← Back to directory
Google DeepMind
Gemini 3.5 Flash
Released May 19, 2026 · knowledge cutoff 2025-01
- Status
- Superseded
- Location
- United States / United Kingdom
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 66K
- Speed
- —
- Price ($/MTok in / out)
- $1.50 / $9
- Cost per task (high effort)
- $0.59
- Open weights
- No
Benchmarks5 of 8 reported
SWE-bench Verified
81%
Terminal-Bench 2.1
76.2%
Humanity's Last Exam
40.2%
LMArena Elo
1481
ARC-AGI-2
72.1%
MMLU-Pro
—
GPQA Diamond
—
AIME (latest)
—
Strengths & weaknesses
Strengths
- The fastest capable option of its moment — iteration speed beats waiting for a better answer
- Near-Pro coding quality at Flash latency
- Reasoning context carries across turns, so multi-turn agent loops don't re-derive everything
Weaknesses
- Instruction-following is its weakest axis; tightly constrained output formats slip
- Output cap low enough to truncate large generated files
- Default effort level shifted at launch, quietly changing behavior for existing prompts
Announced at Google I/O 2026; Terminal-Bench 2.1 76.2%. HLE 40.2 and ARC-AGI-2 72.1 are Google's own figures on the official Gemini 3.5 Flash model card (verified 2026-07-29). SWE-bench Verified 81.0 came from I/O reporting and is NOT on that model card, which reports SWE-Bench Pro (Public) 55.1% instead — treat 81.0 as unconfirmed. LMArena Elo 1481 at launch window.
For developers
API model strings
Not researched
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor