Where to Access Kimi K3: Every Provider and Price Compared (2026)

Start here: OpenCode Go gives you $5 in credits plus $5 off the first month when you use referral code M6HEHM4JM5. It is the cheapest low-friction way to try Kimi K3 inside an agent-first coding interface.

Official sources#

Kimi K3 launched on July 16, 2026. It is Moonshot's flagship 2.8-trillion-parameter model, the first in the open 3T class, with a 1,048,576-token context window, native vision, and always-on reasoning by default. Because the weights are not public until July 27, 2026, every hosted route right now routes through Moonshot's own infrastructure or an authorized partner. There is no unlimited free tier for K3. The Kimi free chat tier covers K2.6, not K3. This post maps the paid and cheap routes so you can pick the one that fits your workflow without surprise billing.

Prices and availability were verified on July 23, 2026. Pricing pages and capacity limits move, so treat the numbers as a snapshot.

Last verified: July 23, 2026.

What Kimi K3 is, in one paragraph#

K3 is a sparse mixture-of-experts model built on Kimi Delta Attention and Attention Residuals. Only 16 of 896 experts are activated per token, which is how Moonshot keeps inference costs practical despite the 2.8T parameter count. It accepts text and images, supports tool calling and structured outputs, and exposes low, high, and max reasoning-effort variants. The max effort is the default at launch. For the full benchmark and capability breakdown, see the Kimi K3 developer guide and the Kimi K3 vs K2.7 comparison. This page is only about access and price.

The access routes#

OpenCode Go (flat fee, agent-first)#

OpenCode Go is a $10 per month subscription with a $5 first-month promo when you use referral code M6HEHM4JM5. It includes K3 through an OpenAI-compatible endpoint, and the model is available in the OpenCode terminal agent. As of July 2026, OpenCode Go lists K3 limits at 110 requests per 5 hours, 250 per week, and 490 per month. That is enough for serious evaluation and light daily coding, but not enough for a high-volume production agent.

This is the best starting point if you want to feel how K3 behaves inside an agent loop before you buy API credits or a Kimi membership. The referral credit covers the first month and then some. Use it as a cheap sandbox.

Kimi first-party API#

The Moonshot API is the reference route. It is fully OpenAI-compatible, with base_url="https://api.moonshot.ai/v1" and model="kimi-k3". You need a $1 minimum top-up to unlock K3. After that, billing is pay-as-you-go:

• $3.00 per million fresh input tokens
• $0.30 per million cached input tokens
• $15.00 per million output tokens

Context caching is automatic, with no cache ID or TTL to manage. Moonshot reports a 90% plus cache-hit rate on programming workloads through its Mooncake split-inference architecture, so the effective input price on repeated code or long context is often much closer to $0.30 than $3.00.

OpenRouter#

If you already centralize model billing through OpenRouter, the model ID is moonshotai/kimi-k3. OpenRouter currently routes K3 requests to Moonshot's hosted INT4 endpoint, so the rates match Moonshot's list price: $3 input, $15 output, $0.30 cached. You get a single API key and failover without managing another provider account. The route is useful if K3 is one model in a larger router setup.

Cloudflare Workers AI#

Cloudflare lists moonshotai/kimi-k3 in its Workers AI catalog. If your stack already routes inference through Workers AI, you can invoke K3 with the same account and billing you use for other models. Cloudflare does not publish fixed per-token prices in the docs; pricing is shown in the Cloudflare dashboard per account. This route matters most if you are already on Cloudflare and want to keep inference inside the same data residency and billing boundary.

Kimi Code and Kimi Work#

Kimi Code is Moonshot's terminal coding agent. K3 is selectable with /model and the model ID k3 inside Kimi Code. Kimi Code is part of a Kimi membership. Membership tiers are Moderato ($19/mo), Allegretto ($39/mo), Allegro ($99/mo), and Vivace ($199/mo). All membership features share a single credit pool, and K3 consumption draws from that pool. Higher tiers include proportionally more credits and faster rate limits.

