Model comparison
Kimi K3 vs Microsoft MAI-Thinking-1
Moonshot AI against Microsoft, compared on context, price, and verified benchmark results.
Catalog record checked September 5, 2026Individual provider fields may changeEvidence confidence: High — see receipts below
Moonshot AI
Kimi K3
Frontier · Open weights
Microsoft
Microsoft MAI-Thinking-1
Frontier
| Specification | Kimi K3 | Microsoft MAI-Thinking-1 |
|---|---|---|
| Provider | Moonshot AI | Microsoft |
| Tier | Frontier | Frontier |
| Context window | Winner: 1.05M | 256K |
| Max output | Winner: 128K | 64K |
| Input / 1M tokens | $3 | Winner: $2 |
| Output / 1M tokens | $15 | Winner: $8 |
| Weights | Open | Closed |
| Parameters | 2.8T (open-weight MoE) | ~1T total / 35B active (sparse MoE) |
| Reasoning levels | low, high, max | low, medium, high |
| Modalities | text, image, video | text |
| API model id | kimi-k3 | mai-thinking-1 |
| Released | July 16, 2026 | June 2, 2026 |
| Artificial Analysis Intelligence Index [max] (2026-08-14) | 60 | Not verifiedUnverified |
| Artificial Analysis Intelligence Index (2026-08-14) | Not verifiedUnverified | 55 |
| Terminal-Bench 2.1 (2026-07-27) | 88.3 | Not verifiedUnverified |
| GPQA Diamond (2026-07-27) | Winner: 93.5 | 84.2 |
| Humanity's Last Exam (2026-07-27) | 56 | Not verifiedUnverified |
| SWE-bench Verified (2026-09-01) | Winner: 93.4 | 73.5 |
| AIME 2025 (2026-08-12) | Not verifiedUnverified | 97 |
| Terminal-Bench 2.0 (2026-06) | Not verifiedUnverified | 46 |
Prices are USD per million tokens at standard rates, excluding batch and caching discounts. Bold indicates the better figure where one is objectively better. Values we could not confirm from the provider are shown as “Not verified” rather than estimated.
Benchmark receipts
- 2026-09-01: vals.ai SWE-bench Verified leaderboard (independent, mini-swe-agent bash-only); not published by Moonshot; a 76.8 aggregator claim is low-reliability — both notedReal GitHub issue resolution: does the model's patch pass the hidden tests. Comparability: comparable with caveat — Post-audit vendor claims and pre-audit scores sit on different task trust levels; scaffolding (agent harness, compute budget) also dominates results. Never aggregate across scaffolds.
- 2026-08-14: Artificial AnalysisComposite index blending reasoning, knowledge, and coding evals into one 0–100 score. Comparability: directly comparable — AA occasionally rebaselines the index scale between snapshots — a score captured on one date is only comparable to same-snapshot scores (check measuredAt).
- 2026-08-12: Microsoft AI — Introducing MAI-Thinking-1 (vendor, 256k output)Competition mathematics: AIME 2025 exam problems, typically pass@1 with tools disallowed or single-attempt code execution per the harness. Comparability: comparable with caveat — 30-question total denominator: one problem = ~3.3 points. Tool-use policy (calculator/code interpreter) must match between compared models.
- 2026-07-27: Kimi K3 HF model card (vendor, with tools; 43.5 no tools)Agentic terminal work: multi-step tasks executed in a sandboxed shell environment. Comparability: comparable with caveat — Scores across Terminal-Bench major versions (v1 vs v2) are NOT comparable; harness configuration (container, time limits) also shifts results.
- 2026-06: Microsoft AI model page (vendor, as labeled: Terminal-Bench 2.0 — NOT comparable to 2.1 observations)Real GitHub issue resolution: does the model's patch pass the hidden tests. Comparability: comparable with caveat — Post-audit vendor claims and pre-audit scores sit on different task trust levels; scaffolding (agent harness, compute budget) also dominates results. Never aggregate across scaffolds.
