text model comparison
GPT-6 Luna vs Kimi K3
GPT-6 Luna and Kimi K3 are both current text models, from OpenAI and Moonshot Kimi. GPT-6 Luna is 97% cheaper per 1M input tokens than Kimi K3, $0.10 vs $3. On Arena Text, GPT-6 Luna scores 1441 Elo and Kimi K3 1488 Elo (Arena, measured October 28, 2026). Both accept about 1.05M tokens of context (1,050,000 and 1,048,576).
- Higher Arena Text
- Kimi K31488 Elo against 1441 Elo
- Lower price per 1M input tokens
- GPT-6 Luna$0.10 against $3
- Longer context window
- GPT-6 Luna1.05M tokens against 1.05M tokens
- Newer release
- GPT-6 LunaSep 2026 against Jul 2026
Data verified November 3, 2026 · Dataset 2026-11-06
Side by side
GPT-6 Luna and Kimi K3, line by line
Prices in effect today, compared at the same tier wherever both list one. The better value in each row is highlighted.
| Fact | GPT-6 Luna | Kimi K3 |
|---|---|---|
| Price | ||
| Price per 1M input tokens | $0.10 (better) | $3 |
| Price per 1M output tokens | $0.50 (better) | $15 |
| Independent benchmarks | ||
| Arena Text | 1441 Elo (#38 of 52) | 1488 Elo (#10 of 52) (better) |
| Arena WebDev | 1579 Elo (#29 of 46) | 1656 Elo (#14 of 46) (better) |
| GPQA Diamond | 90.5% (#26 of 41) | 93.1% (#17 of 41) (better) |
| FrontierMath Tiers 1-3 (v2) | 78.9% (#17 of 41) (better) | 72.2% (#22 of 41) |
| SimpleQA Verified | 41.4% (#28 of 38) | 50.6% (#20 of 38) (better) |
| Specs | ||
| Provider | OpenAI | Moonshot Kimi |
| Release date | September 22, 2026 | July 16, 2026 |
| Access | API only | API and open weights |
| Context window | 1,050,000 tokens | 1,048,576 tokens |
| Max output | 128,000 tokens | Not reported |
| Input | text, image | text, image |
| Output | text | text |
Provider results
What the labs report
Results tables a lab published with one of its models, where it lists both GPT-6 Luna and Kimi K3. These are the lab's own figures, not independent measurements, and each table is read on its own: labs run benchmarks with different harnesses and budgets.
Thinking Machines Lab, Inkling 2 Small model card, November 5, 2026Thinking Machines Lab · 2026-11-06
| Benchmark | GPT-6 Luna | Kimi K3 |
|---|---|---|
| DeepSWE v1.1 | 66.6% | 68.5% (better) |
| Vibe Code Bench v1.1 | 81.7% | 85.0% (better) |
| SWE-Marathon v1.1 | 28.1% | 48.1% (better) |
| SWE-bench Multilingual | 78.2% | 91.4% (better) |
| SWE-bench Multimodal v2 | 37.5% | 44.2% (better) |
| SWE-bench Verified | 76.1% | 93.4% (better) |
| BioMysteryBench | 61.5% | 72.6% (better) |
| MCP Atlaspublic set | 77.8% | 84.2% (better) |
| Finance Agent v2.1 | 56.8% | 60.0% (better) |
| Excel Modeling Benchmark v1.1 | 70.0% | 71.3% (better) |
| Tax Agent Bench | 58.9% | 68.7% (better) |
| APEX-Accounting v1.1pass@1 | 10.8% (better) | 9.5% |
| APEX-Accounting v1.1mean score | 53.9% | 56.7% (better) |
| Harvey LAB v1.1 | 4.2% | 12.9% (better) |
| Legal Research Bench | 30.3% | 46.2% (better) |
| PerceptionBench | 56.2% | 58.5% (better) |
| ScreenSpot-Pro | 81.1% (better) | 73.8% |
| ScreenSpot-Prowith Python | 87.6% (better) | 83.2% |
| CharXiv (RQ) | 81.8% | 84.8% (better) |
| CharXiv (RQ)with Python | 85.3% | 91.3% (better) |
| Chartography | 32.0% (better) | 26.6% |
| SAGE v1.1 | 83.4% | 86.8% (better) |
| LongBench v2 | 65.0% | 69.0% (better) |
| ComplexConstraints v1.0 | 27.3% | 37.9% (better) |
| CyberBench v1.1 | 76.3% (better) | 75.2% |
Thinking Machines' table from the Inkling 2 Small model card, run at effort 0.99 (xhigh) with a 1M-token trajectory limit; where a benchmark's owner publishes scores (Datacurve, Scale AI, Surge, Mercor, Thoughtful, Vals), Thinking Machines uses them. DeepSeek-V4.1-Flash is the 1017 build. Left out here: the Artificial Analysis and Vals indexes, the rows Artificial Analysis ran (FrontierCode, Terminal-Bench, GDP.pdf, MMMU-Pro, HLE, CritPt, AA-LCR, AA-Omniscience, AutomationBench, EnterpriseOps-Gym, Tau 3 Banking), Arena and DesignArena ratings, PMPP-Hard (not a percentage), the safety rows, and JobBench, DAYJOB and Humanity's Sixth Sense, which we could not confirm are public.
