Moonshot’s Kimi K3 arrived in July 2026 as the most serious open-weight challenger yet to Anthropic’s flagship Claude Opus 4.8 — and the two models represent opposite philosophies of frontier AI. One is a 2.8-trillion-parameter open mixture-of-experts you will (from July 27) be able to download; the other is a closed, polished agentic workhorse you rent through an API. Here is how they actually compare on specs, pricing, and practical deployment, using data from our live AI models database.
Wichtige Fakten
- Kimi K3: 2.8T total parameters (MoE, 16 of 896 experts active), 1M context, open weights due July 27, 2026
- Claude Opus 4.8: Anthropic’s proprietary flagship, 1M context, up to 128K output tokens
- API pricing: K3 at $3/$15 per million tokens vs Opus 4.8 at $5/$25 — ~40% cheaper blended ($6.00 vs $10.00 on our 3:1 index)
- Self-hosting K3 is theoretical for most: ~1.4 TB of VRAM at 4-bit — a multi-node H200 cluster, not a workstation
- Moonshot’s published benchmarks show K3 edging past Opus 4.8 on several agentic and coding suites — treat vendor numbers as claims until independent benchmarks land
Specs, Side by Side
Head to Head
| Kimi K3 | Claude Opus 4.8 | |
|---|---|---|
| Entwickler | Moonshot AI | Anthropic |
| Parameter | 2.8T total, MoE (16/896 experts active) | Nicht offengelegt |
| Kontextfenster | 1 Mio. Tokens | 1 Mio. Tokens |
| Maximale Ausgabe | — | 128K tokens |
| Lizenz | Open weights (from Jul 27, 2026) | Proprietary API |
| API price (in/out per 1M) | $3 / $15 | $5 / $25 |
| Run it locally? | ~1.4 TB VRAM (4-bit) — multi-node cluster | Not available |
The Pricing Math
On our blended index (3:1 input-to-output, the ratio of typical production traffic), Kimi K3 costs $6.00 per million tokens against $10.00 for Opus 4.8 — a 40% saving at near-frontier capability. Run a real workload through it: at 100M input + 20M output tokens a month, K3 bills $600 while Opus bills $1,000. Over a year that difference funds an entire additional AI project. You can model your own volumes in our KI-API-Kostenrechner.
The context that matters: both models sit at the premium end of our Preis-Leistungs-Index, where efficiency-tier models deliver far more capability per dollar for routine work. The K3-vs-Opus question is really about the top 10% of tasks — long-horizon agents, complex refactors, high-stakes reasoning — where flagship quality pays for itself.
Open Weights Change the Question
From July 27, K3’s weights are downloadable — but “open” does not mean “runnable at home.” At roughly 1.4 TB of VRAM in 4-bit quantization, you need a multi-node cluster (think two 8-GPU H200 nodes) before the first token flows. For almost everyone, K3 will be consumed through APIs just like Opus. The strategic difference is real, though: open weights mean multiple competing providers, price pressure over time, fine-tuning rights, and no single-vendor kill switch — the same dynamics that made DeepSeek the value king of our index.
Convlys Einschätzung
Choose Opus 4.8 when reliability under agentic load is the product — its 128K output ceiling and mature tooling remain the safest bet for production agents. Choose K3 when cost at frontier scale is the constraint, when you need fine-tuning rights, or when vendor independence matters strategically. And if your workload is routine, choose neither: the efficiency tier does 90% of tasks at a tenth of the price. The most expensive mistake in 2026 isn’t picking the wrong flagship — it’s using any flagship for work a $0.20 model handles.
Häufig gestellte Fragen (FAQ)
Ist Kimi K3 besser als Claude Opus 4.8?
Moonshot’s published results show K3 ahead on several agentic and coding benchmarks, but independent verification is still landing. What is certain: K3 costs ~40% less blended ($6 vs $10 per 1M tokens) and its weights go open on July 27, 2026.
Kann ich Kimi K3 lokal ausführen?
Only with data-center hardware: ~1.4 TB of VRAM at 4-bit quantization, i.e. a multi-node GPU cluster such as two 8×H200 nodes. Check any open model’s requirements in our LLM-VRAM-Rechner.
Which is cheaper for production use?
Kimi K3: $3/$15 per million tokens vs $5/$25 for Opus 4.8. At a typical 3:1 input:output mix, that is $6.00 vs $10.00 blended — 40% cheaper for the same volume.
Go Deeper
- Kimi K3 Explained — the full story of Moonshot’s 2.8T open model
- Kimi K3 vs. ChatGPT und Claude — the three-way user’s view
- Live-Datenbank für KI-Modelle · API Cost Calculator · Price-Performance Index
