阅服务条款及数据使用政策;对开源模型进行二次开发前也要仔细阅读 LICENSE,切勿凭空臆断。
大家好,我是胖猫,一名热衷于逐个安装体验工具并记录实测心得的站长。—— 胖猫小栈 · 甄选并分享实用利器 ——精选往期:任推邦零成本日入400+!关于这个夸克网盘拉新项目,建议大家千万别错过隐退3年的下载利器Motrix再度回归,携54K star重写代码,甚至掌握了让AI分配任务的能力只需丢给OpenClip一个链接,3小时的直播内容就能自动剪辑出高光片段闲鱼捡漏神器[ai-goofish-monitor],24h AI 全天候替你扫货
点赞、在看、转发,THANKS FOR READING

Source: https://mp.weixin.qq.com/s/MsrAzTChwgk5xaComJvaBg For copyright issues, please contact: 13951431913#139.com or submit a comment on the right.

Free API Benchmark (September 2026): Is the free quota only for one-time use? In fact, 5 platforms still offer free usage, but only one boldly marks it on the pricing page. Includes Zhipu, Volcano Engine, Silicon Flow, Alibaba Pailian, Baidu Qianfan, and iFLYTEK. September 2026 free API benchmark horizontal comparison and quota analysis.

📦 8 parts + conclusion 👉 Swipe to view PART 01 Conclusion first: No absolute best, only the best fit PART 02 Zhipu GLM-4.7-Flash: The only one committed to permanent PART 03 Volcano Engine Doubao: The only one that resets daily PART 04 Silicon Flow: One key to call 100+ models PART 05 Alibaba Pailian: Quota is sufficient, but there is an expiration limit PART 06 Baidu Qianfan · iFLYTEK: Claimed permanent, but response speed decreases PART 07 Brief overview of three overseas platforms PART 08 Three tips and suggestions, recommend reading before operating PART /// Summary: My personal advice

Reason for this horizontal benchmark: Half a year ago, the answer was easily available, but now it no longer applies

Last week, a friend asked me which free large model API is currently the most practical. I was about to reply when I realized I couldn't give a definite answer. Half a year ago, I could immediately name a model and its free quota, but when I tried to call it according to the tutorial last month, the system said the model didn't exist — it turned out the batch of models had been offline for a long time. So recently, I re-registered with all the platforms still usable in China and carefully compared their official pricing pages and documentation. The data listed below is the result of checking on September 1st. Since the free quota often changes, it is recommended to recheck the page before actual operation.

2026 September Free LLM API Benchmark: Quota Comparison Across 5 Providers

01

PART

Conclusion first

CONCLUSION · See the result first

There is no so-called "best" option, only the one that "best fits your current needs".

What you want to do

Choose this

Reason

Long-term free use, don't worry about renewal

Zhipu's GLM-4.7-Flash

Its official price is 0 yuan, with no time limit or new user restriction

Run batch scripts daily

Volcano Engine (Doubao)

Daily quota of 2 to 5 million tokens, auto-reset the next day

One key to experience multiple models

Silicon Flow

Aggregates over 100 models

Run batch tasks, need concurrency

Baidu Qianfan's ERNIE-Speed

QPS up to 50

Long documents, ultra-long context

Zhipu (200K) or Gemini (1M)

Depends on your connection

To be safe, create a Zhipu key first, it's the only one I'm sure will likely still work next year

02PARTZhipu GLM-4.7-Flash: The only one committed to permanent freeZHIPU · GLM-4.7-FLASHBased on the investigation, this is the most recommended option, simply because its official price page clearly states 0 yuan for both input and output, with no "limited time" or "new user" labels. The update log also confirms that GLM-4.7-Flash was released as a free model on January 19, 2026.The model uses a 30B-A3B MoE architecture with a context length of 200K. Additionally, new users are given a 2000 million token permanent bonus, truly "permanent", not 30 or 90 days.The API endpoint is https://open.bigmodel.cn/api/paas/v4. Zhipu has recently added support for the Anthropic protocol, which is important for Claude Code users. Switching requires only changing two environment variables, with no need to rewrite any code.{"env": {"ANTHROPIC_BASE_URL": "https://open.bigmodel.cn/api/anthropic", "ANTHROPIC_AUTH_TOKEN": "your key"}}The downside is that the concurrency limit is 1 and the response speed is slow. Tests show that while Claude Sonnet 4 completes "refactor this repo" in seconds, GLM-4.7-Flash takes 15 to 25 seconds, and running multiple agents simultaneously will trigger errors.Two important notes: avoid using outdated model IDs like glm-4.5-flash (discontinued) or glm-4.5-air (paid); the current correct model is glm-4.7-flash. Also, do not use the free quota via OpenRouter, which charges for tokens; direct connection to bigmodel.cn is required for free usage.

03PARTVolcano Engine (Doubao): The only one that resets dailyVOLCANO · DOUBAODoubao provides 2 to 5 million tokens per day, with automatic reset the next morning, and an additional 500K for verified users. This daily reset model is unique among the industry's one-time quota systems. If your workflow involves daily tasks like data scraping or report generation, Doubao is likely the best fit. The model quality is sufficient for regular use, though not expected to be top-tier.