Pricing tiers: Kimi K3: $3/$15 per MTok; cached input $0.30/MTok. Open-weight under the custom Kimi K3 License (not a standard open-source licence). 2.8T MoE with native vision. · Microsoft MAI-Thinking-1: Global Standard $2.00/$8.00 per MTok, cached input $0.20 (Azure Retail Prices API, effective 2026-08-01, uniform across ~22 public regions; US Gov regions carry premiums). Still preview; Global Standard only, no PTU yet. 256K context with a 64K output cap that includes reasoning tokens. ~1T/35B sparse MoE; reasoning is always on (no effort levels). Microsoft's first true reasoning model, trained without OpenAI distillation.
Kimi K3
Kimi K3 is Moonshot AI's open-weight frontier model — 2.8T MoE, 1M context, tied near the top of open-weight leaderboards.
Best for
- Open-weight frontier work
- Long-context
- Multimodal
Watch out
Self-hosting needs large clusters; hosted rates vary.
Microsoft MAI-Thinking-1
MAI-Thinking-1 is Microsoft's first in-house reasoning model, reducing dependence on OpenAI.
Best for
- Enterprise (Azure)
- Reasoning
- Microsoft ecosystem
Watch out
Foundry-gated preview; no MAI-2 exists yet. Of the Build 2026 MAI wave, only MAI-Thinking-1 and MAI-Cyber-1-Flash ($0.60/$3.50, released 2026-07-27) have published token meters.
When the cheaper one wins
Microsoft MAI-Thinking-1 is cheaper on output at $8 per million tokens against $15 for Kimi K3 — about 1.9×. Use the cheaper tier for classification, extraction, summarisation, and any task where the expensive model’s extra score does not change the accepted output. The expensive one only pays if your hardest task actually fails on the cheap tier. These are standard-tier API rates, excluding batch and cache discounts.
Run the model pickerEvidence confidence: High
How strong and complete the evidence behind this comparison is — not a prediction of which model is better.
- Pricing verified on both sides — Input and output rates are verified for both models.
- 5/5 core specs verified on both sides — All core specifications verified for both models.
- 3 shared named benchmarks with differing scores — Measured on: Artificial Analysis Intelligence Index, GPQA Diamond, SWE-bench Verified.
- Verified within the last 90 days — Newest catalog check was 6 days ago.
- Both models carry source citations — Each side has at least two catalog sources on record.
Source receipts
Each catalog figure was checked against the provider or an independent second source on the date shown.
- Kimi K3: HuggingFace — Kimi K3 (accessed 2026-08-29)
- Kimi K3: Moonshot — Kimi K3 guide (accessed 2026-08-29)
- Microsoft MAI-Thinking-1: Azure Retail Prices API (MAI-Thinking-1 $2/$0.20 cached/$8) (accessed 2026-09-04)
- Microsoft MAI-Thinking-1: AzureSpeed — AI model pricing tracker (same figures) (accessed 2026-09-04)
- Microsoft MAI-Thinking-1: Microsoft — Introducing MAI-Thinking-1 (~1T total / 35B active MoE) (accessed 2026-09-05)
- Microsoft MAI-Thinking-1: Microsoft Foundry — MAI-Thinking-1 usage (256K context, 64K output cap) (accessed 2026-09-05)
Related comparisons
- Claude Fable 5.1 vs Kimi K3
- Claude Fable 5.1 vs Microsoft MAI-Thinking-1
- Claude Fable 5 vs Kimi K3
- Claude Fable 5 vs Microsoft MAI-Thinking-1
- Claude Mythos 5.1 vs Kimi K3
- Claude Mythos 5.1 vs Microsoft MAI-Thinking-1
Diving deeper on one model? Kimi K3 · Microsoft MAI-Thinking-1
Common questions
Kimi K3 vs Microsoft MAI-Thinking-1
Answered from the verified figures on this page rather than general guidance.
