With the release of Qwen 3.8 Max and it's insane 98% discount during 'off-peak' hours and 90% discount during regular hours, I decided to download Qoder to give it a go.
For reference, Qoder is essentially Antigravity, but by Alibaba and it blew my mind. Whether it's the fact that they've been throwing freebies at me since I downloaded it (100 usd worth of free SOTA 'ultimate' calls, refreshing every 2 days sofar till the 30th), or the 300 credit 'pro-trial', which goes a LOOOOONG way when you use Qwen 3.8 Max at the 98% discount. Or the fact that Lite mode is free! Like actually free, I've pushed over 20k LOC edits through it today and it hasnt even hit a fair-use cap?
I've been trying out the different platforms lately to see what's the most cost-effective, as my day-job comes with Claude Code, which is great for that, but for any personal projects, that's on my budget, with my accounts. I've been using APIs from various providers to see which is cheaper, more capable, using freebies where possible (eg. Google Startup and Azure Startup free credits) and while they work great while free, reviewing the bill afterwards is why I push them so hard... For perspective, $500 of Azure freebie money for around 50k LOC editing via OpenAI API with the model-router (discounted and routes to the cheapest model that can do the task). $150 in Google freebie money for around 20 LOC editing via Gemini 3.6 flash on Low thinking. Vs 40k LOC on $10 via Qoder (using Qwen 3.8 discounted rate), or $100 for 20k LOC via Ultimate (the expensive one).
Previously, Chinese models were 'decent', but often fell short of their western counterparts, that changed this month... Qwen and Kimi are officially better than anything in their price range, infact they punch high enough that they're the best value for twice their price range at least.
But dont take my word for it, Download Qoder and give it a try yourself. Unlike the western platforms that hide what the actual cost is till you're slapped with a bill and even then they dont actually give you clean breakdowns, Qoder is shockingly open with their free trials, showing you a detailed breakdown of what each task would have cost you. In the modern day of 'gotcha' bills because 'it takes us a while to calculate how much it costs you', Alibaba shows that it's cloak and dagger tactics to nail you with bills you cant predict. Transparent pricing and transparent usages, live. There's no excuse anymore Google and Microsoft, give us live-stats on usages.
Sorry for the poor format, lack of polish, etc. I'm knee deep in work and dont have much time to read or write, but I still wanted to put word out. If you havent downloaded Qoder yet, seriously give it a try. If you like it, you like it, if you dont, it's completely free to try.
P.S. Give the Better Loops feature a try. It's essentially a routine they built that analyzes your codebase for flaws and gives you a breakdown on it with easy to fire scoped 'quests' to fix them. They also have a 'generate wikis' routine, that's actually pretty cool. They treat documenting the codebase as codebase context which is brilliant! Instead of wasting token generating a wiki for people to read, it's actively used to prevent drift.
Top comments (34)
Haven't touched Qoder but I've been living in Qwen the last couple weeks anyway — Qwen3-TTS for a voice cloning pipeline, Qwen Cloud for an invoice-reconciliation agent I built for a hackathon. Same story on pricing from the API side and the quality gap you're describing is real, not hype.
Might actually download Qoder just to see that live cost breakdown you mentioned — that's the part nobody else does.
Definitely give Qoder a try. Could be subjective, but I find that for long-horizon tasks, it does a much better job staying consistent. Eg. I told it to start pushing to git, that was 20 prompts ago and in that time it's written over 50k LOC on Lite, the tier that's supposed to be bad? I've also been hammering Lite, fully free and it's blowing my mind how I've never heard of this platform yet! Seriously, it's worth downloading for the Better Loop and Wiki Generator alone. Every other AI orchestrator platform has just been a copy-paste of eachother, Qoder is the first 1 I found that actually adds value from a platform side. Also it's fully free for BYOK (havent dug too much, but looks like just Chinese models supported), but still cool that it's available for free with what I'd describe as the most robust orchestration layer I've seen to date. The models dont drift and they've been exceptionally consistent with writing unit tests. It also picked up gaps that Sonnet, Opus, Gemini Pro, GTP SOL, all missed. That alone makes it worth giving a shot atleast.
Do note, all of this is my honest, unaffiliated opinion. I wholeheartedly recommend giving Qoder a try
50k LOC over 20 prompts on the free tier without drift is the number that'll actually get me to download it... Thanks
And the usage isnt hidden in some obscure subsystem. You click on your usage, you click view details, it takes you to this page.
And to just show you I'm not kidding about it actually being LIVE usage stats.
the screenshots land better than the free-tier claim did .
Quest Mode showing the same log format on both Lite and Ultimate means it's not a stripped-down free version, it's the same system with a different bill attached.
$5.76 for 38 minutes on Ultimate is the number I'd actually budget against, not the $0.00 rows. Is that Ultimate rate pre- or post-discount??
Ultimate isnt discounted at all. Those were free Ultimate calls they gave (200 to claim once off, 1000 every 2nd day sofar. Ultimate is priced at 1x credit rate, whereas different tiered models work at different % of credit rate. Eg. Qwen 3.8 Max is at 90% discount during peak hours, 98% off-peak. Translating to 0.1 and 0.01x credit rate respectively. But I get what you mean.
Here's some I did during off-peak. So they show real rate and the discounted rate. Qwen charges directly to credits, which I used via the pro-trial, so these are the real numbers you'd get billed, with vs without (striked out) the discount.
Pay attention to rate before vs after 4pm. Off-peak by me translates to 4pm-2am.
I read it wrongly. The 90/98 split you're describing is the actual mechanism and it's more interesting than what I had: the discount is tied to when you run the model, not which tier you bought.
that's less "this platform is cheap" and more "this is cheap if you can schedule around Alibaba's off-peak window."
Does Qoder let you queue a task to fire after 4pm automatically or is timing it on you manually???
Once you let it spec a task, it gives you the option to build or schedule. So yes, you can most definitely schedule it! It also tells you exactly when (in your local time) the off-peak discount starts.
Overall, the platform's perks are that they're transparent and they offer actual features, not just gimmicks. Better Loop searches for architectural flaws, not just tiny bugs. The Generate Wikis keeps your codebase documented, while keeping the model context tight and fresh. Both really good features that I honestly wish all AI platforms had.
That being said, their pricing isnt terrible either, though it depends what you use. Eg. Kimi K2.7 is 0.3x credit rate, vs Kimi K3 is 0.8x credit rate. K2.7 costs 1/3rd what K3 costs. I'd still need to reverse engineer a way to find out if that's a good thing for K3, or a bad thing for K.2.7, but given that GLM is 0.5x, I'm leaning towards good thing for K3, not so great thing for K2.7. Though with the new Qwen models, I'd honestly suggest just using K3 for UI design and rely on Qwen for the rest.
Also, looks like they just released an update. Better Loop got expanded into Better Harness. Essentially a full on codebase auditing tool, that finds any security, performance, reliability, etc. issues, instead of just optimization suggestions. So they're definitely hard at work on the platform...
the credit rate alone won't settle K3 vs K2.7 .you'd need the same task on both to see if that 2.7x gap is a real capability jump or just repricing. GLM at 0.5x is the better anchor: if K2.7 undercuts even that, it's priced to move, not priced to match.
Better Loop becoming Better Harness is the bigger tell though . going from "here's an optimization" to full security/performance/reliability auditing is a scope change, not a feature bump. Curious if it's still fast enough to run per-commit, or if that's now a "kick it off and check back" tool.
Better Harness is more a once-off audit system, rather than a continuous system. Though I imagine they'll likely add it as gating logic at some point, given that it's still in Beta.
