[Issue Report] Severe slowdowns and frequent 503 errors on Gemini Flash 3.8 during specific hours: Are server resources being intentionally throttled?

Every day when I use Antigravity for work between 16:00 and 22:00 UTC, Gemini Flash 3.8 constantly goes on strike.

Even when it’s not down, my tests show that during this timeframe, Flash’s speed isn’t even 1/10th of what it usually is. A 2-minute task takes over 20 minutes and still isn’t done—it’s absolutely infuriating!

Google Cloud model server 503 capacity exhausted (throwing errors multiple times):

Yesterday at noon and early this morning, the API frequently returned: UNAVAILABLE (code 503): No capacity available for model gemini-3.8-flash-high on the server

Along with network connections being cut off remotely: write tcp ... wsasend: An existing connection was forcibly closed by the remote host dial tcp: lookup daily-cloudcode-pa.googleapis.com: no such host

This happens about 90% of the time.

Right now, as I’m writing this post, Flash has been running for 30 minutes and hasn’t finished a single task. It’s painfully slow and lagging.

Wait a minute… the news says most AI developers have already migrated to ChatGPT, DeepSeek, and Grok, with the ChatGPT API being particularly cheap and highly rated. Given this, it makes no sense that a decrease in Gemini API’s overall user base would contradictorily result in service suspensions due to “overcapacity.” I can only think of one explanation: Google is intentionally throttling the gateway capacity and diverting the freed-up compute power to internally test Gemini 4

If that’s true, I absolutely support Google hustling to catch up with ChatGPT and Claude Code, but please don’t squeeze the gateway capacity this hard, okay? Thanks.

Honestly, Google could just drop Claude Code Opus 4.6. It’s outdated, and its performance and benchmark scores are bottom-tier—even Flash High beats it by a mile. If Google isn’t going to update to Opus 5.5, there’s really no need to keep paying Anthropic. It’s a waste of money; Gemini is already great on its own.

Completely agree. Besides, we have quota, we shouldn’t be getting 503 and 429 errors when we have quota, that defeats the point. If Google can’t handle the compute, then maybe they should stop adding AI into every product? I don’t need AI in my search engine, but that’s undoubtedly the one that 503’s or 429’s very often, even after just 3 web searches.

It makes no sense to the customer! I understand that maybe the server-side quota for that model on that particular end-point is being rate-limited, but the customer should not pay an extra price by causing queries to fail, and this is ESPECIALLY straight up BS when you have quota.

Semi-offtopic:
I hope my release Sunday of dyngov, a cybernetic dynamic effort and reasoning budget governor handled via hooks, is enjoyed by many. I have empirical proof of it running very lean, but here are some approximate numbers:

Information on the dynamic governor

dyngov preliminary results

Script used Token Delta (%)
Simple conversational loop about my day with no context. -90...-95
Simple code task -40...-60
Complex philosophical debate. -14...-60
Advanced code task +10...-30
Very complex code task +44... +4
Note:
All tests were performed on `Flash 3.6 Low` due to `3.7/3.8` having severe issues regarding overthinking. Pro on Medium (possible by running `dyngov --lock 0.5`) is also rather good. Keep in mind that these tests were performed multiple times with varying scripts to test... well, variety.
Important:
Every time you see a percentage that is higher than baseline, this means that effort went over `1.0`. You can set a ceiling of `1.0` with `dyngov --ceil 1.0` to make that impossible.

Watch my GitHub (look for Natrox Github or Raptor86 GitHub), I will announce the release on the forums too, but this way, you see it right away.

I really hope this bug is fixed soon.

Are you guys using Vertex or AI studio? Which SDK and API?

I have implemented all the advised jitter, hedge, fallback, backoff, etc. and still getting crazy error rates. Running out of ideas and started benchmarking competition to migrate…

Totally down today no respond in antigravity chat .