Gemini 3.7 Flash High Scores 78.8 on LiveBench at $0.157 Per Task — Cheapest in the Top Six
The August 2026 LiveBench leaderboard shows Gemini 3.7 Flash High at 78.8 overall — sixth globally — at $0.157 per successful task. That cost is the lowest among the top six ranked models by a wide margin: the next cheapest is Kimi K3 open at $0.348, more than twice the price.
| Model | Overall | Reasoning | Mathematics | Cost/Task |
|---|---|---|---|---|
| Claude Fable 5 Max Effort | 83.0 | 89.7 | 96.0 | $1.439 |
| GPT-5.6 Sol Max Effort | 81.0 | 91.7 | 96.2 | $0.515 |
| GPT-5.5 Thinking xHigh | 80.2 | 89.7 | 95.9 | $0.435 |
| Claude Opus 5 Thinking Max | 80.1 | 91.2 | 95.7 | $0.699 |
| Kimi K3 open | 79.2 | 90.7 | 84.4 | $0.348 |
| Gemini 3.7 Flash High | 78.8 | 87.8 | 93.5 | $0.157 |
For 2.2 fewer overall points than GPT-5.6 Sol, Gemini 3.7 Flash High costs one-third as much per task. Against Claude Fable 5, the gap is 4.2 points — at one-ninth the cost.
Where It Holds and Where It Drops
Gemini 3.7 Flash High’s category breakdown is not uniform. Two areas are close to full-frontier performance:
- Mathematics: 93.5, trailing GPT-5.6 Sol by 2.7 points and Claude Fable 5 by 2.5
- Reasoning: 87.8, within 2 points of Claude Fable 5’s 89.7 and GPT-5.5 Thinking’s 89.7
Two areas lag:
- Data Analysis: 68.0, roughly 10–13 points below models ranked above it
- Agentic Coding: 58.3, consistent with the wider frontier pattern of agentic coding underperforming overall scores
The instruction-following score of 79.9 is actually higher than all five models ranked above it — a dimension where Flash-tier optimisation for following constraints precisely pays off.
The Cost-Efficiency Framing
Cost-per-task on LiveBench includes inference compute across all benchmark attempts, including failed ones. A model that scores higher per dollar is not just cheaper at the API level — it succeeds more often per unit of spend in actual evaluation conditions.
At $0.157/task, Gemini 3.7 Flash High’s mathematics and reasoning performance creates a practical case for workloads that are heavy on those two categories. A system handling quantitative analysis, structured reasoning, or mathematical problem-solving can approach top-six benchmark performance for roughly what lower-ranked models cost on inference alone.
The exception is agentic coding. At 58.3, Gemini 3.7 Flash High trails Claude Opus 5 Thinking Max by 6.9 points in the category most representative of autonomous software work. For teams running coding agents rather than reasoning or maths pipelines, the cost advantage narrows because success rates diverge in that task class.
Context: The Flash Tier in 2026
The LiveBench top six in August 2026 includes Gemini 3.7 Flash High — a model from Google’s inference-optimised tier — alongside five models from flagship or extended-compute inference configurations. That was not the case a year ago, when Flash-class models tracked 10–15 points below full-flagship performance on comparable evaluations.
The narrowing has come from targeted post-training for specific task categories rather than raw parameter count. Google’s approach with the Gemini 3.x Flash series has prioritised reasoning and mathematics gains, which are visible in the 87.8 and 93.5 scores. General data analysis, which requires broader world knowledge and longer-context table operations, has not seen the same acceleration.
Gemini 3.7 Flash launched August 13 with a focus on DeepSWE coding performance. The LiveBench data reflects a different evaluation surface — one where the model’s cost-efficiency advantage is clearest.