Well... Guess they added security gating into the IDE now.
one more on the sub-agent thread . does Qoder let you define your own named agents (something like Claude Code's .claude/agents/ markdown files, scoped tools and system prompt per agent), or is Better Harness a single built-in auditor you can't decompose or customize??
Better Harness is a dedicated audit mode. They have the security gating which you can enable and they have the agents and experts systems which lets you either use a template one, or create your own custom one and unlike Google's Agents Platform, it's free. It has all the bells and whistles from the top Western agentic IDEs, without the bloated costs.
No way, you're just finding out too... Well, I guess it's because we have way too many tools like this over here — hard to keep track.🤣
Honestly, I'm loving it! I've been building V.E.L.O.C.I.T.Y. IDE and it didnt just help me build out the entire agentic browser (from scratch), but using their system made me realize just how 'stupid' the go-to agentic IDEs are. Seriously, the fact they built in codebase analysis with patch generating and an entire Wikis generator. It made me expand the scope of velocity to include those. Cuz a wikis tied to a merkle root site-map as agent context means it's a VC system that tracks state, not just code. The codebase analysis, while they use AI to do most of the analysis, you can see they have a pretty solid setup for pattern detection that isnt inherent to a model. They did a really good job with it and it blows my mind the more I use it. For perspective, Lite mode, I've now pushed close to 100k LOC through it, that's 10s of millions of tokens... For free, on an unpaid account, without ever hitting 'high traffic' or 'fair-usage' blocks. It just works and it works well... Writing Rust isnt easy and it's doing a phenomenal job. It's faster and smarter than Gemini 3.6 flash, that's something I can confidently say.
Maybe you could also give DevEco Code by Huawei a try — GLM 5.1 is free to use, up to 50 requests per minute. Not sure if there'd be any network restrictions on your end though.
Download started. Thanks for the headsup! Though it seems to lean heavily on HarmonyOS, is it OS restrictive in terms of what platform it can build for?
Nah, you're good. The CLI tool (npm install -g @deveco/deveco-code) works as a standalone AI coding assistant — GLM-5.1 is built-in, free, unlimited, 50 req/min. No HarmonyOS toolchain required to use the AI features.
The HarmonyOS lock-in is only for the full IDE (DevEco Studio) when you need to compile, debug, or run on actual devices. But if you just want to try GLM-5.1 for coding, the CLI alone does the job. Worth a shot since it's literally free 😏
And you've been gatekeeping this all this time why? That's incredible! I'm definitely gunna give it a run and see what it can do. Honestly Qoder was the turning point for me... The Cloudflare Workers AI hook I built for Kimi K2.7 was a good experiment, but it started me down the rabbit hole of exploring Eastern models more and IDEs too, usually it's a bit iffy, cuz of the language barrier and some models being trained almost explicitly on Chinese and it tends to through characters out mid session (surprisingly though, none more than Google's Gemini... Odd...). But Qoder and just the features they built in, like the Wikis and knowledge cards. I've been exploring it a bit and honestly, I can see why the East is winning the AI race... While Codex, Cursor, Antigravity, Claude, Copilot all fight for user-base, they do it by doing a few token refreshes, or giving you the 'privilege' of using their most expensive model on a subscription you pay for, at the rate it's charged at... Vs Qoder actually adds useful features AND freebies and the models supported are cheaper in general. I signed a sub today for Qoder, fact that the wikis and knowledge cards stay up to date is a game-changer, cuz that's persisted doesnt matter what model you use and it just works. But Imma give DevEco a try, GLM 5.1 for free sounds like it can be pretty useful.
Honestly, it's mostly an information gap. There's a certain level of hostility toward Chinese models internationally. The reason I never mentioned it before is that I assumed everyone already knew about this stuff — but it turns out only we know about it. Over here, there are just too many AI companies popping up everywhere — models, freebies, giveaways, everything is comfortable. AI services are basically becoming infrastructure here. Like the model I use the most right now — DeepSeek V4 Flash.
