Compare / Gemini 3.8 Flash vs GPT-6 Luna

Gemini 3.8 Flash vs GPT-6 Luna

Which is cheaper, which is more capable, and which does better at each kind of work, from 109 results on 10 sources that measured both models.

GoogleGemini 3.8 Flash Released 2 Sep 2026 · $1.50 in · $7.50 out
OpenAIGPT-6 Luna Released 22 Sep 2026 · $0.10 in · $0.50 out
Gemini 3.8 Flash wins 3719 too close to callGPT-6 Luna wins 53

Of the 109 results both models have, the better value on each. Results whose reported ranges overlap are too close to call.

Results as of 16 November 2026 (catalogue release 2026-11-16-49bdfa045dee). See also Google vs OpenAI, every model from both labs.

Gemini 3.8 Flash or GPT-6 Luna? The short answer

Pick Gemini 3.8 Flash if quality matters most: it scores higher (Spring Prompt overall 58 against 39). Pick GPT-6 Luna if cost matters more: it costs 22 times less per product listing on CatalogBench.

  • Spring Prompt overall: Gemini 3.8 Flash 58, GPT-6 Luna 39 (out of 100; how the score works).
  • GPT-6 Luna is cheaper to run: $0.000 against $0.010 per product listing on CatalogBench (22 times less).
  • Gemini 3.8 Flash scores higher on the Artificial Analysis Intelligence Index: 40.9 against 38.1.
  • GPT-6 Luna is newer: released 22 Sep 2026, against 2 Sep 2026 for Gemini 3.8 Flash.
  • On our own benchmarks, on the results both models have: Gemini 3.8 Flash does better for speed under pressure; GPT-6 Luna for product listings, decks from an analysis, user surveys and marketing planning.
  • On third-party benchmarks, on the results both models have: Gemini 3.8 Flash does better for agents and tool use, professional work, reasoning and knowledge, writing and sticking to the facts.

Head to head

58
Spring Prompt overall
39
40.9
Intelligence Index
38.1
$3.00
Price per million tokens, blendedlower is better
$0.20
141
Output speed, tokens a second
127
1,048,576
Context window, tokens
1,050,000

Every fact side by side

Gemini 3.8 FlashGPT-6 Luna
Price per million tokens$1.50 in · $7.50 out$0.10 in · $0.50 out
Blended (3 in : 1 out)$3.00$0.20
Cheapest host, blended$1.50 (Google)$0.20 (Amazon Bedrock)
Cost per product listing (CatalogBench)$0.010$0.000
Cached input, per million–$0.010 cached input
Intelligence Index40.938.1
Spring Prompt overall5839
Output speed, tokens a second141127
Context window1,048,576 tokens1,050,000 tokens
Released2 Sep 202622 Sep 2026 (newer)
WeightsClosedClosed
Developer based inthe USthe US

Prices are the developer's list price where we have read it, otherwise the typical host on OpenRouter; see every model's price. Intelligence and speed from Artificial Analysis.

Which is better for what

Each model's rank among the models on sale today, for each kind of work, from our own benchmarks and licensed sources. Further right is better; the better rank is in green.

Product listingsTurning a sparse product feed and photos into listings that can go live 22 of 32 7 of 32
Decks from an analysisTurning a finished analysis into a deck you could present as it is 26 of 32 10 of 32
User surveysPlanning a user survey and reading its results without being misled 16 of 23 12 of 23
Marketing planningPlanning a year of ad spend without overspending 11 of 32 6 of 32
Agents and tool useMulti-step tasks with tools: support desks, coding agents, function calls 35 of 159 53 of 159
Professional workReal tasks from banking, consulting and law, business documents and freelance projects 11 of 51 25 of 51
Reasoning and knowledgeHard questions across science, maths and general knowledge 15 of 187 48 of 187
WritingWhat people prefer in blind comparisons, and judged writing quality 3 of 169 61 of 169
Sticking to the factsSummarising without inventing things, and factual answers 12 of 139 89 of 139
Speed under pressureGood decisions against a real clock (fast chess) 13 of 35 24 of 35

Their slides, side by side

The same brief for both: Manchester store investment review for Hearthside Coffee, an invented company. First slide of each deck from DeckBench.

