What TSK-1 Found
TSK-1 hasn't run these two head-to-head in the same test. We tested Kimi on Jul 31, 2026; Claude's deepest tests were Jul 30, Aug 1 and Aug 3, 2026. What those tests do show: Kimi K3's steps failed least often of the models tested that day (Jul 31, 2026 — 7.2%, fewest steps, most accurate account of its own work). Claude Sonnet 5 writes the cleanest code we have measured (Aug 3, 2026), and the family has produced the best-looking build we have measured (Jul 30, 2026), the most complete app of its test (Aug 1, 2026), and that cleanest code. Both are available in Taskade; automatic model routing picks the right one per task.
- Kimi: Jul 31, 2026 — fewest failed steps of the models tested (7.2%), fewest steps overall, most accurate account of its own work.
- Claude: Aug 3, 2026 — the cleanest code we have measured; plus the best-looking build we have measured (Jul 30, 2026) and the most complete app of its test (Aug 1, 2026).
See the full evidence at /tsk/kimi, /tsk/claude, and the TSK-1 hub.
The Headline
This is a distribution comparison before it is a capability comparison. Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters used per token, a 1,048,576-token context, and weights you can download. Claude is a closed family with a published rate card, a 1M-token context that carries no long-context surcharge, and Anthropic's safety training on top.
That difference decides more of the design than any benchmark does. If the weights have to live in your region, on your hardware, or inside an air-gapped network, only one of these two is even a candidate. If you need a supported gateway with contractual terms and a per-token price you can forecast, the other is the safer buy.
TL;DR: Kimi K3 is open-weight at 2.8T total, 104B used per token, with a 1,048,576-token context, released under Moonshot's bespoke Kimi K3 License — permissive enough to self-host, specific enough to read before you redistribute. Claude is closed, with published rates from $1/$5 per 1M on Haiku 4.5 to $10/$50 on Fable 5, and no surcharge on its 1M context. Route by task inside Taskade Genesis instead of standardising on one.
The Routing Matrix
The useful question is never "which model is best". It is "which model for this step, and what does that choice cost".
| Task | Reach for | Why | Cost consequence |
|---|---|---|---|
| Long-context reasoning over a whole repo or corpus | Either | Both reach ~1M tokens | Claude bills the full window at standard rates, but its 4.7+ tokenizer emits ~30% more tokens for the same text — budget per document, not per token |
| Agentic coding, multi-step tool use | Kimi K3 or Kimi K2.7 Code | Open weights let the loop run on your own infrastructure; K2.7 Code is purpose-built for coding | Self-hosting turns per-token spend into fixed GPU spend — good above sustained volume, bad below it |
| Conversational pair programming, terminal agent work | Claude | Claude Code plus the Opus tier | $5 in / $25 out per 1M on the Opus tier |
| Cheap high-volume classification and extraction | Claude Haiku 4.5, or a self-hosted open-weight model | Haiku is the cheapest published first-party Claude rung | $1 in / $5 out per 1M, halved again with the Batch API |
| Customer-facing writing and nuanced instruction following | Claude | Anthropic's strongest published differentiator, plus Constitutional AI on the output | Sonnet 5 at $2/$10 per 1M through 31 Aug 2026, then $3/$15 |
| Vision-to-text work | Either | K3 is image-text-to-text; Claude takes vision + text | Image tokens bill as input tokens on both sides |
| On-premise, air-gapped, or region-locked deployment | Kimi only | Claude does not ship weights | Infrastructure cost replaces API cost entirely |
| Fine-tune on proprietary data and keep the result | Kimi only | Weights are downloadable | Read the Kimi K3 License before redistributing the result |
| Repeated prompts over a stable prefix | Claude | Cache hits bill at 0.1× base input | A 90% saving on the cached portion |
| Bulk asynchronous jobs | Claude | Batch API is 50% off in both directions | Halves the rate card on anything that can wait |
Read the matrix as two columns of one system. Most real setups end up using an open-weight model where volume and control dominate, and a governed closed model where the output reaches a person.