04PARTSilicon Flow: One key to call 100+ modelsSILICONFLOWNew users get 2000 million tokens, with models under 9B being permanently free. The API URL is https://api.siliconflow.cn/v1/chat/completions. This is a model aggregation platform supporting Qwen, DeepSeek, and GLM with a single key, making it convenient for comparing model performance. Note that the deduction rate depends on the chosen model — switching from the default Qwen3 to DeepSeek-V3 may consume tokens faster due to different rate coefficients.

05PARTAlibaba Pailian: Ample quota, but with expirationALIBABA · DASHSCOPEEach model offers 100 million tokens, with Qwen and DeepSeek quotas being cumulative (up to tens of millions). The API is https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions. However, the significant drawback is the 90-day validity period — after which all unused tokens are forfeited. This is suitable for immediate, large-scale tasks, not long-term hoarding.

06PARTBaidu Qianfan and iFLYTEK: Permanent free, but slow responseQIANFAN · SPARKBaidu Qianfan's ERNIE-Speed offers a permanent free plan with a high QPS of 50, which is generous for a free tier. iFLYTEK's Spark Lite is also permanently free and claims unlimited tokens, but its QPS is only 2 — usable for chat but too slow for batch processing. Both are best viewed as backup options rather than primary tools.

07PARTOverseas services: Brief overviewOVERSEASDomestic users may find overseas connections unreliable, so this is just a reference. Google AI Studio provides free Gemini Flash with ~1500 requests/day, 15 RPM, and 1M token/minute window; context length up to 1M. Groq offers 1000 requests/day, 30 RPM, and 500K tokens for Llama 4 Scout. OpenRouter's free tier is limited to 50 requests/day unless payment is made — it hosts >10 free models for comparison but not production use. Note: Cerebras' free tier has been discontinued since Aug 2026; new accounts receive $5 credit valid for 30 days with mandatory payment binding.

08PARTThree tipsPITFALLS1. Use expiring quotas first, then permanent ones. For example, consume the 90-day Pailian quotas before relying on Zhipu's permanent offer, to avoid wasting limited-time resources.2. Do not use shared API keys from forums. They are unstable, often expired, and pose security risks — always register your own account.3. Be aware that free APIs may use your data for training (e.g., Google AI Studio in some regions). For sensitive data, consider local inference with open-source models instead.

///LASTFinal thoughtsADVICEIf I had to pick one service for long-term use, it would be Zhipu GLM-4.7-Flash. Not because of its quota, but because its free status is clearly and permanently stated in the price list — unlike others that label themselves "limited" yet still charge. This transparency builds confidence. Doubao is a good secondary choice for daily routines, while Pailian is ideal for short-term, high-volume tasks before expiration.Ultimately, free quotas exist for testing and learning — not for production-grade reliability. When your load increases, upgrading to a paid plan is more cost-effective than dealing with downtime or overage charges. Also, always read the service terms and privacy policies before using any API to ensure compliance and data protection.

Hello everyone, I'm Fat Cat, a station master passionate about installing and testing tools, sharing my real experiences.—— Fat Cat's Den · Select and share useful gadgets ——Selected previous articles:Renmaibang earns 400+ per day at zero cost! Don't miss this Quark drive referral project.The download artifact Motrix, absent for 3 years, returns with 54K stars and now lets AI assign tasks.Just give OpenClip a link, and it automatically clips the highlights from 3 hours of live stream.The second-hand market bargain hunter [ai-goofish-monitor], 24/7 AI scanning for you.

Like, follow, share, THANKS FOR READING

}

}

Its disadvantages are equally pronounced: not only is the concurrency limit restricted to 1, but the response speed is also relatively slow. According to experimental data, for the same command to "refactor this repo," Claude Sonnet 4 takes only a few seconds to return a result, whereas GLM-4.7-Flash requires 15 to 25 seconds. If multiple Agent workers are run simultaneously, the concurrency limit will be triggered immediately.

另外两个坑:

When configuring the model ID, avoid using old versions. glm-4.5-flash has already been deprecated, and glm-4.5-air belongs to a paid subscription plan. Many tutorials on the internet still recommend these two, but blindly following them will inevitably lead to errors. The current effective version should be glm-4.7-flash.

It is recommended not to call Zhipu through OpenRouter. On the OpenRouter platform, the pricing for z-ai/glm-4.7-flash is $0.06 per million input tokens and $0.40 for output, which includes a middleman markup. If you want to use it for free, you must connect directly to bigmodel.cn.

03PART Volcano Engine (Doubao): The only service with daily replenishment VOLCANO · DOUBAO provides 200,000 to 500,000 tokens per day, with the quota automatically resetting the next day. After completing real-name verification, an additional 500,000 tokens are awarded. This model is currently unique in the industry. While other platforms typically offer a one-time allowance that expires once used or after a certain duration, Doubao consistently replenishes your balance every day. This makes it an excellent fit for users with fixed, daily task requirements—such as regular data collection, batch content rewriting, or generating daily reports—as it is more reliable than any single-use large-quota plan. As for model performance, while not cutting-edge, it is sufficient for everyday needs; just don’t expect it to handle heavy workloads.

