Complaints that English-speaking developers posted publicly around Nov 7–9, 2026, paraphrased below with links to the original posts. It was a relatively light weekend for mass API outages — status.claude.com recorded a short, resolved Claude Fable 5.1 error window on Nov 8 (08:18–08:32 UTC) and no incidents on Nov 9. The louder pain was surprise cloud-agent burn, split quota pools, a benchmarked model you still cannot call, silent quality shifts behind the same model ID, default API watermarks, overflow credits that stay gated, and expensive video decision loops. 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. Cloud agents reroute the job — and a second wallet shows up.

What developers are complaining about: Work that used to stay inside a coding bot is being handed to a Cursor Cloud Agent without a clear heads-up, and the meter jumps. One tester said the same kind of edit went from about 1–2% of usage to about 20% after the reroute (@maz3839, follow-up). Another watched a cloud-agent run burn 55% to 100% of the weekly pool on Saturday in about 1.5 hours, then add roughly $30 of extra usage (@Robokoch56). A single cloud-agent PR ate 52% of a weekly Premium limit, versus about 5% of a Claude five-hour window for similar work (@ThatMorneGuy). On Cursor Ultra, the usage page can show almost nothing used (0.9%) while the agent still refuses with “buy credits or switch to a Grok model” (@PodobnikLovro, @BeebsMemes). Others say unsupervised bots nuked a month of usage (@washedkline), that a $200 Ultra weekly pool moved from 28% to near 60% overnight (@GrokBotRadar), and that a local Claude session on an existing subscription was later blocked when automation insisted on the cloud path (@LostNDtiredmeow).

How PoYo answers: PoYo does not sell a shared weekly subscription pool that can silently hand work to another runtime. Chat and media share prepaid credits from one wallet; empty balance returns 402, not a surprise extra-usage invoice (homepage, error codes).

Practical controls that match what those posts want:

• Set hourly / daily / total caps on the API key page so an overnight agent cannot keep spending past the ceiling.
• Poll GET /api/user/balance and stop on a hard 402 insufficient_credits_error (user balance, error codes).
• For multi-tenant or multi-bot setups, send X-End-User-ID and block one caller without rotating the whole key (end-user identification).
• Pass an explicit model ID and read live pricing at https://poyo.ai/models/<id>/llms.txt rather than inheriting a hidden default (llms.txt, API overview).

Public docs do not claim Cursor / Grok Bot cloud-agent routing controls, a separate “extra usage” wallet beside prepaid credits, or an automatic spend firewall for unsupervised agents. Build the hard stop in your client.

Q2. Gemini 4 Argon burns its own weekly pool.

What developers are complaining about: Paying for Gemini access is not the same as being able to spend it on Gemini 4 Argon. One Antigravity user asked for a hotfix because even the High reasoning tier was burning the restrictive “Gemini 4 Argon Weekly” pool instead of the primary “Gemini Models” 5-hour / weekly allowance (@Qazcto7). Others asked whether Gemini and Argon quotas had reset (@JamesMalsawm), said a paid AI Pro plan still left them stuck on a Gemini 4 quota wall (@realJackHalder), or hit a raw “You exceeded your current quota… limit: 20000000, model: gemini-4-argon” error (@nayeemrahman_).

How PoYo answers: PoYo does not split one chat model into a named weekly pool and a separate primary pool. You pick a documented ID and draw from one prepaid wallet (homepage, llms.txt).

If you want Gemini 4-class chat on PoYo today, the catalog ID is gemini-4-argon (Gemini 4 Argon). Live rates in that card’s schema are 320 credits ($1.60) input and 1600 credits ($8.00) output per 1M tokens, published as 20% below official — that is per-card, not a sitewide discount (gemini-4-argon/llms.txt). Call it on Gemini-native /v1beta/models/{model}:generateContent or the documented Chat Completions path (chat overview). Check GET /api/user/balance before large batches (user balance); empty wallet is 402, not a “wait for the weekly Argon bar” state (error codes).

Public docs do not describe Antigravity quota-pool routing, a dual “Gemini Models vs Gemini 4 Argon Weekly” meter, or automatic weekly resets that match Google’s IDE.

Q3. The model tops the leaderboard — and you still cannot call it.

What developers are complaining about: Gemini 4 Carbon was announced late October and sits at the top of public intelligence indexes, but builders still cannot get an endpoint. @Yuchenj_UW asked the blunt question: Free, Pro, Ultra, and API all say not available. Others repeated the same gap — a price or rank is not an API ID (@411apt, @klaivor, @jaymos, @onefinalprompt, @austindixson_, @NeaByteLab).

How PoYo answers: Treat the live catalog as source of truth, not a launch tweet. Today’s index lists 147 models and includes gemini-4-argon; it does not list a Gemini 4 Carbon ID (llms.txt, chat overview). An unknown chat model ID returns a documented 404 model_not_found (error codes).

If you need a Gemini-class chat card that is published today, start with Gemini 4 Argon and re-read https://poyo.ai/models/<id>/llms.txt before you lock a unit-cost model. Other documented Google chat IDs sit on the same chat overview.

Public docs do not promise that an announced third-party model will appear on PoYo, publish a Carbon launch date, or automatically substitute a nearby Gemini ID when you request one that is not in the catalog.

Q4. Same model ID, different brain.

What developers are complaining about: The name on the request did not change, and that is exactly what people distrust. Weekend posts circulated Shane Melaugh’s review of about 1,400 Claude Code sessions: Opus 5.5 reasoning is said to have flipped on Nov 3, with no new version string and no changelog. One summary called it “same model ID, different brain” and “silent downgrades are the worst updates” (@ThorTheUndying, @AGTPinsights, @dhruval_ramani). The alleged shift date is Nov 3 — before this weekend’s window — but the API-name complaint is what builders kept posting on Nov 8–9. Anthropic has not publicly confirmed the analysis in those threads.