What Claude Actually Costs
Anthropic publishes both sides of its pricing, which makes it the easy half of this comparison to plan against.
| Claude plan | Month-to-month | Billed annually |
|---|---|---|
| Free | $0 | — |
| Pro | $20/mo | $17/mo, billed $200 up front |
| Max | From $100/mo — 5x or 20x more usage than Pro | Not disclosed |
| Team | $25/seat/mo, premium seat $125/seat/mo | $20/seat/mo, premium seat $100/seat/mo |
| Enterprise | Seat price plus usage at API rates | — |
API rates, per 1 million tokens:
| Model | Input | Output |
|---|---|---|
| Claude Haiku 4.5 | $1 | $5 |
| Claude Sonnet 5 — introductory through 31 Aug 2026 | $2 | $10 |
| Claude Sonnet 5 — from 1 Sep 2026 | $3 | $15 |
| Claude Opus tier | $5 | $25 |
| Claude Fable 5 | $10 | $50 |
Three modifiers matter more than the headline numbers. The Batch API is 50% off in both directions. A cache hit bills at 0.1× base input. And web search on the API is $10 per 1,000 searches, which is a line item agentic workloads discover late.
The tokenizer footnote nobody prices in
Claude 4.7 and later use a newer tokenizer that produces roughly 30% more tokens for the same text. A rate card comparison between two labs is therefore not a cost comparison unless both sides are measured on the same corpus. Run your own text through both tokenizers before you conclude that a lower per-token price is a lower bill.
The good news in the other direction: Anthropic bills the full 1M-token context at standard rates. A 900,000-token request costs the same per token as a 9,000-token one. There is no long-context tier, which is unusual and genuinely worth something on document-heavy work.
And Kimi's price?
Moonshot AI publishes Kimi's API rates on its developer platform, and the honest answer is: read them at the source, at platform.kimi.ai, because a number copied into a comparison page starts aging the day it is published. Note that the older platform.moonshot.ai address now redirects to platform.kimi.ai — update any integration still pointing at the old host.
What is structurally true regardless of the current figure is that an open-weight model has a self-hosting floor a closed model cannot have. Above sustained volume that floor, not the rate card, is the number that decides the architecture.
Licence & Distribution: The Real Divide
This is the row most comparisons get wrong, so it is worth stating precisely.
| Dimension | Kimi K3 | Claude |
|---|---|---|
| Licence | Kimi K3 License — a bespoke Moonshot licence covering the code repository and the weights | Commercial terms of service |
| Is it MIT or Apache 2.0? | No. Modified MIT applied to the earlier K2 line and does not carry forward | n/a |
| Weights downloadable | ✅ Yes, Hugging Face | ✗ No |
| Self-host | ✅ Yes, licence permitting | ✗ Anthropic gateway only |
| Fine-tune on your data | ✅ Yes | ✗ No |
| Redistribute a fine-tune | ⚠️ Read the licence first | ✗ No |
| Audit the weights | ✅ Yes | ✗ No |
| Data residency / air-gap | ✅ Achievable | ✗ Not achievable |
Say it plainly: Kimi K3's weights are downloadable and self-hostable, under a vendor-specific licence your legal team should read before you redistribute or resell. That is a real and meaningful difference from a closed model. It is not the same freedom as a standard permissive licence — DeepSeek ships MIT across both code and weights, and Qwen's open weights are Apache 2.0. On the open-weight spectrum, Kimi sits between those and Claude, not alongside them.
Anyone who tells you K3 is "MIT licensed" is repeating a K2-era fact.
Architecture: One Published, One Undisclosed
Kimi K3's structure is public, and it is unusual enough to be worth understanding before you design around it.
- Sparse by design. 104B of 2.8T parameters are used per token, so running cost tracks the number used, not the headline count. The headline number is a capacity claim, not a compute claim.
- A hybrid attention design. 69 of 93 layers are Kimi Delta Attention, a linear-cost variant, with 24 gated multi-head latent attention layers interleaved. The intent is holding quality across a million-token window without paying full attention cost at every layer.
- Vendor claims are vendor claims. Moonshot makes stronger comparative statements about how this design performs against full attention. Benchmark on your own workload before you build a roadmap on them.
Claude's architecture is undisclosed. What Anthropic publishes instead is the safety framework — Constitutional AI — the tier ladder from Haiku through Opus and Fable, and a rate card. For a lot of buyers that trade is the point: less to inspect, more to hold someone accountable for.
Self-Host vs Managed Gateway
Kimi K3 can be self-hosted. Claude cannot. But "can" and "should" are different questions, and a 2.8-trillion-parameter model is not a weekend deployment.
| Kimi K3 self-hosted | Claude | Taskade Genesis | |
|---|---|---|---|
| Possible at all | ✅ Yes, licence permitting | ✗ No | ✅ Managed routing across providers |
| Hardware | Serious multi-GPU, high-VRAM infrastructure | None | None |
| Cost shape | Fixed infrastructure + ops headcount | Per token, published | Included in the subscription |
| Data residency | ✅ Fully under your control | Anthropic's regions | Managed |
| Who owns uptime | You | Anthropic | Taskade |
| Breaks even when | Volume is high and sustained | Always available | Immediately, at low and mid volume |
At low and moderate volume a managed gateway wins on every measure except control. Self-hosting K3 makes sense when volume is genuinely high and sustained, or when data residency makes the question moot — and in the second case it is not a cost decision at all.
