Gemini 3 Flash
A multimodal language model developed by Google, optimized for speed and cost efficiency. It features strong visual understanding capabilities and serves as the high-efficiency variant within the Gemini 3 model family, targeting use cases that prioritize rapid inference and reduced computational overhead.
Core Facts
Timeline (last 90 days)
On 30 real receipt photos from the CORD public dataset, Gemini 3 Flash's fabrication rate jumped from 21% on synthetic bills to 73%, with confidence still at 95, while its own detection correctly stated the total was uncalculable in all 30/30 cases.
The median confidence difference across models when fabricating totals is enormous: Gemini 3 Flash and Flash-Lite at 95, GPT-5.4 nano at 62, and Claude Sonnet 5 at only 20.
When asked directly whether the total was visible, Claude, Gemini 3 Flash, and Flash-Lite answered across all 48 receipts that the total was not visible and could not be calculated, yet when asked directly for the total, they still fabricated one 21–44% of the time.
When the total was obscured but each line item was visible, Claude and Gemini 3 Flash got all 48 correct, with all failures coming from the smaller models (Flash-Lite failed 31%, nano failed 79%), and nano got only 12/48 correct even when explicitly instructed to add up the items.
Confirm-then-answer (probing before extracting within the same conversation) reduces each model's fabrication rate to 0–2%, while Gemini 3 Flash is also able to correctly calculate 98% of recoverable totals.
The test covered four vision models: Claude Sonnet 5 (Anthropic), Gemini 3 Flash (Google), Gemini 3.1 Flash-Lite (Google), and GPT-5.4 nano (OpenAI).
Fabrication rates when the total is fully obscured: Claude Sonnet 5 at 38%, Gemini 3 Flash at 21%, Gemini 3.1 Flash-Lite at 44%, GPT-5.4 nano at 73%; fabrication rates when the total is half-obscured: Claude at 92%, the other three models all at 100%.
The author used Gemini 3 Flash to generate audience-style questions for 120 video caption passages to form "question–passage" data pairs, and after less than three minutes of training, the top-1 hit rate improved from approximately two-thirds to approximately three-quarters.
The author built a custom agent around the open-source coding agent PI, running Gemini 3 Flash, equipped with Shell and SQL tools, operating a Postgres database running on Neon.
On Cursor Bench, Sonnet 5.5 consumed 271,920 tokens per task, surpassing Opus's approximately 218k and roughly double that of Gemini 3 Flash.
1 more timeline events
All Facts (10)
The test covered four vision models: Claude Sonnet 5 (Anthropic), Gemini 3 Flash (Google), Gemini 3.1 Flash-Lite (Google), and GPT-5.4 nano (OpenAI).
65%UnverifiedOn 30 real receipt photos from the CORD public dataset, Gemini 3 Flash's fabrication rate jumped from 21% on synthetic bills to 73%, with confidence still at 95, while its own detection correctly stated the total was uncalculable in all 30/30 cases.
50%UnverifiedWhen asked directly whether the total was visible, Claude, Gemini 3 Flash, and Flash-Lite answered across all 48 receipts that the total was not visible and could not be calculated, yet when asked directly for the total, they still fabricated one 21–44% of the time.
50%UnverifiedWhen the total was obscured but each line item was visible, Claude and Gemini 3 Flash got all 48 correct, with all failures coming from the smaller models (Flash-Lite failed 31%, nano failed 79%), and nano got only 12/48 correct even when explicitly instructed to add up the items.
50%UnverifiedConfirm-then-answer (probing before extracting within the same conversation) reduces each model's fabrication rate to 0–2%, while Gemini 3 Flash is also able to correctly calculate 98% of recoverable totals.
50%UnverifiedFabrication rates when the total is fully obscured: Claude Sonnet 5 at 38%, Gemini 3 Flash at 21%, Gemini 3.1 Flash-Lite at 44%, GPT-5.4 nano at 73%; fabrication rates when the total is half-obscured: Claude at 92%, the other three models all at 100%.
50%UnverifiedThe median confidence difference across models when fabricating totals is enormous: Gemini 3 Flash and Flash-Lite at 95, GPT-5.4 nano at 62, and Claude Sonnet 5 at only 20.
50%UnverifiedThe author used Gemini 3 Flash to generate audience-style questions for 120 video caption passages to form "question–passage" data pairs, and after less than three minutes of training, the top-1 hit rate improved from approximately two-thirds to approximately three-quarters.
50%UnverifiedThe author built a custom agent around the open-source coding agent PI, running Gemini 3 Flash, equipped with Shell and SQL tools, operating a Postgres database running on Neon.
50%UnverifiedOn Cursor Bench, Sonnet 5.5 consumed 271,920 tokens per task, surpassing Opus's approximately 218k and roughly double that of Gemini 3 Flash.
50%