Gemini 3.5 Pro Spotted on LM Arena — Google's Long-Delayed Flagship May Finally Launch

The timeline of a very delayed flagship

Let's recap the journey. At Google I/O 2026 on May 19, Google announced Gemini 3.5 Flash as the first member of the 3.5 family, calling it "frontier intelligence with action." The Pro version was promised for June. On stage, Google said it was "showing great improvements."

June came and went. On July 16, Bloomberg reported that Google was "taking time to try to improve [Gemini 3.5 Pro's] capabilities, particularly in coding." According to the report, a late-June training data update aimed at improving coding "results were disappointing." 9to5Google confirmed that development saw what looked like a reset between I/O and the missed launch.

Then on July 21, Google surprised everyone — not with 3.5 Pro, but with an entire triple-launch: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The blog post acknowledged Pro's absence: "Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it's ready." It also dropped the news that Google had started its "most ambitious pre-training run yet, for Gemini 4."

Now, on August 1–2, Gemini 3.5 Pro has been spotted on LM Arena (formerly Chatbot Arena) — the crowdsourced benchmark platform run by LMSYS. Historically, Google uses Arena for final-stage blind testing before public launches. We saw this pattern with Gemini 3 Flash (tested May 2025, launched days later), Gemini 3.5 Flash (appeared on Arena shortly before I/O), and most recently Gemini 3.6 Flash (spotted July 1, launched July 21).

What to expect from 3.5 Pro

Multiple leaks and reports over the past two months have painted a picture of what Gemini 3.5 Pro brings to the table:

• 2 million token context window — double the 1M tokens of 3.1 Pro and 3.6 Flash, putting it on par with Grok 4.20's 2M window.
• Deep thinking mode — extended reasoning with multiple thinking levels, an escalation from the thinking modes already available in the Gemini app.
• Improved math and reasoning — Google reportedly delayed the model specifically to "upgrade math" performance (per Geeky Gadgets, July 6), aiming to compete with Claude Opus 4.8 Thinking and o3.
• Coding that actually works — the whole reason for the delay. Bloomberg's report noted the 3.5 Pro team was targeting "75% of all new code at Google being AI-generated," a milestone Google had already crossed internally with existing tools.

If Pro matches the Flash trend, expect it to be more token-efficient than its predecessor. Gemini 3.6 Flash already uses 17% fewer output tokens than 3.5 Flash. Pro could follow the same pattern — fewer words, more precision.

Where it fits in the competitive landscape

The text leaderboard on LM Arena (updated August 1) has been dominated by Anthropic's Claude Fable 5 (1509 Elo), followed by Claude Opus variants. Google's best showing is Gemini 3 Pro at rank 10 (1486 Elo), with 3.1 Pro Preview at rank 11 (1485). The recently launched Gemini 3.6 Flash sits at rank 15 (1483 Elo, preliminary), while 3.5 Flash High is at rank 18 (1476).

If 3.5 Pro can crack 1500 Elo — Fable 5 territory — it would re-establish Google as a top-tier contender in the text quality race. Early Arena testers are already comparing outputs on social media, though the model appears under a codename (typical for pre-launch testing).

Here is how Google's current model lineup stacks up on cost:

*EUR conversions at approximate August 2026 rates (~1 USD = 0.91 EUR). 3.5 Pro pricing is an estimate based on the Pro-tier position between Flash and competitors' premium tiers.

For context, Anthropic's Claude Opus 5 costs $5/$25 per 1M tokens (~€4.50/€22.80), while GPT 5.5 is $5/$30 (~€4.50/€27.30). If Google prices 3.5 Pro aggressively — say $3/$15 — it could undercut both while delivering competitive quality.

The European perspective

Gemini is fully available in the EU, including through the Gemini app, Google AI Studio, and Vertex AI. Google operates cloud regions in the EU (Belgium, Finland, Netherlands, Germany, France) with data residency options, which matters for companies navigating GDPR compliance.

The EU AI Act entered into force in 2024, with the rules for general-purpose AI models (GPAI) entering application in August 2025. Gemini falls under the GPAI tier — Google, as a systemically important provider, must comply with transparency requirements, model documentation, and risk assessments. The recent Frontier Safety framework update for 3.6 Flash (covering CBRN and cyber offense safeguards) is likely a preview of what 3.5 Pro's compliance posture will look like.

For European developers and companies, the pricing matters: Google's AI Ultra subscription runs at $100/month (~€91), and AI Pro at $19.99/month (~€18.20). EU businesses using the API pay in USD with European cloud region options. The aggressive pricing of the Flash series — $0.30/1M input for 3.5 Flash-Lite — already makes Gemini one of the cheapest frontier-quality APIs available in Europe, undercutting even DeepSeek V4 Pro ($0.43/1M).

Why the Arena appearance matters

Google's Arena testing pattern is well-established. In the past year alone:

• Gemini 3 Flash appeared on Arena in late April 2025, launched weeks later at I/O
• Gemini 3.5 Flash was tested on Arena in the days leading up to its I/O 2026 announcement
• Gemini 3.6 Flash was spotted on Arena July 1 (TestingCatalog caught it), and Google launched it July 21

Each time, the gap between Arena appearance and official launch has been roughly 1–3 weeks. If that pattern holds, we are looking at a mid-August 2026 launch for Gemini 3.5 Pro — roughly three months after the original June target, but still in time to matter before the rumored Gemini 4 pre-training run starts producing results.

The question now is not if 3.5 Pro ships, but whether the extra two months of tuning were enough to close the gap with Anthropic's and OpenAI's current flagships. The Arena blind tests will give us the first real answer — and users are voting right now.