sugra.blog

BlogLLM Benchmark Scores as Published, Side by Side

Gemini 3.8 Flash: Benchmark Scores as Published

Every result the catalog holds for Gemini 3.8 Flash: 51 tests in 9 areas, from 8 pages. 45 printed by Google itself, the rest by the company named beside each number. Each number is as printed; nothing is ranked or rescored.

Developer GoogleReleased 2026-09-02Context window 1.05MMax output 65.5KInput Text, image, video, audio, PDFOutput TextWeights ClosedData version 2026103001, as of 2026-10-30

Compare in the table

Knowledge Work

TestValuePrinted bySetup as printed
GDPval-AA v21545Google, Gemini 3.8 Flash model cardchecked 2026-10-15-
Vals Finance Agent v261.4%Google, Gemini 3.8 Flash model cardchecked 2026-10-15-
Harvey's Legal Agent Benchmark10%Google, Gemini 3.8 Flash model cardchecked 2026-10-15-

Agentic Coding

TestValuePrinted bySetup as printed
Terminal-Bench 2.189.4%Google, Gemini 3.8 Flash model cardchecked 2026-10-15-
DeepSWE v1.173.7%Google, Gemini 3.8 Flash model cardchecked 2026-10-15; printed on 1 more page-
Terminal-Bench 4.019.1%Google, Gemini 3.8 Flash model cardchecked 2026-10-15; printed on 1 more page-
FrontierCode v1.1 Main43.6%OpenAI, GPT-6 Astrachecked 2026-10-14-
FrontierCode v1.1 Extended56.3%OpenAI, GPT-6 Astrachecked 2026-10-14-
Artificial Analysis Coding Agent Index v1.461.2OpenAI, GPT-6 Astrachecked 2026-10-14-

Science and Math

TestValuePrinted bySetup as printed
GPQA Diamond95.3%OpenAI, GPT-6 Astrachecked 2026-10-14-
BioMysteryBench Human Difficult56.5%Google, Gemini 3.8 Flash model cardchecked 2026-10-15-
BioMysteryBench Human Solvable88.8%Google, Gemini 3.8 Flash model cardchecked 2026-10-15-
LAB-Bench 286.2%Google, Gemini 3.8 Flash model cardchecked 2026-10-15-

Computer Use

TestValuePrinted bySetup as printed
OSWorld 2.0 partial59%Google, Gemini 3.8 Flash model cardchecked 2026-10-15Tools: batch tool enabled

Multimodal Understanding

TestValuePrinted bySetup as printed
CharXiv Reasoning (no tools)86.2%Google, Gemini 3.8 Flash model cardchecked 2026-10-15Tools: none
LVBench87.1%Google, Gemini 3.8 Flash model cardchecked 2026-10-15Model setup as reported: Model card results table ("Results as of September, 2026"); effort not stated. Methodology page linked. Value printed as "87.1% (static)".
87.8%Google, Gemini 3.8 Flash model cardchecked 2026-10-15Model setup as reported: Model card results table ("Results as of September, 2026"); effort not stated. Methodology page linked. Value printed as "87.8% (agentic)".
GDP.pdf (all-pass rate)35%Google, Gemini 3.8 Flash model cardchecked 2026-10-15-

General Knowledge and Reasoning

TestValuePrinted bySetup as printed
Humanity's Last Exam (Verified)54.9%Google, Gemini 3.8 Flash model cardchecked 2026-10-15-

Healthcare

TestValuePrinted bySetup as printed
HealthBench Professional (length-adjusted)52.1%OpenAI, GPT-6 Astrachecked 2026-10-14-

Safety and Alignment

TestValuePrinted bySetup as printed
Gray Swan IPI (attack success rate after k attempts)attack success rate after k attempts, lower is better5.5%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19; printed on 3 more pages-
Primary misbehavior judge score, overall (18 tasks)lower is better4.07Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Agent Harnesses: Helpfulness68.81%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Agent Harnesses: Safe Action Rate84.95%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Child Sexual Abuse and Exploitation Safety100%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Coding Agent Robustness: survival at 40 attempts37%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Computer Use Robustness (adaptive scenarios): survival at 100 attempts69%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Contextual Security Robustness: baseline (no adversary)63%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Contextual Security Robustness: survival at 1 attempt (static)63%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Contextual Security Robustness: survival at 15 attempts (adaptive)23%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Dangerous Content Safety (multi-turn)97.8%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Dangerous Content Safety (single-turn)94.9%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Grounding Robustness (RAG indirect prompt injection): survival at 40 attempts81%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Hate and Harassment Safety (multi-turn)98.9%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Hate and Harassment Safety (single-turn)95.6%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Helpfulness Across Benign Over-Refusal Benchmarks (multi-turn)91%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Helpfulness Across Benign Over-Refusal Benchmarks (multimodal)69.9%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Helpfulness Across Benign Over-Refusal Benchmarks (single-turn)95.2%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Personal-Device Assistant Robustness: adaptive survival (ARR@5)91%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Personal-Device Assistant Robustness: clean control task utility89%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Personal-Device Assistant Robustness: task utility under attack89%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Sexually Explicit Content Safety (multi-turn)94.9%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Sexually Explicit Content Safety (single-turn)94.4%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Turn injections in tool use: survival at 1 attempt (static)68%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Turn injections in tool use: survival at 100 attempts51%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Turn injections in tool use: survival at 150 attempts36%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Computer Use: Manual Red Teaming Safety Score100%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Computer Use: Safe Action Rate (prohibited actions)98.9%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Prompt injection robustness, general tool calls: survival at 100 attempts66%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Prompt injection robustness, general tool calls: survival at 150 attempts60%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-
Prompt injection robustness, general tool calls: survival at 50 attempts78%Google, Gemini 4 Argon Safety and Security Reportchecked 2026-10-19-

Preference and Composite Rankings

TestValuePrinted bySetup as printed
Artificial Analysis Intelligence Index v4.1.158.7OpenAI, GPT-6 Astrachecked 2026-10-14-

Price

API list prices in US dollars per one million tokens, as printed on the developer's pricing page on the day checked. Cost per task is measured by Vals AI.

PriceValuePrinted bySetup as printed
Input$0.75Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Cached input$0.075Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Output$3.75Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Input, batch$0.375Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Cached input, batch$0.0375Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Output, batch$1.875Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Input, flex$0.375Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Cached input, flex$0.0375Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Output, flex$1.875Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Input, priority$1.35Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Cached input, priority$0.135Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Output, priority$6.75Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Input, contributor$0.375Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Cached input, contributor$0.0375Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Output, contributor$1.875Google, Gemini API pricingchecked 2026-10-15Offer: promotional price, through 2026-12-31
Cost per taskVals AI, USD, lower is better$6.87Vals AI, Vibe Code Bench v1.1 leaderboardchecked 2026-10-27Harness: OpenHands

More from Google

Every number here is read from the same catalog file as the benchmark table, data version 2026103001. Where one page printed the model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.