How PoYo answers: Always send an explicit model ID and read that card’s live schema at https://poyo.ai/models/<id>/llms.txt (llms.txt). For Claude Opus 5.5, the documented ID is claude-opus-5-5; live rates are 640 credits ($3.20) input and 3200 credits ($16.00) output per 1M tokens, published as 20% below official standard on that card only (claude-opus-5-5/llms.txt). Use OpenAI-compatible /v1/chat/completions or Claude Messages /v1/messages as marked for that ID (chat overview, API overview). Unknown IDs return 404 model_not_found, not a silent remap (error codes).

Getting Started already sets a realistic stability expectation versus official providers. Public docs do not claim cryptographic attestation that a response was produced by an unchanged upstream checkpoint, publish a changelog when a lab updates a named model in place, or offer automatic substitution if quality drifts.

Q5. Grok API text is watermarked by default.

What developers are complaining about: Since Nov 7, builders say xAI turned on an invisible statistical watermark for Grok API text by default — including, in some write-ups, JSON and tool-call arguments — across several Grok 4.x models. Opt-out is described as a team-admin action tied to a legal exception (EU AI Act), and the signal to look for is an x-watermark header. The worry is not only the mark itself, but that production agents may already be shipping watermarked email, posts, and reports without the app author noticing (@thecsguy, @Osu_most, @Drerki, @getloadoutai).

How PoYo answers: PoYo lists Grok chat cards you can call with an explicit ID — grok-4.7 and grok-4.6 (chat overview, llms.txt). For Grok 4.7, live rates are 320 credits ($1.60) input and 960 credits ($4.80) output per 1M tokens, published as 20% below official standard on that card (grok-4-7/llms.txt). The endpoint documented there is OpenAI-compatible /v1/chat/completions.

That same llms.txt — and the rest of the public docs we checked today — do not mention watermarks, an x-watermark header, or an opt-out switch. If watermark behavior matters for your product, verify it against upstream xAI docs and your own response headers; do not assume PoYo strips, preserves, or attests the mark.

If you need a different documented lab instead, pick another published chat ID (for example Claude Opus 5.5, Claude Sonnet 5.5, or Gemini 4 Argon). Public docs do not claim cryptographic attestation of the upstream lab, watermark removal, or a default-off watermark policy.

Q6. Extra credits you already bought still will not run.

What developers are complaining about: Paying for overflow capacity does not always mean you can spend it when the wall hits. One customer said they bought thousands of OpenAI credits and were told they cannot use them when they need them, only when they do not — support replies about Astra usage and purchased credits did not clear the block (@peterspear1, follow-up). A Codex user was down to 25% of the weekly limit by Sunday and planned to spend the API credits bundled with the coding agent to finish the week (@SteveGalvisX). Another reported Codex timeouts with credits still on the account after the banked reset was used up (@VibeMediaStudio). Separately, a weekend explainer reminded people that an OpenAI 429 can be credit_balance_exhausted rather than RPM — retries do nothing until the balance moves (@bufcapy).

How PoYo answers: Prepaid credits live in one wallet. There is no subscription window that later decides whether those credits are “allowed” today, and credits never expire (homepage).

Generation defaults: up to 20 new /api/generate/submit requests per 10 seconds per account; overage returns 429 and does not enter the queue. LLM endpoints are currently not rate limited (Getting Started, chat overview). Inspect GET /api/user/balance for a real-time credits_amount (user balance). Cap keys on the API key page.

Docs do not describe ChatGPT / Codex weekly usage walls that block already-purchased overflow credits, or OpenAI-style credit_balance_exhausted inside a 429 body — on PoYo, empty balance is 402.

Q7. Video decision loops still pay keeper prices.

What developers are complaining about: The expensive part of AI video is often not the final clip. One builder argued that Kling 4.0 should not be the model deciding framing, and that a 1-credit draft should carry those clicks — most of an hour-long session is decision-making, not footage (@francescoinweb3). Another warning: some catalogs price video in tokens, not seconds, so “the number on the page rarely predicts the number on the bill” (@web3l0var).

How PoYo answers: Video uses the unified async path: POST /api/generate/submit → status poll or webhook (video overview, webhooks). Failed generation tasks are not charged (homepage, llms.txt).

Live per-second prices for the IDs people were arguing about (re-check the file before you budget):

• Kling 4.0 — kling-4.0: 28.5 credits ($0.142)/s at 720p with no video input, 42.5 credits ($0.212)/s at 1080p (kling-4-0/llms.txt)
• Kling 4.0 Flash — kling-4.0-flash: 18.8 credits ($0.094)/s at 720p with no video input (kling-4-0-flash/llms.txt)
• Other draft-friendly catalog entries include Seedance 2.0 Mini, Wan, and Veo family IDs (llms.txt, video overview)

Ops tips for decision-loop waste:

• Use a cheaper documented ID for framing tests, then an explicit keeper ID for the final submit.
• Prefer webhooks over aggressive polling.
• Do not blindly resubmit: a retry can create another billable task if the first already accepted (Getting Started).
• Download finished files within 3 days retention (Getting Started).
• Cap spend with key limits and stop on 402 (API key page, error codes).

Docs do not promise a “1-credit draft” SKU, a dislike / regenerate refund for finished-but-undesired clips, or automatic multi-lab video failover — only that failed tasks are not charged. PoYo’s published Kling prices are per second, not per token; do not infer a token bill from another catalog.

Quick map

Browse models and live prices on poyo.ai, schemas via llms.txt, and the full docs index at docs.poyo.ai/llms.txt.