For comparison, DeepSeek V4 Flash costs $0.14 per 1M input tokens and $0.28 per 1M output tokens. Cache hits? $0.0028 per 1M tokens — basically free. Claude Opus 4.5 charges ~$15 input and ~$75 output per 1M tokens. That's roughly 100x more expensive.
And Flash supports 1M context, tool calling, and thinking mode. So yeah, the price is really nice — makes you wonder what the West is charging for 😏
It really is an information gap. Search top 10 Agentic IDEs, or Top 10 most efficient Agentic IDEs. You see 10 different western copy-paste solutions. Not a single 1 from the East. Admittedly it really does depend on where you operate, regions like the EU have data-compliance laws (specifically to isolate the market and boost local LLM usage and growth), eg. if you want to process financial documents, it cant leave the region. I did look into DeepSeek when I was building the IDE originally, cuz the cost effectiveness is just unmatched for raw API cost. Honestly, the western companies are charging a premium for 'SOTA' when reality is, they havent been a step ahead for a while... Kimi K2.7 Code already bridged the gap at 1T params, now with Qwen and Kimi's new models, that gap flipped to where the Eastern models are faster, smarter and open weights... But why support a chinese company, when you can push Meta's Llama, who wont even disclose what they used to train it with... Western companies are likely sitting at a 90%+ Profit margin, while the Eastern companies are sitting at a modest 20%, but the difference for developers is you dont break the bank getting a real amount of work done. I made the switch officially, I sub'd to Qoder yesterday to really test the platform out, see how expensive the Generate Wikis maintenance is, etc. It reset Lite usage cap, so it's cost me a total of 80c sofar between Qwen Max and Wikis. For comparison, You cant even generate a single file with Haiku for that...
Haha, if you ever come to China, KFC's on me anytime 😄
The thing that really peeves me off, is the fact that everyone wholeheartedly agree that Qwen's tiny models are astonishing and the go-to for any local LLM needs... Yet they only advertise HuggingFace for downloading it for local inference, they never mention Alibaba Cloud, or Qoder. It really feels like the information gap is intentional to gatekeep and prevent people from realizing that the Chinese models are cheaper, faster, smarter. Or even infrastructure wise. Look at Antigravity, the times I've used it, it's a buggy mess. You give a command, switch to a different task and the entire app freezes, need to close it and reopen. I havent had a single hiccup with Qoder, sure it'll get slow when you really hammer it, but it gets slow, not crippled and broken. Which means it's foundationally built better, with better failover and async guarding to make sure it recovers from errors, not crash. That's not something you can retroactively implement, that's a design choice from the start and it'll take Google with all their money over a year to achieve and they'd have to release it as a V3, instead of a patch...
🤣 Bet!
Alright, I'll share what I know. Here's a list of Chinese cloud AI platforms — whether you can access them depends on your network:
Alibaba Cloud — Qwen / Tongyi: tongyi.aliyun.com/
Qoder (the one you mentioned): qoder.cloud/
Baidu — ERNIE Bot: yiyan.baidu.com/
Tencent Cloud — Hunyuan: cloud.tencent.com/product/hunyuan
ByteDance Volcano Engine — Doubao: volcengine.com/product/doubao
Huawei Cloud — Pangu: huaweicloud.com/product/pangu.html
DeepSeek: platform.deepseek.com/
Moonshot AI — Kimi: kimi.moonshot.cn/
Zhipu AI — GLM / ChatGLM: open.bigmodel.cn/
01.AI — Yi: lingyiwanwu.com/
Baichuan AI: baichuan-ai.com/
iFlytek — Spark: xinghuo.xfyun.cn/
MiniMax: minimaxi.com/
SiliconFlow (unified API): cloud.siliconflow.cn/
Hope your network holds up. Take your time exploring the free tiers. I used to think the Western internet was all about "freedom" — guess it's not quite what they advertise, huh 😏
qoder.com for me, kimi.com for me, other than that, rest of links pass. Thank you very much!