04PART SiliconFlow: A Single Key for Access to Over 100 Models SILICONFLOW offers 20 million tokens to new users, with models below 9B parameters being permanently free. The API access URL is https://api.siliconflow.cn/v1/chat/completions. This is essentially a model aggregation platform where you can use a single key to call Qwen, DeepSeek, GLM, and others. This convenience is invaluable when you need to compare the performance of different models, eliminating the need to register for multiple accounts. However, note that the rate of quota deduction depends on the selected model. Reports indicate that while the default quota is tied to Qwen3-30B-A3B, using DeepSeek-V3 may result in faster depletion due to differences in billing coefficients. Although this specific claim hasn’t been cross-verified, it’s advisable to monitor your usage closely.

05PART Alibaba Cloud DashScope: Ample Quota but with Expiration Limits ALIYUN · DASHSCOPE

Each model is granted 1 million tokens, and the quotas for the Qwen series and DeepSeek are stackable, accumulating to well over ten million tokens in total.

接口地址https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions

The drawback is that this quota has a very short validity period of only 90 days; anything unused after this time is forfeited.

This makes it suitable for users with immediate, high-volume task needs in the near future, rather than for long-term hoarding.

06PART Baidu Qianfan and iFlytek: Permanently Free, but Slow QIANFAN · SPARK Baidu Qianfan’s ERNIE-Speed offers a permanent free service with a QPS of 50, which is quite generous for a free tier. iFlytek’s Spark Lite is also permanently free, with claims of no token limit. However, its QPS is only 2—meaning you can process about one request every two seconds. While acceptable for casual conversation, it is far too slow for batch processing tasks. Therefore, these two are best viewed as backup options rather than primary workhorses.

07PART Three Overseas Services: A Brief Overview OVERSEAS Direct connectivity from China tends to be unstable, so here is basic information without further analysis. Google AI Studio The Gemini Flash series offers free services with approximately 1500 requests/day, 15 RPM, and 1M tokens/minute, supporting a context length of up to 1M. The quota is substantial, though accessibility depends on your network setup. Groq Offers 1000 requests/day and 30 RPM, with Llama 4 Scout providing 500K tokens/day. Thanks to LPU hardware, this is the fastest free service available. OpenRouter Without pre-payment, you are limited to 50 requests/day. Purchasing a $10 credit increases this to 1000 requests/day. The platform hosts several models with “free” in their names, making it a good option for comparison testing but unsuitable for serious production work. Additionally, Cerebras has discontinued its free-tier model and now offers a $5 trial credit for new accounts with a 30-day expiry, requiring a payment method for verification.

08PART Three Important Reminders PITFALLS Use time-limited quotas first, then permanent ones. Consume the 90-day quotas (like DashScope) before relying on permanent tokens (like Zhipu), to avoid wasting perishable credits. Never use shared keys from forums. Publicly shared keys are prone to revocation and pose security risks. It is safer to spend a few minutes setting up your own account. Be cautious of data usage in free tiers. Google AI Studio, for example, explicitly states that data may be used for training. Do not input sensitive information (such as passwords or confidential business data) into free services; using locally hosted open-source models is a safer alternative.

///LAST Final Thoughts ADVICE If I had to choose one service to use long-term, I would pick Zhipu GLM-4.7-Flash. This decision is not driven by its highest quota, but by its lack of an expiration date. Over the past two years, I have seen too many “limited-time free” offers turn into “free until you run out, then you pay.” Zhipu’s clear labeling of “0 RMB” on its price page feels much more reliable. Doubao, with its daily reset, is a solid fallback. For immediate, large-scale tasks—like heavy Agent work or code refactoring runs—I recommend claiming DashScope’s accumulated millions of tokens and using them promptly to avoid the 90-day expiry. Lastly, a reminder: free quotas are meant for experimentation, not for building a sustainable production pipeline. Once your volume increases, a reasonably priced paid plan is often more cost-effective than waking up to a “service down” message at 3 AM because your free quota has been exhausted. Remember to review the terms of service and data policies for each provider, especially regarding local regulations, before proceeding.

Hi everyone, I'm Fat Cat, a webmaster who enjoys testing and recording the experience of installing various tools. — Fat Cat's Workshop · Curating and sharing practical gadgets — Selected from previous editions: RenTuibang earns over 400 per day at zero cost! Regarding this KuaKe cloud storage user acquisition project, I highly recommend you don't miss it; the download powerhouse Motrix, absent for 3 years, has returned with rewritten code and 54K stars, even mastering AI task assignment. Simply throw an OpenClip link and it will automatically edit the highlights from 3 hours of live content. The idle fish bargain-hunting artifact [ai-goofish-monitor] provides 24/7 AI scouting for you.

点赞、在看、转发,THANKS FOR READING

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry