Gemini 3.8 Live: Google Still Trails OpenAI, Anthropic

Google just dropped two new voice AI models — Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking — and the pitch sounds impressive: real-time visual processing, 97 languages with seamless mid-conversation switching, and a thinking mode that reasons out loud while it works. But the AI landscape in September 2026 isn’t the same one Google dominated with search. Here’s where Gemini actually stands, and why impressive specs aren’t the same as industry leadership.

What Gemini 3.8 Live Actually Does

At its core, Gemini 3.8 Live is a voice-first AI model designed for natural conversation. Think of it as Google’s answer to the assistant experience OpenAI demonstrated with GPT-4o’s voice mode last year — except Google claims to go further.

The standout features:

97-language support with automatic detection. You can start a conversation in English, switch to Japanese, then jump to Portuguese, and the model follows without missing a beat. No manual language switching required.

Near real-time visual processing. Point your camera at something and the model can discuss what it sees while you’re talking — not after a delay, but during the conversation.

Background tool execution. The model can search the web, pull up documents, and call APIs without interrupting the voice conversation. You keep talking; it keeps working.

Meanwhile, the Extended Thinking variant adds verbal reasoning. When it hits a complex problem, it narrates its thought process — “Let me check that,” “I’m comparing these two options” — while working through multi-step tasks in the background. It’s the AI equivalent of thinking out loud, and it feels more natural than watching a loading spinner.

Google says all AI-generated audio includes SynthID watermarking, which means the output can be identified as machine-generated. That’s a responsible move, and one that OpenAI and Anthropic haven’t fully matched yet.

The Benchmarks Tell a Selective Story

Google’s announcement highlights some genuinely strong numbers. The Extended Thinking model hit 82.6 on the Artificial Analysis Speech-to-Speech Index, taking the top spot. On Big Bench Audio reasoning, it scored 97.7%. On ServiceNow’s EVA-Bench for complex workflows, both models “push the Pareto Frontier.”

At first glance, that sounds like Google is winning. But I’d encourage skepticism about which benchmarks a company chooses to feature in a press release.

Here’s what Google didn’t highlight. According to independent comparisons, Claude Sonnet 4.6 scores 82.1% on SWE-bench Verified — one of the most respected coding benchmarks — while Gemini 3.1 Pro manages 63.8%. That’s an 18-point gap in the task category developers care about most. The newest Gemini 3.8 Flash closes this gap somewhat, but Google notably didn’t publish SWE-bench numbers for its latest models.

When a company highlights audio reasoning scores but avoids coding benchmarks, it tells you where the confidence lies — and where it doesn’t.

More broadly, the benchmark landscape paints a picture of genuine three-way competition, with each player owning different territory:

Category Leader Score Coding (SWE-bench) Claude 82.1% Scientific Reasoning (GPQA) Gemini 94.1% Expert Reasoning (HLE-Verified) Statistical tie ~54.5% Speech/Audio Gemini 82.6 Agentic Tasks (τ-Voice) Gemini 68.6%

Gemini leads in scientific reasoning and voice/audio tasks. Meanwhile, Claude leads in coding and software engineering. At the low end, OpenAI’s GPT-5.6 offers the best price-to-performance ratio. Each model has a genuine claim to “best” — it just depends on what you’re measuring.

Why Developers Still Reach for the Competition

Benchmarks measure capability. Developer adoption measures usefulness. And right now, Claude and GPT-5 dominate the developer ecosystem in ways Gemini hasn’t matched.

Interestingly, part of this is pricing — but not in the way you’d expect. Gemini Flash is actually the cheapest option at $0.75 per million input tokens during the introductory period (it doubles in January 2027). That undercuts Claude Haiku at $1.00 and GPT-5.6 Luna at $0.20 on output tokens. Price isn’t Gemini’s problem.

The problem is ecosystem lock-in and trust. OpenAI had a multi-year head start with developers who built their products on GPT-3, GPT-4, and now GPT-5. Anthropic earned developer loyalty by focusing relentlessly on code quality and tool-use reliability. Switching AI providers isn’t like switching text editors — it requires rewriting prompts, adjusting output parsing, and re-testing edge cases. The switching cost creates inertia that Google’s better benchmarks haven’t overcome.

Feature Gemini 3.8 Flash Claude Opus 5 GPT-5.6 Sol Input Price (per 1M tokens) $0.75 $5.00 $4.00 Output Price (per 1M tokens) $3.75 $25.00 $20.00 Context Window ~1.05M 1M ~1.05M Max Output 65,536 tokens 128,000 tokens 128,000 tokens Input Modalities Text, image, video, audio, PDF Text, image, PDF Text, image

There’s one area where Google holds an unmatched advantage: multimodal input. Gemini is the only major model that processes video and audio natively alongside text and images. If your application needs to analyze a video clip or transcribe a podcast in real time, Gemini is still the only serious option. Claude and GPT-5 don’t touch video or audio as input modalities.

Google’s Real AI Problem Isn’t Technical

Here’s what I think Google’s actual challenge is, and it has nothing to do with benchmarks or pricing.

Google has a positioning problem. When you think “AI coding assistant,” you think Claude or GitHub Copilot. For AI chatbots, ChatGPT is often the first name that comes to mind. In the enterprise space, Microsoft Copilot might be the first product you consider. Google sits in an uncomfortable middle — strong everywhere, dominant nowhere.

Gemini 3.8 Live is technically impressive. The 97-language support is best-in-class. Its voice quality also stands out. Meanwhile, the pricing is competitive. But “competitive” doesn’t translate into market share when competitors are already embedded in people’s workflows. Google needs something that makes developers and users say “I have to use Gemini for this” rather than “Gemini could also do this.”

The multimodal edge might be that something. If Google leans harder into video and audio processing — building tools that only Gemini can power — it creates a moat the others can’t cross quickly. The question is whether Google’s product teams will capitalize on that technical advantage or continue playing catch-up on text-based benchmarks where Claude and OpenAI already own the narrative.

For now, Gemini 3.8 Live is a strong product launch that won’t meaningfully change the competitive dynamics. Google’s AI division needs a category-defining moment, not another incremental improvement.

FAQs

Is Gemini 3.8 Live free to use?

General users can access Gemini 3.8 Live through Google Search Live at no cost. Extended Thinking features require a Gemini subscriber plan. Developers access both models through the Gemini API and Google AI Studio with usage-based pricing.

How does Gemini 3.8 Live compare to ChatGPT’s voice mode?

Both offer conversational voice AI, but Gemini supports 97 languages with automatic switching — significantly more than ChatGPT. Gemini also processes video and audio natively, which ChatGPT currently doesn’t support as input modalities.

What does Extended Thinking do differently?

Extended Thinking lets the model reason out loud while processing complex tasks. Instead of silent processing, it narrates its approach — explaining what it’s checking and comparing — while simultaneously running background tasks. It’s designed for multi-step problems that need sequential reasoning.

Is Gemini better than Claude for coding?

No. Independent benchmarks show Claude leading by 18+ percentage points on SWE-bench Verified, the standard coding evaluation. Gemini 3.8 Flash is competitive on some agentic coding tasks, but Claude remains the developer favorite for software engineering in 2026. Google’s own AI smart glasses showcase Gemini’s strengths in visual and voice processing, not code generation.

Will Gemini 3.8 Live replace Google Assistant?

Google hasn’t announced plans to retire Google Assistant, but Gemini is increasingly taking over Assistant’s functions on newer Android devices. The 3.8 Live model’s conversational capabilities significantly exceed what traditional Assistant offered, making a gradual transition likely.