Kimi Work is the desktop knowledge-work environment. It requires version 3.1.0 or later to use K3. The same membership and credit rules apply.

For a solo developer doing exploratory coding, the cheapest Kimi membership is more expensive than OpenCode Go or the API. It becomes competitive when you are already using Kimi's chat, slides, deep research, and other membership features and can amortize the subscription across all of them.

Local and self-host#

Full weights are promised by July 27, 2026. Moonshot has not published the final license as of July 23, so self-hosting details are still pending. Once the weights are out, the model will be servable through vLLM, SGLang, Transformers, KTransformers, and llama.cpp. At 2.8T total parameters, a 4-bit quant still targets high-RAM multi-GPU hardware. For consumer laptops, a smaller dense model is the practical local option. See the best local coding LLMs for alternatives.

Price comparison#

The table is a blunt instrument. OpenCode Go and Kimi membership are flat-fee subscriptions, so their per-token cost depends on how much you use. The Kimi API and OpenRouter are pure per-token. Because K3's context caching is automatic, the effective API input cost on repeated or long-context coding workloads is usually closer to the $0.30 cached rate than the $3.00 headline rate.

Which route should you pick?#

• Just trying K3: OpenCode Go with referral code M6HEHM4JM5. It is the lowest-cost way to test the model inside an agent.
• API or automated workloads: Kimi API directly if you want the simplest first-party relationship; OpenRouter if you want one key for many models.
• Already on Cloudflare: Workers AI keeps inference inside your existing stack.
• Daily coding in Kimi Code: a Kimi membership only makes sense if you also use chat, slides, or other membership features.
• High-volume production: wait for the weight drop on July 27 and evaluate self-hosting costs against steady per-token spend.
• Truly local or air-gapped: wait for weights, but budget for serious GPU memory or step down to a smaller model.

FAQ#

Is Kimi K3 free?#

No. The free tier on Kimi covers K2.6 chat. K3 requires a $1 minimum top-up on the API, an OpenCode Go subscription, or a Kimi membership. There is no permanent, unlimited free hosted route.

What is the cheapest way to use Kimi K3?#

For light use, OpenCode Go is $5 for the first month with referral code M6HEHM4JM5 and then $10 per month. For token-based workloads, the Kimi API and OpenRouter both list $3 input and $15 output, but automatic context caching often drops the effective input price toward $0.30 on repeated code contexts.

Can I run Kimi K3 locally?#

Not yet. Moonshot has promised full weights by July 27, 2026. The license and exact serving requirements are still pending. At 2.8T parameters, even a 4-bit quant targets high-RAM multi-GPU hardware, not consumer laptops.

What is the correct model ID?#

• Moonshot API: kimi-k3
• OpenRouter: moonshotai/kimi-k3
• Cloudflare Workers AI: moonshotai/kimi-k3
• OpenCode Go: kimi-k3 (or opencode-go/kimi-k3 depending on your OpenCode config)
• Kimi Code: k3

How does OpenCode Go relate to the Moonshot API?#

OpenCode Go is a separate subscription. It pays for access to a pool of open-weights models, including K3, through its own endpoint and rate limits. You do not need a Moonshot API key, but you also do not get the direct per-token billing or the full context-caching economics of the Moonshot API.

Continue Reading#

• Kimi K3 Developer Guide: What the 2.8T Open Model Changes
• Kimi K3 vs K2.7: Is the Upgrade Worth It for Coding?
• Kimi K3 Drops: Moonshot's 2.8T Frontier Model
• OpenCode Developer Guide 2026
• Model Routers and the Optionality Advantage
• AI Coding Tools Pricing in 2026

Sources#

• Kimi K3 tech blog - fetched July 23, 2026
• Kimi API Platform: Kimi K3 quickstart - fetched July 23, 2026
• Kimi K3 pricing - fetched July 23, 2026
• OpenCode Go docs - fetched July 23, 2026
• OpenRouter: moonshotai/kimi-k3 - fetched July 23, 2026
• Cloudflare Workers AI: Kimi K3 - fetched July 23, 2026
• Kimi membership pricing - fetched July 23, 2026