Is Kimi K3 or Microsoft MAI-Thinking-1 cheaper for input?
Microsoft MAI-Thinking-1 is cheaper at $2 per million input tokens, against $3 for Kimi K3 — roughly 1.5× the price. Output tokens usually dominate a real bill, so weigh the output rate more heavily than the input rate. These are base rates; Kimi K3 has tiered pricing: $3/$15 per MTok; cached input $0.30/MTok. Open-weight under the custom Kimi K3 License (not a standard open-source licence). 2.8T MoE with native vision. Microsoft MAI-Thinking-1 has tiered pricing: Global Standard $2.00/$8.00 per MTok, cached input $0.20 (Azure Retail Prices API, effective 2026-08-01, uniform across ~22 public regions; US Gov regions carry premiums). Still preview; Global Standard only, no PTU yet. 256K context with a 64K output cap that includes reasoning tokens. ~1T/35B sparse MoE; reasoning is always on (no effort levels). Microsoft's first true reasoning model, trained without OpenAI distillation.
Is Kimi K3 or Microsoft MAI-Thinking-1 cheaper for output?
Microsoft MAI-Thinking-1 is cheaper at $8 per million output tokens, against $15 for Kimi K3 — roughly 1.9× the price. Output tokens usually dominate a real bill, so weigh the output rate more heavily than the input rate. These are base rates; Kimi K3 has tiered pricing: $3/$15 per MTok; cached input $0.30/MTok. Open-weight under the custom Kimi K3 License (not a standard open-source licence). 2.8T MoE with native vision. Microsoft MAI-Thinking-1 has tiered pricing: Global Standard $2.00/$8.00 per MTok, cached input $0.20 (Azure Retail Prices API, effective 2026-08-01, uniform across ~22 public regions; US Gov regions carry premiums). Still preview; Global Standard only, no PTU yet. 256K context with a 64K output cap that includes reasoning tokens. ~1T/35B sparse MoE; reasoning is always on (no effort levels). Microsoft's first true reasoning model, trained without OpenAI distillation.
Which has the larger context window, Kimi K3 or Microsoft MAI-Thinking-1?
Kimi K3 accepts 1.05M tokens against 256K for Microsoft MAI-Thinking-1. This only matters if you routinely send very long documents or large codebases.
Do Kimi K3 and Microsoft MAI-Thinking-1 support the same reasoning levels?
Kimi K3 exposes low, high, max, while Microsoft MAI-Thinking-1 exposes low, medium, high.
Should I use Kimi K3 or Microsoft MAI-Thinking-1?
Both sit in the frontier tier, so the choice usually comes down to price and context rather than capability. Kimi K3 suits open-weight frontier work; Microsoft MAI-Thinking-1 suits enterprise (azure).
Can I self-host Kimi K3 or Microsoft MAI-Thinking-1?
Kimi K3 publishes open weights, but self-hosting and third-party provider access depend on the licence, hardware, serving support, and availability. Microsoft MAI-Thinking-1 is a closed model whose supported access paths are controlled by its provider.
Next step
Choosing between them
The verified figures that separate this pair, computed from the catalog rather than restated boilerplate.
- Input price: Microsoft MAI-Thinking-1 costs $2 per 1M tokens versus $3 for Kimi K3 — a 1.5x difference at the headline tier.
- Context: Kimi K3 takes 1.05M against 256K for Microsoft MAI-Thinking-1 — only decisive if your prompts approach the smaller window.
- Measured capability: Kimi K3 leads Artificial Analysis Intelligence Index 60 to 55 (measured 2026-08-14).
- Deployment: Kimi K3 publishes weights you can self-host; the other is API-only.
Choosing a harness rather than a model — Cursor, Copilot, Claude Code, Windsurf, or Lovable? Compare agentic harnesses · Latest releases.