Two Leaks, Two Verdicts
| Leak 1 (Bullish) | Leak 2 (Bearish) | |
|---|---|---|
| vs Claude Fable 5 | Beats it on reasoning and coding | Trails it |
| vs GPT-5.6 | Beats it in internal evals | Trails it, weak on code infilling |
| Source | X user citing internal evals, plus a separate LMSYS Arena sighting | Differently-sourced leak, no data published |
| Reproducible data | None | None |
The bullish leak, amplified widely on X, claims Gemini 3.5 Pro beats GPT-5.6 and Claude Fable 5 in internal evaluations, with major zero-shot gains over Gemini 3.1 Pro. A separate developer earlier claimed an unreleased model tagged “Gemini 3.5 Pro” beat Fable 5 on a front-end coding task in a public LMSYS Arena test. The bearish leak says the opposite: Pro is struggling on advanced reasoning, coding, and long-horizon tasks, landing behind both rivals, consistent with reports that Google delayed the launch from June to July specifically to fix quality issues enterprise testers flagged.
What’s Actually Confirmed
Google unveiled the Gemini 3.5 family at I/O on May 19, 2026 and shipped Gemini 3.5 Flash the same day with full published benchmarks. Pro was promised “next month,” slipped past June, and sits in limited Vertex AI enterprise preview. July 17 is a leaked internal target, not an announced date, and June already slipped once. Reported specs, a 2 million-token context window and pricing near $15/$60 per million tokens, come from leaks, not Google.
Why the Contradiction Matters More Than Either Leak
Stealth-testing frontier models under disguised names inside public arenas is now routine enough that X communities track it as a recurring beat, reposting sightings and shaping opinion months before any official number exists. GPT-5.6’s own uneven benchmark performance and Meta’s cherry-picked Muse Spark comparison chart show the same pattern this month: labs and communities alike are shaping model reputations through selective or unverified benchmarks before anyone can reproduce the numbers. Two leaks that flatly contradict each other on the same unreleased model is the clearest evidence yet that neither should move anyone’s opinion. Wait for July 17, if it holds, and for independent evals after that.
Follow us on Bluesky, LinkedIn, X, and Telegram to Get Instant Updates