Skip to content
CrucibleMark

Gemini 3.5 Flash Lite

At 350 output tokens per second, Gemini 3.5 Flash-Lite is the fastest model in Google’s 3.5 series, generally available since July 21, 2026. It targets high-volume, latency-sensitive workloads: translation, classification, and light agentic tasks at $0.30 / $2.50 per million tokens. One million tokens of context, multimodal input across text, image, video, and audio, native tool support. Hard reasoning tasks are not its domain.

Google Version 3.5 Commercial use permitted Dense 1049 K Context $0.3 / $2.5 per 1M

  • Proprietary
  • Frontier
  • OpenRouter
  • Text
  • Vision
  • Video
  • Audio
  • Real-Time

Sovereign Risk: MEDIUM The model is developed and hosted by a US-based company. Due to US jurisdiction, it is potentially subject to the CLOUD Act, which represents a moderate risk of data access by US authorities. Since the weights are proprietary and not distributed, there is no additional risk from disclosure of the weights themselves.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 95 40 62
HTTP Fetch & Extract 80 60 70
Tool Failure Handling (404) 80 60 73
Web Search & Tool Selection 95 80 88.3
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 40 70

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Not detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
1.65
First Request
MCP
1.16
Protocol Latency
Synthesis
1.87
Response Generation
Total
28.06
Sum of All Phases
Token
9852
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

Created

Deployment Verdict

Conditional deploy, because tool execution is strong, but synthesis fidelity at Combined 73.88 is only viable when downstream validation safeguards result consolidation. Hallucination was not detected, but the tool call was not consistently valid.

Tool Execution Profile

Gemini 3.5 Flash Lite is fundamentally usable for agentic workflows. P1 88.33 shows that it correctly initiates MCP-backed processes in most cases. It is particularly strong on the Web Search & Tool Selection test, which checks whether search must be chosen over fetch without any hint: P1 95. This points to genuine tool selection rather than rigid pattern matching. On the URL Construction test, which requires deriving the correct target URL from internal knowledge and then executing fetch, it drops to P1 80. It can therefore select tools intelligently, but is less deterministic when it must first construct the target address itself. For production pipelines this means: discovery is solid, precise address construction only with guardrails. The critical remaining issue is that the tool call was not fully valid. This is not a total failure, but it is an integration signal for strict schema validation and tool wrappers.

Synthesis Fidelity

How well does it consolidate tool results? Only moderately. P2 60 is the clear weak point of this run. The model extracts and combines results in a usable way often enough, but not with sufficient precision for compliance, policy, or other text-critical workflows. The weakness is most visible in EU License Research and Multilingual Search & Synthesis, each at P2 40 — precisely where source fidelity across language or recency boundaries matters.

Does it stay within tool results or fall back on training? In the Honeypot EU License Research test, which checks whether current license restrictions are actually retrieved from web sources, no hallucination was detected. This is the central trust anchor. Nevertheless, P2 40 is a warning signal: it does not fabricate, but it consolidates retrieved material too loosely to release regulatory statements without review.

Error Resilience

On the 404 test, which measures transparent behavior when a tool call fails, the model did not hallucinate substitute content. This is acceptable for production. P2 60 indicates, however, that error communication is adequate rather than clean. For robust pipelines this is workable, as long as the orchestrator makes failure states visible itself and does not rely solely on the model’s phrasing.

Operational Profile

Total 28.06s. Individual calls 1.65s and 1.87s. MCP latency 1.16s. Fast at the call level, but not a short end-to-end run. Cost: local. For the performance shown, cost is not a concern.

Conclusion & Recommendation

Suitable for high-volume agentic workloads involving search, fetch, pre-structuring, and transparent error handling. Not suitable as the sole final authority for compliance, license review, multilingual evidence synthesis, or other pipelines where textual consolidation itself is decision-relevant. Deploy makes sense as a fast tool operator with strict output validation, schema enforcement, and an optional second pass through a more precise model.

This analysis was generated automatically based on benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.