First slide of Gemini 3.8 Flash's deck for Hearthside Coffee
Gemini 3.8 Flash · deck rating 788 · needs work
First slide of GPT-6 Luna's deck for Hearthside Coffee
GPT-6 Luna · deck rating 1,267 · needs work

Every slide, side by side →

From our research

Every shared result

Each model's best result on every measure both have, grouped by benchmark, with the setting that produced it. Under each measure, every model that benchmark has measured is a grey tick, best to the right, so you can see where the two sit in the field. The headline measures of each benchmark come first; the rest are under "Show all". Green marks the better value; where a source reports ranges that overlap, neither is marked.

MeasureGemini 3.8 FlashGPT-6 Luna
BulletBench measured by Spring Prompt · 3 results
Cost per game · Blitz 3+2US dollars, lower is better 11 · 1 of 20 measured $0.0795 $0.0072
Cost per game · Bullet 60sUS dollars, lower is better 43 · 21 of 59 measured $0.0154 $0.0016no reasoning
Ladder Elo · Bullet 60sladder Elo, higher is better 1 · 49 of 59 measured 1,068low reasoning 249no reasoning
CatalogBench measured by Spring Prompt · 3 results
Cost per productUS dollars, lower is better 18 · 2 of 33 measured $0.0103 $0.0005
Reliably publish-ready% of products, higher is better 20 · 10 of 32 measured 35.7% 55.4%
Reliably publish-ready · sales brief% of products, higher is better 23 · 5 of 32 measured 3.6% 51.8%
DeckBench measured by Spring Prompt · 3 results
Cost per deckUS dollars, lower is better 15 · 3 of 33 measured $0.0613 $0.0087
Deck ratingrating, higher is betterThe rating weighs each win by how strong the other deck was; the win rate counts wins against whichever decks a model happened to meet, so the two can disagree. 26 · 10 of 33 measured 788 1,267
Accurate decks% of tasks, higher is better 24 · 15 of 33 measured 16.7% 50.0%
ROASBench measured by Spring Prompt · 3 results
Overall scorescore out of 100, higher is better 11 · 6 of 33 measured 51.4 53.4
Audience scorescore out of 100, higher is better 10 · 5 of 33 measured 67.1 68.3
Business scorescore out of 100, higher is better 11 · 6 of 33 measured 51.0 54.8
SurveyBench measured by Spring Prompt · 3 results
Analysis ratingrating, higher is betterThe rating weighs each win by how strong the other analysis was; the win rate counts wins against whichever analyses a model happened to meet, so the two can disagree. 16 · 12 of 24 measured 875 1,093
Cost per taskUS dollars, lower is better 14 · 1 of 24 measured $0.0943 $0.0088
Analysis scorepoints of 100, higher is better 15 · 6 of 24 measured 74.7 90.8
APEX-Agents reported by APEX-Agents · 1 result
Tasks passed% of tasks, higher is better 8 · 27 of 37 measured 64.3% 44.3%max reasoning
Arena (formerly LMArena) reported by Arena (formerly LMArena) · 3 results
Overall · TextArena rating, higher is better 7 · 78 of 387 measured 1,495high reasoning 1,443max reasoning
Overall · Text factualityArena rating, higher is better 16 · 98 of 163 measured 1,477high reasoning 1,430max reasoning
Business, management and finance · TextArena rating, higher is better 21 · 71 of 380 measured 1,481high reasoning 1,447max reasoning
Artificial Analysis reported by Artificial Analysis · 3 results
Artificial Analysis Intelligence Indexindex score, higher is better 25 · 39 of 249 measured 40.9high reasoning 38.1max reasoning
Humanity's Last Exam% of questions, higher is better 18 · 53 of 250 measured 47.8%high reasoning 38.5%max reasoning
Terminal-Bench 4.0% of tasks, higher is better 30 · 37 of 120 measured 19.7%high reasoning 12.6%max reasoning
GDP.pdf reported by GDP.pdf · 1 result
Professional document tasks% of rubric, higher is better 12 · 14 of 32 measured 23.4%medium reasoning 23.0%max reasoning
SimpleQA Verified (Epoch AI) reported by SimpleQA Verified (Epoch AI) · 1 result
Correct answers% of questions, higher is better 7 · 40 of 76 measured 69.7%high reasoning 41.4%max reasoning
Show all 109 results (85 more)
MeasureGemini 3.8 FlashGPT-6 Luna
BulletBench measured by Spring Prompt · 7 results
Games lost on time · Blitz 3+2% of games, lower is better 7 · 13 of 20 measured 12.5% 50.0%