Yeah and this from Africa. China has actually had quite alot to do my country, including being the providers of the only road that hasnt had a single pothole in the past 10 years 🤣 meanwhile my government puts speedbumps on a highway... (true story, search it up, Namibia B1 speedbumps)
I'll definitely look into all of these, thank you very much. Cuz I am honestly sick and tired of how expensive AI has gotten and how pathetic western models and IDEs are at completing a full task. Case in point, a manufacturing system, standard BOM and WO, I spend 3 hours troubleshooting why Claude Fable (the golden child apparently), decided to mess why my formulas. Cuz I run under the assumption I missed something, so I have to validate why it made the choice and whether I was wrong... I wasnt, it just falls apart at anything more complex than 'make me a cyberpunk landing page for my website'...
Here's an updated list — added the ones I missed, including international URLs that might be more accessible:
Already works for you:
• Qoder: qoder.cloud/ ✅
• Kimi (Moonshot AI): kimi.moonshot.cn/ ✅
Global-friendly / should work:
• DeepSeek: platform.deepseek.com/ (fully global)
• Alibaba Cloud Model Studio (Qwen): alibabacloud.com/product/model-studio
• Zhipu AI (GLM / ChatGLM): open.bigmodel.cn/
• MiniMax: minimaxi.com/
• 01.AI (Yi): lingyiwanwu.com/
May need a workaround (your network situation varies):
• Baidu ERNIE Bot: yiyan.baidu.com/
• Tencent Hunyuan: cloud.tencent.com/product/hunyuan
• ByteDance Doubao: volcengine.com/product/doubao
• Huawei Cloud Pangu: huaweicloud.com/intl/product/pangu...
• Baichuan AI: baichuan-ai.com/
• iFlytek Spark: xinghuo.xfyun.cn/
• SiliconFlow (unified API): cloud.siliconflow.cn/
Bonus — open models you can run locally (HuggingFace / Ollama):
• Qwen2.5 (72B)
• DeepSeek V4
• GLM-4-9B
• Yi-1.5
• MiniMax-01
Most of these have free tiers that'll give you more tokens than Claude's entire paid plan 😂 Happy exploring!
Thank you very much! I will definitely be exploring them all and testing them out!
The live cost breakdown is the part I’d trust more than the headline model ranking. For personal projects, the failure mode is usually not 'can it edit 20k LOC once'; it is whether the tool keeps a clean diff, remembers the constraint, and makes the next rollback obvious when it gets too confident.
It does an honestly exceptional job. To put in perspective, I threw it at Dwarven Stronghold on Lite mode. That's a 150k LOC Rust codebase, with deeply integrated simulation logic and over 700 Tests. It keeps it accurate and keeps implementations clean... Even on Lite.
And that's where it started at in Qoder. It's up to 968 tests and pushing closer to 230k LOC now. Still keeping the instructions clean that I gave it at step 1. Which is strict Zero-Allocation and flat arrays. I added inbetween for it too, like adding Git and saying it should commit at stable states and keeping documentation up to date. It's been flawless at it.
Ok, update, I just hit usage cap on Lite. Apparently, they have a monthly cap on it. But what's interesting, is that Lite finished the entire set of tasks first, committed and pushed... Then only did it enforce the usage cap. Which means even when you eventually (for me it was probably 100k+ LOC across the 4 projects concurrently) hit the Lite cap, they dont leave you dead in the water, it finishes first, then stops. Which other platform finishes up before it hits cap?
Here's an idea of what it did when cap hit... Those requirements were set during the first prompt, a few days ago. It still enforces it and 20 of this round of edits were done after I had hit the Lite usage cap on another agent. Instead of stopping dead in it's tracks, it kept going, finishing not only the active edit, but all the edits and still pushed to git before it enforced the cap. And that's just what Qoder offers for free, to everyone... Even without any billing details or anything added beyond a generic 'sign in with google'. I am thoroughly impressed.