Charts
Against every current text model
Where GPT-6 Luna and Kimi K3 land among the current models of their task, on every independent benchmark, on price and on context.
Where each stands
On Arena Text, GPT-6 Luna ranks #38 and Kimi K3 #10 of 52 current models. Each bar is the share of current models it beats: on score, price and context.
Independent benchmarks
Arena Text
GPT-6 Luna:#38of 52
1,441 Elo · ahead of 25%
Kimi K3:#10of 52
1,488 Elo · ahead of 82%
Arena WebDev
GPT-6 Luna:#29of 46
1,579 Elo · ahead of 38%
Kimi K3:#14of 46
1,656 Elo · ahead of 69%
GPQA Diamond
GPT-6 Luna:#26of 41
90.5% · ahead of 35%
Kimi K3:#17of 41
93.1% · ahead of 60%
FrontierMath Tiers 1-3 (v2)
GPT-6 Luna:#17of 41
78.9% · ahead of 60%
Kimi K3:#22of 41
72.2% · ahead of 48%
SimpleQA Verified
GPT-6 Luna:#28of 38
41.4% · ahead of 27%
Kimi K3:#20of 38
50.6% · ahead of 49%
Price and context
Price per 1M input tokens
GPT-6 Luna:#1of 58
$0.10 · cheaper than 96%
Kimi K3:#45of 58
$3 · cheaper than 23%
Price per 1M output tokens
GPT-6 Luna:#3of 58
$0.50 · cheaper than 93%
Kimi K3:#44of 58
$15 · cheaper than 23%
Context window
GPT-6 Luna:#1of 67
1.05M tokens · longer than 85%
Kimi K3:#12of 67
1.05M tokens · longer than 67%
Quality against price
GPT-6 Luna scores 1,441 at $0.2; Kimi K3 scores 1,488 at $6 (blended price per 1M tokens). The line joins the current models no other beats on both price and Arena Text score; hover any point to name it.
Quality against price: the data
| Model | Provider | Blended price per 1M tokens | Arena Text (Elo) |
|---|---|---|---|
| Gemini 4 Carbon | $4 | 1,528 | |
| Gemini 4 Argon | $4 | 1,520 | |
| Claude Opus 5.5 | Anthropic | $8 | 1,509 |
| Muse Wonder | Meta | $8 | 1,507 |
| Claude Fable 5 | Anthropic | $20 | 1,504 |
| Claude Fable 5.1 | Anthropic | $20 | 1,504 |
| Gemini 3.8 Flash | $1.50 | 1,496 | |
| Muse Spark 1.3 | Meta | $2 | 1,495 |
| Claude Opus 5 | Anthropic | $10 | 1,492 |
| Kimi K3 | Moonshot Kimi | $6 | 1,488 |
| GPT-5.6 Sol | OpenAI | $8 | 1,485 |
| Gemini 3.6 Flash | $3 | 1,483 | |
| GPT-5.5 | OpenAI | $11 | 1,482 |
| Claude Opus 4.8 | Anthropic | $10 | 1,482 |
| Qwen3.8-Max | Alibaba Qwen | $3 | 1,482 |
| GPT-6.1 Sol | OpenAI | $4 | 1,481 |
| GLM-5.3 | Z.ai | $2.15 | 1,481 |
| MiMo-V2.6-Pro | Xiaomi MiMo | $0.54 | 1,480 |
| Claude Sonnet 5.5 | Anthropic | $4 | 1,477 |
| GLM-5.3-Flash | Z.ai | $0.24 | 1,477 |
| GPT-6 Astra | OpenAI | $20 | 1,476 |
| DeepSeek-V4.1-Flash | DeepSeek | $0.52 | 1,476 |
| GPT-5.4 | OpenAI | $5.63 | 1,475 |
| Gemini 3 Flash Preview | $1.13 | 1,473 | |
| Grok 4.20 | xAI | $1.56 | 1,472 |
| GPT-5.6 Terra | OpenAI | $4.50 | 1,467 |
| Grok 4.5 | xAI | $3 | 1,465 |
| DeepSeek-V4-Pro | DeepSeek | $1.98 | 1,465 |
| Claude Sonnet 5 | Anthropic | $4 | 1,462 |
| GPT-6 Sol | OpenAI | $4 | 1,458 |
| Gemini 3.5 Flash-Lite | $0.85 | 1,456 | |
| Grok 4.6 | xAI | $3 | 1,455 |
| GPT-5.6 Luna | OpenAI | $0.45 | 1,453 |
| MiMo-V2.6-Flash | Xiaomi MiMo | $0.18 | 1,448 |
| Step 5 Preview | StepFun | $1.43 | 1,448 |
| Grok 4.3 | xAI | $1.56 | 1,442 |