Games lost on time · Bullet 60s% of games, lower is better 39 · 39 of 59 measured 50.0%low reasoning 50.0%no reasoning
Invalid moves · Blitz 3+2% of moves, lower is better 1 · 1 of 20 measured 0.0% 0.0%
Invalid moves · Bullet 60s% of moves, lower is better 1 · 1 of 59 measured 0.0% 0.0%no reasoning
Ladder Elo · Blitz 3+2ladder Elo, higher is better 2 · 11 of 20 measured 1,126 557
Median move time · Blitz 3+2milliseconds, lower is better 9 · 16 of 20 measured 4.0 s 7.0 s
Median move time · Bullet 60smilliseconds, lower is better 46 · 39 of 59 measured 1.8 slow reasoning 1.5 sno reasoning
CatalogBench measured by Spring Prompt · 28 results
Channel compliance% of products, higher is better 25 · 1 of 33 measured 99.4% 100.0%
Channel rules broken% of products, lower is better 20 · 1 of 33 measured 0.6% 0.0%
Channel rules broken · sales brief% of products, lower is better 17 · 1 of 33 measured 0.6% 0.0%
Claims to check% of products, lower is better 1 · 1 of 31 measured 0.0% 0.0%
Claims to check · sales brief% of products, lower is better 26 · 1 of 31 measured 5.4% 0.0%
Conflicts caught% of conflicts, higher is better 1 · 1 of 33 measured 100.0% 100.0%
Content quality% of checks, higher is better 9 · 18 of 33 measured 97.7% 97.0%
Content quality · sales brief% of checks, higher is better 14 · 15 of 33 measured 97.6% 97.4%
Decision accuracy% of decisions, higher is better 19 · 14 of 33 measured 96.4% 97.0%
Failed outputs% of products, lower is better 1 · 1 of 33 measured 0.0% 0.0%
Failed outputs · sales brief% of products, lower is better 1 · 1 of 33 measured 0.0% 0.0%
Field accuracy% of missing fields, higher is better 24 · 15 of 33 measured 92.7% 93.7%
Invented values% of filled values, lower is better 4 · 16 of 33 measured 0.3% 1.7%
Missing UK information% of products, lower is better 1 · 1 of 33 measured 0.0% 0.0%
Missing UK information · sales brief% of products, lower is better 1 · 1 of 33 measured 0.0% 0.0%
Not findable% of products, lower is better 27 · 29 of 33 measured 8.3% 10.1%
Not findable · sales brief% of products, lower is better 28 · 29 of 33 measured 8.9% 10.1%
Publish-ready listings% of products, higher is better 21 · 14 of 33 measured 54.8% 67.9%
Publish-ready listings · sales brief% of products, higher is better 23 · 6 of 33 measured 16.7% 67.9%
Time per productseconds, lower is better 19 · 7 of 33 measured 18.7 8.20
Unsupported claims% of products, lower is better 24 · 10 of 33 measured 31.6% 8.9%
Unsupported claimsclaims per product, lower is better 22 · 10 of 33 measured 0.79 0.31
Unsupported claims · sales brief% of products, lower is better 28 · 6 of 33 measured 81.6% 8.3%
Unsupported claims · sales briefclaims per product, lower is better 24 · 6 of 33 measured 3.08 0.35
Wrong attributes% of products, lower is better 11 · 21 of 33 measured 15.5% 18.5%
Wrong attributes · sales brief% of products, lower is better 2 · 20 of 33 measured 11.9% 19.1%
Wrong category or variant% of products, lower is better 27 · 6 of 33 measured 3.6% 1.8%
Wrong category or variant · sales brief% of products, lower is better 23 · 1 of 33 measured 3.0% 0.0%
DeckBench measured by Spring Prompt · 14 results
Caveat dropped% of tasks, lower is better 1 · 1 of 33 measured 0.0% 0.0%
Clean layout% of tasks, higher is better 24 · 8 of 33 measured 16.7% 66.7%
Design quality% of the maximum, higher is better 19 · 8 of 33 measured 61.7% 78.1%
Draft figure quoted% of tasks, lower is better 26 · 19 of 33 measured 50.0% 33.3%
Findings missing% of tasks, lower is better 1 · 1 of 33 measured 0.0% 0.0%
Head-to-head win rate% of comparisons, higher is betterThe rating weighs each win by how strong the other deck was; the win rate counts wins against whichever decks a model happened to meet, so the two can disagree. 26 · 10 of 33 measured 38.7% 70.9%
Layout defects% of tasks, lower is better 24 · 8 of 33 measured 83.3% 33.3%
Misleading metric used% of tasks, lower is better 1 · 1 of 33 measured 0.0% 0.0%
Presentable decks% of tasks, higher is better 15 · 15 of 33 measured 0.0% 0.0%
Recommendation late or wrong% of tasks, lower is better 20 · 15 of 33 measured 33.3% 16.7%
Slides needing work% of tasks, lower is better 10 · 3 of 33 measured 33.3% 16.7%
