2026-10-21: 8 API complaints we heard today

Complaints that English-speaking developers posted publicly on X between Oct 19 and Oct 21, 2026, paraphrased below with links to the original posts. Each answer uses only what PoYo's public site and docs say today, and links to the page it comes from. Where the docs don't cover something, we say so.

---

Q1. A tool I trusted leaked my API key, and I woke up to a bill. How do I limit the damage next time?

What developers are complaining about: The OpenCode Console breach was the biggest API story of the past two days. OpenCode says provider keys saved by about 1.4% of Console users may have been exposed, and some developers reported unauthorized OpenAI, Anthropic, and Google charges, with one report of $1,700. Developers kept giving the same advice: give every tool its own key with a spending cap, and remember that deleting a key inside a third-party tool doesn't revoke it. Another dev posted about a $420 bill from a forgotten 2024 key.

How PoYo answers: Our docs list three controls you can put on a PoYo key:

• Per-key usage caps (hourly, daily, and total) and an IP whitelist, so only your approved server IPs can use the key. Both are on the API key page and explained in Getting Started.
• Per-end-user blocking. Your app sends an X-End-User-ID header with each request, and you can then block one abusive caller on a key without rotating or disabling the key. You can also make the header required for every request on that key. See End-user identification.
• Keys stay on the server. The docs say never to put keys in browser code, mobile apps, or public repos.

Our public docs don't promise refunds for usage made with a leaked key. If you think a PoYo key was misused, revoke it and contact support ([email protected]).

---

Q2. I prepaid for API credits, and a year later they were gone.

What developers are complaining about: One widely shared post argued that Anthropic and OpenAI prepaid credits expiring after 12 months has no technical reason behind it, because you're holding a dollar balance, not reserved GPU time. The author said he had personally lost about $50 to expiry.

How PoYo answers: The PoYo homepage says plainly: credit-based pricing, no subscriptions, one-time purchases, and credits never expire. You can read your current balance at any time from the user balance endpoint (GET /api/user/balance), which returns credits_amount in real time.

---

Q3. Gemini 3.8 Flash has been returning 503 "no capacity" for days. What do I do when one upstream stalls?

What developers are complaining about: One developer said the gemini-3.8-flash free tier had been returning 503 for 2.5 days. Another saw 60+ second replies or 503 capacity errors in Antigravity. On the coding-agent side, one analysis counted 42.7 hours of open Codex incidents in two weeks and argued that outages, not taste, are why many people pay for two coding tools.

How PoYo answers: Here is what the docs support, and what they don't:

• One key, many models. Gemini 3.8 Flash, Gemini 3.7 Flash, Claude, GPT, and DeepSeek models all run on the same PoYo key. If one model is struggling, you can switch to another by changing the model value. Every model is listed in llms.txt.
• Separate error codes. The error codes page distinguishes 502 upstream_error from 503 service_error ("server is busy") and recommends retrying both with exponential backoff.
• What we don't claim. The docs describe no automatic failover between providers, so if you want a fallback, build it into your client. Getting Started also says directly that PoYo's "overall stability may be slightly lower than official providers." We'd rather tell you that up front.

---

Q4. My 429 wasn't actually a rate limit. Why do providers reuse the same error for different things?

What developers are complaining about: One developer explained that a 429 they hit was really a tier cap on long context, not a rate limit, and that it took about an hour of digging in dashboards to figure that out. A related tutorial on Gemini quota errors (Oct 16) walks through a single 429 RESOURCE_EXHAUSTED that can mean "no quota for this model at all," "per-minute limit," or "per-day limit," and each needs a different fix.

How PoYo answers: Our error codes page gives each status a single meaning:

The same page tells you to keep an eye on your credit balance so you don't hit a 402 in the middle of critical work. The balance endpoint is the way to check it.

---

Q5. My batch jobs keep hitting rate limits halfway through.

What developers are complaining about: A developer building a lead-generation agent with n8n said API rate limits during batch runs were their biggest problem, and asked others how they handle backoff and deduplication.

How PoYo answers: Getting Started publishes the default limit: up to 20 new generation requests per 10 seconds, per account, which "typically allows 100+ concurrent running tasks." Requests over the limit get an HTTP 429 and do not enter the queue, so pace your submissions instead of firing everything at once. If you hit 429s consistently, the docs say you can ask support for a higher limit, though approvals are cautious.

Generation is asynchronous: you submit a task, get a task_id, and then poll or receive a callback (API overview). For large batches, the docs recommend signed webhooks (HMAC-SHA256) over frequent polling. On deduplication, the docs give one rule: treat every resubmission as potentially new paid work, and check whether the original request already created a task before you retry.

---

Q6. AI video APIs cost wildly different amounts, and I keep paying for clips I can't use.

What developers are complaining about: One price comparison found video generation prices varying by more than 3x for the same model across providers. Another developer said Higgsfield is still too expensive even with Seedance 2.5 and asked for cheaper options. A third said they spent about $12 of $40 in credits on an app's video generator without getting one usable 15-second clip.

How PoYo answers: All video models are collected on the AI Video API page, and they all use the same async submit-and-status flow. Each model publishes per-second prices. For example, Seedance 2.5 lists 480p at $0.140/s and 1080p at $0.570/s without video input, with lower rates when the request includes a reference video. The pricing page puts PoYo's price next to the official price wherever one is listed.

On failures: both llms.txt and the model pages say failed tasks are not charged. To be clear, a task that finishes but gives you a clip you don't like still counts as a successful output and is billed. You can test prompts in the Playground on each model page before calling the API.

---

Q7. Is the cheap API I'm using actually serving the model it claims?

What developers are complaining about: One developer said OpenRouter started sending their GLM requests to an FP4-quantized endpoint, which broke their project before they figured out what had changed. Another post described "100% official direct supply" API relays that sell at steep discounts but turn out slow, and how to spot a fake upstream.

How PoYo answers: Here is what you can check on our public pages:

• Labeled official variants. Some models are offered as clearly named "official" variants, such as Sora 2 Official and Veo 3.1 Official. They are listed separately from the other variants of the same model, as shown in llms.txt.
• Public schemas and prices. Every model publishes its current input schema and price in its own llms.txt.
• Task history. The task history page shows the model, input parameters, status, and result or error for every task you run (see Getting Started).

Our public docs don't describe how each request is routed upstream, so we won't make claims about routing here.

---

Q8. My agent's token bill is out of control.

What developers are complaining about: One developer compared a month of research-agent costs: $220 on Opus 5.5 versus $4.80 on DeepSeek. Another argued that if Gemini 4 Argon matches Opus for less money, "brand loyalty won't save your API bill."

How PoYo answers: All these models run on one key, so moving a workload to a cheaper model means changing the model field, not starting a new integration. Current listed prices:

• Claude Opus 5.5: $3.20 input / $16.00 output per 1M tokens, which the model page describes as 20% below official standard rates.
• Gemini 4 Argon: $1.60 / $8.00 per 1M tokens (20% off official).
• DeepSeek V4.1 Flash: $0.30 / $1.20 per 1M tokens.

Check the pricing page before committing a workload. Prices change, and each model's llms.txt always has the current numbers.

---

Quick map