Editorial illustration for Google's Gemini 3.5 Flash, pricier, adds 11 Omniscience points hallucinations 61%
Google's Gemini 3.5 Flash, pricier, adds 11 Omniscience...
Google's Gemini 3.5 Flash just slashed its hallucination rate by a staggering 31 points. Don't break out the champagne yet, though. At 61 percent, it’s still hallucinating more than twice as often as front-runners like MiMo-V2.5-Pro. And according to fresh data from The Decoder, there’s a brutal efficiency tax: getting this model to complete agentic work requires a marathon 49 interaction turns, burning through input tokens and erasing any benefit from its lower per-token price.
In practice, though, the math flips. Gemini 3.5 Flash burns through so many more tokens on agent-based tasks that total costs end up 75 percent higher than Gemini 3.1 Pro, according to Artificial Analysis.
So the GDPval-AA benchmark shows Flash can now nearly match GPT-5.4 on complex tasks. But that capability comes at a perverse cost: running it is pricier than using Google's own Gemini 3.1 Pro, all thanks to those 49 turns. And in the crucial coding arena, it's simply not in the race. What's left is a paradox—a model that finally works, just woefully inefficiently, arriving just as rivals like MiMo and Grok are proving you can have high accuracy without the operational headache.
Related Reading
Browse all Google news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks
From Archives Alibaba launches Qwen3.5‑LiveTranslate‑Flash: 60‑language translation in 2.8 s
From Archives I/O 2026 unveils Gemini Omni for universal creation, Gemini 3.5 Flash debut
From Archives Google to roll out AI agents this week, adopting OpenClaw features
Open Source
From Archives Google plans USD 180‑190 B capex—six‑fold rise—to power AI agents like Spark
Business & Startups
Common Questions Answered
How much did Google's Gemini 3.5 Flash reduce its hallucination rate?
Google's Gemini 3.5 Flash reduced its hallucination rate by 31 points. However, at 61 percent, it still hallucinates more than twice as often as leading competitors like MiMo-V2.5-Pro, indicating significant room for improvement.
Why is Gemini 3.5 Flash more expensive to operate despite its lower per-token price?
According to The Decoder's data, completing agentic work with Gemini 3.5 Flash requires 49 interaction turns, which burns through input tokens and eliminates any cost savings from its lower per-token pricing. This efficiency tax makes it ultimately pricier than using Google's own Gemini 3.1 Pro.
How does Gemini 3.5 Flash perform on complex tasks compared to GPT-5.4?
The GDPval-AA benchmark shows that Gemini 3.5 Flash can now nearly match GPT-5.4 on complex tasks. However, this capability comes at a significant operational cost due to the high number of interaction turns required.
What is Gemini 3.5 Flash's main weakness in the coding arena?
Gemini 3.5 Flash is not competitive in the crucial coding arena compared to its rivals. This limitation, combined with its operational inefficiency, positions it poorly against competitors like MiMo and Grok that offer high accuracy without the operational overhead.
Further Reading
- Gemini 3.5 Flash | Gemini API - Google AI for Developers — Google AI for Developers
- Gemini 3.5 Flash | Gemini Enterprise Agent Platform — Google Cloud Documentation
- Gemini 3.5 Flash - Google DeepMind — Google DeepMind
- Gemini 3 Flash: frontier intelligence built for speed — Google Blog
- Models | Gemini API - Google AI for Developers — Google AI for Developers