Time per deckseconds, lower is better 13 · 15 of 33 measured 93.0 101
Unsupported claims% of tasks, lower is better 30 · 1 of 33 measured 50.0% 0.0%
Unsupported numbers% of tasks, lower is better 26 · 1 of 33 measured 16.7% 0.0%
ROASBench measured by Spring Prompt · 6 results
Consistency scorescore out of 100, higher is better 6 · 8 of 33 measured 36.4 35.8
Contribution profitsimulated US dollars, higher is better 5 · 8 of 33 measured $842,432 $738,879
Cost of a runUS dollars, lower is better 15 · 2 of 33 measured $0.16 $0.0142
Months over budgetmonths of 12, lower is better 1 · 1 of 33 measured 0 0
Planning scorescore out of 100, higher is better 21 · 18 of 33 measured 54.7 54.8
Return on ad spendprofit per $1 spent, higher is better 12 · 9 of 33 measured $0.65 $0.66
SurveyBench measured by Spring Prompt · 18 results
Head-to-head win rate% of comparisons, higher is betterThe rating weighs each win by how strong the other analysis was; the win rate counts wins against whichever analyses a model happened to meet, so the two can disagree. 6 · 14 of 24 measured 62.5% 48.5%
Leading or double-barrelled questions% of tasks, lower is better 21 · 3 of 24 measured 66.7% 11.1%
Mistaken next steps% of tasks, lower is better 5 · 5 of 24 measured 11.1% 11.1%
Numbers right% of answers, higher is better 1 · 18 of 24 measured 100.0% 98.0%
Objectives unanswerable% of tasks, lower is better 1 · 1 of 24 measured 0.0% 0.0%
Output tokens per tasktokens, lower is better 3 · 2 of 14 measured 21,981 14,333
Precision ignored% of tasks, lower is better 18 · 1 of 24 measured 44.4% 0.0%
Research scorepoints of 100, higher is better 15 · 6 of 24 measured 77.6 91.7
Sound research% of tasks, higher is better 15 · 4 of 24 measured 0.0% 33.3%
Stakeholder's view presumed% of tasks, lower is better 15 · 10 of 24 measured 22.2% 11.1%
Survey plan scorepoints of 100, higher is better 18 · 7 of 24 measured 84.4 94.0
Survey structure% of tasks, lower is better 11 · 11 of 24 measured 22.2% 22.2%
Time per taskseconds, lower is better 8 · 10 of 24 measured 131 138
Traps handled% of traps, higher is better 18 · 8 of 24 measured 78.3% 94.2%
Traps missed% of tasks, lower is better 16 · 6 of 24 measured 88.9% 33.3%
Unsupported findings% of tasks, lower is better 14 · 5 of 24 measured 77.8% 22.2%
Wrong numbers% of tasks, lower is better 1 · 13 of 24 measured 0.0% 11.1%
Wrong recommendation% of tasks, lower is better 1 · 5 of 24 measured 0.0% 11.1%
Arena (formerly LMArena) reported by Arena (formerly LMArena) · 8 results
Confirmed task success · AgentIPS effect estimate, higher is better 11 · 29 of 48 measured 0.07high reasoning, arena agent -0.00max reasoning, arena agent
Creative writing · TextArena rating, higher is better 7 · 77 of 385 measured 1,485high reasoning 1,413max reasoning
Expert prompts · TextArena rating, higher is better 13 · 40 of 338 measured 1,526high reasoning 1,498max reasoning
Instruction following · TextArena rating, higher is better 12 · 57 of 387 measured 1,486high reasoning 1,448max reasoning
Praise over complaint · AgentIPS effect estimate, higher is better 19 · 30 of 48 measured 0.05high reasoning, arena agent -0.02max reasoning, arena agent
Steerability · AgentIPS effect estimate, higher is better 17 · 16 of 48 measured 0.03high reasoning, arena agent 0.04max reasoning, arena agent
Tool grounding · AgentIPS effect estimate, higher is better 17 · 39 of 48 measured 0.00high reasoning, arena agent -0.00max reasoning, arena agent
Writing, literature and language · TextArena rating, higher is better 8 · 79 of 386 measured 1,487high reasoning 1,421max reasoning
Artificial Analysis reported by Artificial Analysis · 4 results
Long-context reasoning (AA-LCR)% of questions, higher is better 7 · 13 of 241 measured 84.0%medium reasoning 83.3%max reasoning
Output speedtokens per second, higher is better 31 · 37 of 78 measured 141high reasoning 127max reasoning
SciCode% of problems, higher is better 20 · 31 of 121 measured 56.6%high reasoning 54.6%max reasoning
Time to first answer tokenseconds, lower is better 58 · 20 of 78 measured 15.5high reasoning 0.74no reasoning