| GPT-6 Luna | OpenAI | $0.20 | 1,441 |
| Grok 4.7 | xAI | $3 | 1,441 |
| Mistral Large 4 | Mistral AI | $1.05 | 1,432 |
| Claude Haiku 5.5 | Anthropic | $0.20 | 1,425 |
| GPT-4.1 | OpenAI | $3.50 | 1,415 |
| Solar Pro 4 | Upstage | $0.52 | 1,388 |
| GPT-4.1 mini | OpenAI | $0.70 | 1,383 |
| GPT-4.1 nano | OpenAI | $0.18 | 1,322 |
Cost
What each costs
Example workloads at each provider's listed rates, then every other rate either one lists. The cheaper figure in each row is highlighted.
The same work, priced
For 10M input and 2M output tokens, Kimi K3 costs 30 times as much as GPT-6 Luna, at each provider's listed rates.
10M input and 2M output tokens
Kimi K3 costs 30× as much
- GPT-6 Luna
- $2
- Kimi K3
- $60
The same, with 9M of the input tokens cached
Kimi K3 costs 30× as much
- GPT-6 Luna
- $1.19
- Kimi K3
- $35.70
| Workload or rate | GPT-6 Luna | Kimi K3 |
|---|---|---|
| Example workloads | ||
| 10M input and 2M output tokens | $2 (better) | $60 |
| The same, with 9M of the input tokens cached | $1.19 (better) | $35.70 |
| Other listed rates | ||
| 1M input tokens, cached | $0.01 (better) | $0.30 |
| 1M input tokens, prompts over 272K tokens | $0.20 | Not listed |
| 1M output tokens, prompts over 272K tokens | $0.75 | Not listed |
Your own bill will differ. These figures use list prices and example workloads. The AI Model Cost Audit reviews your real usage, shortlists models that score as well for each task for less, and estimates the savings.
Launch price $100, regular price $150
Verdict
When to choose each
GPT-6 Luna
Where it leads
- Lower input and output prices, $0.10 and $0.50 vs $3 and $15 per 1M tokens
- Higher FrontierMath Tiers 1-3 (v2) score, 78.9% vs 72.2%
- Cheaper cached input, $0.01 vs $0.30 per 1M tokens
Best for
- High-volume summarisation, extraction and routing
- Long-context budget work, where the 272k threshold is more forgiving than Haiku 5.5's 100k
Watch out for
- Prompts over 272,000 tokens are rebilled at $0.20 input and $0.75 output for the whole request
- A twentieth of GPT-6 Sol's price reflects a capability gap, not a discount on the same model
Kimi K3
Where it leads
- Higher Arena Text score, 1488 Elo vs 1441 Elo
- Higher Arena WebDev score, 1656 Elo vs 1579 Elo
- Higher GPQA Diamond score, 93.1% vs 90.5%
- Higher SimpleQA Verified score, 50.6% vs 41.4%
- Open weights you can run yourself
Best for
- Teams that want a frontier-scale model they can also self-host, with one architecture across both paths
- Long-context work at mid-tier hosted pricing
Watch out for
- Hosted rates match Claude Sonnet 5 exactly at $3 / $15 while offering a larger context window, so the comparison turns on capability rather than price
- Running the open weights shifts cost to infrastructure, which the token rates recorded here do not represent
- Moonshot’s self-reported results place it ahead of Claude Opus 4.8 and GPT-5.5 but behind Claude Fable 5 and GPT-5.6 Sol; none of this was reproduced independently here
From benchmark to production
Building on GPT-6 Luna or Kimi K3?