Choose Claude If…
A fair comparison names the cases where the other side is simply the better buy. There are several here.
- Your output is customer-facing prose. Long-form writing quality and nuanced instruction following are Anthropic's most consistently cited strengths, and there is no self-hosting workaround for "the draft reads better".
- You need someone contractually accountable. A published rate card, enterprise agreements, and a named safety framework are things an open-weight download cannot give you. For regulated buyers that is often the whole decision.
- You want a finished product, not a model. Projects, Artifacts and Claude Code are polished surfaces your team can use today without any infrastructure work at all.
- Your volume is modest. Below meaningful sustained throughput, the API is cheaper and vastly simpler than standing up multi-GPU inference — and Haiku 4.5 at $1/$5 per 1M is inexpensive by any measure.
Choose Kimi If…
- The weights have to be yours. Data residency, air-gapped environments, or an audit requirement that a vendor description cannot satisfy.
- You are fine-tuning on proprietary data and need to keep the resulting model.
- Your volume is high and sustained enough that a fixed infrastructure bill beats a per-token one.
- You want optionality. A downloadable model cannot be deprecated out from under you the way a hosted endpoint can — Moonshot's own V1 series sunsets on 31 August 2026, which is exactly the risk open weights insure against.
The Taskade Angle: Route, Don't Standardise
Most comparison pages end with "pick one". The operating reality of 2026 is that serious teams run several models and route between them.
Taskade routes across 15+ frontier models from OpenAI, Anthropic, Google and open-weight providers inside one workspace, with the AI allowance included in the subscription rather than a separate per-seat consumer subscription with each lab. Paid plans start at Pro $10/mo billed annually, with Business $25, Max $100 and Enterprise $250 per month billed annually. Routing is managed centrally and updates as new models ship, so a lab's roadmap is not your roadmap.
Four patterns that hold up:
- Open-weight in the loop, governed model on the output. The agent's tool-using steps run on a cost-efficient model. The paragraph a customer actually reads runs on the model you are willing to be accountable for.
- Cheap tier for triage, premium tier for the hard 10%. Bulk classification and extraction go to the cheapest capable rung. Escalations route up.
- Long-context ingest, short-context reasoning. One step reads the whole corpus. The next step reasons over a distilled summary at a fraction of the tokens.
- Everything behind Workspace DNA. Whichever model runs a step, the result lands in the same project graph, so the next agent inherits the context instead of re-deriving it.
See 10 Best Open-Source AI LLMs in 2026 for how Kimi sits against the rest of the open-weight field.
Final Word: Different Distribution, Different Job
Kimi K3 is a 2.8-trillion-parameter open-weight model with a million-token context and a licence that is genuinely permissive but genuinely bespoke. It is the choice when control — of hardware, of region, of the model itself — is the requirement.
Claude is a closed frontier family with a published rate card, no long-context surcharge, a tokenizer quirk worth measuring, and a safety framework Anthropic will stand behind. It is the choice when accountability and finish are the requirement.
Neither is the winner. The winner is the setup that puts each one where it belongs and does not have to renegotiate a contract to change its mind.
▲ Memory feeds Intelligence. ■ Intelligence triggers Execution. ● Execution creates Memory. Two distribution models. One workspace. No single point of vendor failure.
This is the origin of living software. 🌱
Build with open-weight and closed models side by side →
Related reading
- Claude Fable 5 & Mythos 5 Explained. Anthropic's newest model tier: what it is for and what it costs.
- 10 Best Open-Source AI LLMs in 2026. The full open-weight ranking.
- Kimi vs DeepSeek. Bespoke licence versus MIT, head to head.
- Qwen vs DeepSeek. The other open-weight duel.
- Multi-Model AI Access. How Taskade routes across providers.
- Multi-Agent Teams. Specialists with different model picks.
- Taskade MCP Server. Use Claude Desktop or Cursor with your workspace.
- Free Claude Alternative. How Taskade compares to Claude as a workspace.
- TSK-1 Kimi profile — Full benchmark evidence for the Kimi family.
- TSK-1 Claude profile — Full benchmark evidence for the Claude family.
- TSK-1 hub — The complete model benchmark dataset.