Breaking

Unverified Gemini 3.5 Pro Benchmark Leak Claims Edge Over Claude and GPT, but Google Has Not Confirmed It

July 12, 2026 at 12:15 EDT

Foundation Models
AI Agents

A set of unverified benchmark figures circulating this month claims Google's still-unreleased Gemini 3.5 Pro outperforms Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 in internal evaluations, but none of the numbers have been confirmed by Google and the model remains unpublished. The leaked material describes a model in private validation and testing, with a purported public rollout target of July 17, alongside claims of sharply improved zero-shot performance over the earlier Gemini 3.1 Pro.

Unverified Leak · Gemini 3.5 Pro

Leaked benchmarks claim Google's unreleased flagship beats its rivals — none of it is confirmed

Circulating figures pitch Gemini 3.5 Pro as topping Claude Fable 5 and GPT-5.6, with a 2M-token window and a July 17 target. Google has shipped none of it — only the smaller Gemini 3.5 Flash is public and documented.

✓ Confirmed & Shipped

Gemini 3.5 Flash

Generally available since ~May 19, 2026. Model card, API and verified benchmarks published. Priced at $1.50 / $9 per 1M input/output tokens.

⚠ Unverified & Unpublished

Gemini 3.5 Pro

No API, no model ID, no model card, no verified benchmark sheet. Pricing, availability and release date all unannounced.

The only real numbers

Confirmed Flash gains over Gemini 3.1 Pro (agent & coding)

Terminal-Bench 2.1

70.3%

3.1 Pro

76.2%

3.5 Flash

MCP Atlas

78.2%

3.1 Pro

83.6%

3.5 Flash

Flash gained on agentic and coding tasks but reportedly slipped on some reasoning and long-context work — the gap Pro is expected to close.

Leaked claims — awaiting official confirmation

token context window claimed (Flash: 1M, confirmed)

Jul 17

purported public rollout target

Deep Think + UI/game gen

alleged new capabilities

The optimistic read

A genuine opening for Google to retake a performance lead — if the Pro variant delivers long-horizon stability and deeper reasoning on top of Flash's agent focus.

The cautious read

Internal evals are self-selected and benchmark choice can bias comparisons. June checkpoint leaks drew harsh verdicts vs Claude Fable 5 and GPT-5.5. Trust nothing until Google publishes.

The figures hint at an agent- and coding-oriented flagship — but "beats Claude Fable 5 and GPT-5.6," the 2M window and the July 17 date remain unconfirmed.

Leaks have run ahead of official Gemini announcements before, later corrected by Google's own materials. Confirmed pricing exists only for Flash.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.