Public scores rarely hold on your own data. We test both models on your real tasks, then build the agent or pipeline around the one that wins.
AI Model Cost AuditLaunch price $100, regular price $150What the audit includes
FAQ
GPT-6 Luna vs Kimi K3 FAQ
Which is cheaper, GPT-6 Luna or Kimi K3?
GPT-6 Luna is 97% cheaper per 1M input tokens than Kimi K3, $0.10 vs $3. For 10M input and 2M output tokens, GPT-6 Luna costs $2 and Kimi K3 $60 at standard rates.
Which is cheaper with prompt caching, GPT-6 Luna or Kimi K3?
GPT-6 Luna charges $0.01 per 1M cached input tokens, 90% below its standard input rate, and Kimi K3 $0.30, 90% below. With 9M of 10M input tokens cached and 2M output tokens, GPT-6 Luna costs $1.19 and Kimi K3 $35.70.
Which scores higher on independent benchmarks, GPT-6 Luna or Kimi K3?
Kimi K3 leads on Arena Text, 1488 Elo to 1441 Elo; Kimi K3 leads on Arena WebDev, 1656 Elo to 1579 Elo; Kimi K3 leads on GPQA Diamond, 93.1% to 90.5%. Scores are independent evaluations, not either vendor's launch figures.
How do GPT-6 Luna and Kimi K3 compare in Thinking Machines Lab's own results?
In the Inkling 2 Small model card (November 5, 2026), Kimi K3 leads on DeepSWE v1.1, 68.5% to 66.6%; Kimi K3 leads on Vibe Code Bench v1.1, 85.0% to 81.7%; Kimi K3 leads on SWE-Marathon v1.1, 48.1% to 28.1%. These are Thinking Machines Lab's figures, not independent measurements.
Which has the larger context window, GPT-6 Luna or Kimi K3?
Both accept about 1.05M tokens of context (1,050,000 and 1,048,576).
Can I run GPT-6 Luna or Kimi K3 myself?
Kimi K3 is released as open weights you can run yourself; GPT-6 Luna is available through its provider's API only.
Further reading
More on GPT-6 Luna and Kimi K3
Articles on GPT-6 Luna and Kimi K3
- OpenAI Models on Google Cloud: What Changes for AI TeamsOctober 22, 2026
- Claude Sonnet 5.5 vs GPT-6 Luna: Cost, Speed and AccuracySeptember 29, 2026
- GPT-6 Sol vs Claude Opus 5.5, Plus Luna: Best Value per TaskSeptember 24, 2026
- GPT-6 Sol and Luna: Half the Price, and Where Each One FitsSeptember 23, 2026
- Qwen vs DeepSeek vs Kimi for Agents and CodingAugust 6, 2026
- Kimi K3 vs Opus 5 vs GPT-5.6 Sol: which model should you use?July 31, 2026
Related comparisons
- Claude Haiku 5.5 vs GPT-6 Luna
- Claude Opus 5.5 vs Kimi K3
- GPT-6.1 Sol vs GPT-6 Luna
- DeepSeek-V4.1-Flash vs Kimi K3
- Claude Sonnet 5.5 vs GPT-6 Luna
- GPT-6 Astra vs Kimi K3
- Claude Opus 5.5 vs GPT-6 Luna
- Gemini 3.8 Flash vs Kimi K3
- DeepSeek-V4.1-Flash vs GPT-6 Luna
- GLM-5.3 vs Kimi K3
- GPT-5.4 vs GPT-6 Luna
- DeepSeek-V4-Pro vs Kimi K3
Production evaluation
Choosing between GPT-6 Luna and Kimi K3?
Public benchmarks narrow the shortlist. We run both on your own tasks, data and cost limits, then build the routing and review controls to run the winner in production.
- A head-to-head evaluation on your real workload
- Routing between models by cost and difficulty
