Skip to content
CrucibleMark

Gemini 3.8 Flash

Gemini 3.8 Flash is Google’s most intelligent Flash model (GA since September 2, 2026), built on 3.7 Flash for long-horizon software engineering, autonomous agents, and complex enterprise workflows. Three thinking levels (low, medium, high, default medium) control reasoning depth; one million tokens of context and 65,536 output tokens fit entire repositories in a single request. Introductory pricing of 0.75 / 3.75 USD per million tokens through year-end, then 1.50 / 7.50.

Google Version 3.8 Commercial use permitted Dense 1000 K Context $0.75 / $3.75 per 1M

  • Proprietary
  • Frontier
  • OpenRouter
  • Text
  • Vision
  • 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 100 40 64
HTTP Fetch & Extract 80 100 90
Tool Failure Handling (404) 80 80 80
Web Search & Tool Selection 100 80 91
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 80 90

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
3.39
First Request
MCP
1.18
Protocol Latency
Synthesis
8.58
Response Generation
Total
78.9
Sum of All Phases
Token
11911
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: tool usage is strong, but tool call validity was not consistently clean, and synthesis fidelity on current-events research is not stable enough. The combined score is good, but for production MCP pipelines the trust edge matters more than the average.

Tool Execution Profile

Gemini 3.8 Flash demonstrates genuine tool intelligence, not just rigid pattern behavior. On the Web Search & Tool Selection test — which checks the choice between search and direct retrieval without an explicit hint — it selects the right tool reliably. That speaks to robust orchestration capability in dynamic pipelines. On the URL Construction & Fetch test, which measures deriving a target URL from internal knowledge and then fetching it, it remains usable but not deterministic enough. That is exactly where the practical risk lies: not in whether it wants to use tools, but in whether the concrete call is MCP-compliant and reproducible across every variant. The fact that tool call validity is false overall is a clear warning sign for production chaining.

Synthesis Fidelity

How well does it compress tool results? Solid, but not reliably precise. The strong performance on HTTP Fetch & Extract shows it can extract and consolidate structured content from retrieved material well. The overall P2 score is nonetheless only middling, because compression loses sharpness on knowledge-adjacent research.

Does it stay within the tool result or fall back on training? Not reliably enough. On the Honeypot EU License Research test — which checks whether current license restrictions are actually retrieved from web sources — synthesis drops sharply to P2 40. It does not hallucinate overtly, but there is also no strong trust signal that the answer remains strictly anchored to the retrieved sources. For compliance, legal, and policy pipelines, that is too weak.

Error Resilience

On the 404 test, which checks for transparent behavior when a tool call fails, the model responds in a production-appropriate way. It does not fabricate page content and keeps error communication clean. That is acceptable for operations, because a pipeline controller can reliably continue working with responses like that.

Operational Profile

Call 1: 3.39s. Call 2: 8.58s. MCP latency: 1.18s. Total: 78.90s. For a Flash model, that is slow for an end-to-end run. Price: $0.75/1M input, $3.75/1M output, web search separate. Cost profile: cheap per token, but not cheap per complex agentic run.

Conclusion & Recommendation

Suitable for agentic research and retrieval pipelines where tool selection, search strategy, and clean error handling matter more than highly precise, legally or compliance-sound final synthesis. Not the first choice for MCP pipelines with strict requirements around current factual accuracy, citable policy statements, or deterministic tool calls. Deploy only with guardrails: tool output logging, schema validation, URL and call checks, and downstream verification before any externally visible response.

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.