Quick answers

Is Gemini 3.8 Flash better than GPT-6 Luna?

Pick Gemini 3.8 Flash if quality matters most: it scores higher (Spring Prompt overall 58 against 39). Pick GPT-6 Luna if cost matters more: it costs 22 times less per product listing on CatalogBench.

Which is cheaper, Gemini 3.8 Flash or GPT-6 Luna?

GPT-6 Luna is cheaper to run: $0.000 against $0.010 per product listing on CatalogBench (22 times less).

Which is faster, Gemini 3.8 Flash or GPT-6 Luna?

Gemini 3.8 Flash writes about 141 tokens a second on its usual API, against 127 for GPT-6 Luna (Artificial Analysis).

Which is better for product listings, Gemini 3.8 Flash or GPT-6 Luna?

GPT-6 Luna ranks 7 of 32 models on sale for turning a sparse product feed and photos into listings that can go live, against 22 for Gemini 3.8 Flash.

Which is better for decks from an analysis, Gemini 3.8 Flash or GPT-6 Luna?

GPT-6 Luna ranks 10 of 32 models on sale for turning a finished analysis into a deck you could present as it is, against 26 for Gemini 3.8 Flash.

Which is better for user surveys, Gemini 3.8 Flash or GPT-6 Luna?

GPT-6 Luna ranks 12 of 23 models on sale for planning a user survey and reading its results without being misled, against 16 for Gemini 3.8 Flash.

Which is better for marketing planning, Gemini 3.8 Flash or GPT-6 Luna?

GPT-6 Luna ranks 6 of 32 models on sale for planning a year of ad spend without overspending, against 11 for Gemini 3.8 Flash.

Which has the bigger context window, Gemini 3.8 Flash or GPT-6 Luna?

GPT-6 Luna: 1,050,000 tokens, against 1,048,576 for Gemini 3.8 Flash.

Run Gemini 3.8 Flash and GPT-6 Luna on your own prompt.

Benchmarks aren't your task. Test both side by side on your prompt, with every answer checked against your rules and the cost of each one.