TL;DR: Kimi Code is a legitimate free-to-install MIT-licensed terminal coding agent backed by K2.6 — the same open-weight model family that powers Cursor Composer 2.5 under the hood. The 88% cost savings figure is real, but it compares K2.6 API rates against Claude Opus 4.7 API rates, not against Cursor’s $20/month flat subscription. For API-heavy teams burning through usage caps, the math gets genuinely interesting. For developers who want a single frictionless IDE workflow, Cursor still wins.

	Kimi Code (Moderato)	Cursor Pro	Claude Code (Pro)
Best for	Bulk agentic tasks, cost-sensitive teams	IDE-native daily driver	Complex multi-file refactors
Price / Cost	$19/mo + API at $0.95/$4.00 per M tokens	$20/mo flat	$20/mo (Claude Pro)
The catch	Free tier has no Kimi Code quota; 5-hr rolling cap stings at scale	Composer usage caps	Rate limits hit fast on Pro

Honest take: If you’re already on Cursor and hitting Composer caps weekly, Kimi Code’s $19/month Moderato plan with pay-as-you-go API is the more cost-predictable stack for batch refactors and test generation. If you’re not hitting those caps, stay on Cursor.

The product confusion, cleared up

“Kimi Code” labels two distinct things: a feature inside kimi.com’s subscription that gives you a browser-based coding agent, and kimi-code — an MIT-licensed terminal CLI built by Moonshot AI (think Claude Code, but backed by Moonshot’s model stack). Both exist simultaneously, share the same underlying K2.6 model, and are sold under a single pricing page. Half the contradictory forum posts about pricing trace back to people conflating them.

This review covers both, with emphasis on the CLI because that’s where the interesting technical decisions live.

What K2.6 actually is

Kimi K2.6 launched April 20, 2026. Under the hood it’s a 1-trillion-parameter Mixture-of-Experts model — same architectural family as K2 and K2.5 — rebuilt with heavier agentic training on multi-step coding workflows.

Benchmarks that matter for coding work:

Benchmark	K2.5	K2.6	GPT-5.4	Claude Opus 4.6
SWE-Bench Verified	~72%	80.2%	—	—
SWE-Bench Pro	50.7%	58.6%	57.7%	53.4%
Terminal-Bench 2.0	50.8%	66.7%	—	—
BrowseComp (Agent Swarm)	78.4%	86.3%	—	—

K2.6 leads the SWE-Bench Pro leaderboard as of June 2026. That’s the benchmark designed to resist shortcut-finding, so it’s a better proxy for real agentic work than SWE-Bench Verified.

The signal that counts more than any benchmark: Cursor Composer 2.5 was built on the K2.5 base, with extensive continued pretraining and large-scale RL on top. Moonshot AI’s model passed production scrutiny at the company charging $20/month for AI coding. K2.6 is a step further up from K2.5 on every coding dimension.

Context window: 1M tokens, up from 256K on K2.5. That matches Gemini 2.5 Pro and roughly doubles Claude Sonnet 4.6’s usable window at API scale.

The 88% cost savings claim — what it actually means

The number comes from comparing K2.6 API rates against Claude Opus 4.7 API rates directly:

Model	Input ($/M tokens)	Output ($/M tokens)
Kimi K2.6 (Official API)	$0.95	$4.00
Kimi K2.5 (Official API)	$0.60	$3.00
Claude Opus 4.7	~$5.00	~$25.00

On input alone that’s an 81% discount. Coding workloads skew toward longer outputs (diffs, test suites, full file rewrites), so the blended savings land closer to 88% on a realistic tokens-in / tokens-out ratio — the number holds up.

What it doesn’t tell you: Cursor Pro at $20/month is often cheaper than either for solo developers doing moderate volume. At $0.95/M input tokens you need roughly 21 million input tokens per month before K2.6 API billing beats Cursor Pro on pure cost. A developer running 50 agentic sessions a month at 5,000 input tokens each moves 250K tokens — that’s $0.24 at K2.6 rates. For that developer, the comparison isn’t “which API is cheaper.” It’s “does Kimi Code match Claude Code quality?”

The cost advantage becomes real in these scenarios:

• Batch refactoring a 300K-line codebase
• Nightly test-generation pipelines running unattended
• Teams with 10+ developers routing through an API budget instead of per-seat subscriptions

At $0.95/M input with 80.2% SWE-Bench Verified quality, K2.6 is a credible API backend for those workloads.

Installing Kimi Code CLI (v1.12.0, tested June 2026)

# macOS/Linux — no Node.js runtime dependency on end user
curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash

# Windows (PowerShell)
irm https://code.kimi.com/kimi-code/install.ps1 | iex

After install, kimi opens the TUI. First-run authentication:

$ kimi

? How would you like to authenticate?
> Kimi Code OAuth (kimi.com subscription — uses your plan quota)
  Moonshot API key (platform.moonshot.ai — pay per token)

OAuth ties to your kimi.com subscription and uses your Moderato (or higher) plan’s Kimi Code quota. The API key path routes to Moonshot’s billing console on platform.moonshot.ai — that’s where you pay $0.95/M tokens directly.

If you’re on the $19/month Moderato plan, OAuth is the right choice. It avoids double-billing. If you want the pay-as-you-go path with no subscription, use an API key.

One real friction point that bit users in GitHub discussion #1147: the VS Code extension requires a compatible Node.js version on the extension host. If the “VS Code plugin not responsive” bug hits you, check your extension host’s Node version — the CLI binary itself is self-contained, but the IDE integration layer is not.

What the CLI does well

Three built-in subagent modes cover the main workflow shapes:

• /coder — modifies files, runs shell commands, reads directory structure. This is the default agentic mode.
• /explore — reads and summarizes without touching files. Useful before giving the agent edit access.
• /plan — decomposes the task, writes a plan document, waits for your approval before executing.

MCP server configuration is conversational:

/mcp add
> MCP server URL: http://localhost:3000
> Server added: local-db. It will be available in your next session.

No JSON config editing. This is meaningfully better than Cursor’s manual mcp.json edits or Claude Code’s claude mcp add syntax.

Agent Swarm — Moonshot’s parallel subagent system — runs up to 100 simultaneous subagents on Moderato and 300 on Vivace. Moonshot reports 4.5x wall-clock speedup on parallelizable tasks like batch refactoring across independent modules. In practice on a 50-file TypeScript refactor, parallel execution brings a 20-minute sequential job down to around 5 minutes. That’s the most genuinely differentiated capability versus Cursor, which runs agents sequentially.

The VS Code extension (moonshot-ai.kimi-code on the Marketplace) surfaces the agent in a side panel but does not yet offer Tab autocomplete. For inline completions you still need Cursor or GitHub Copilot running alongside. The agent side panel works well for longer tasks you hand off and come back to.

Where it breaks

The 5-hour rolling quota. Moderato allocates between 300 and 1,200 API calls per 5-hour window, refreshing on a rolling basis. There’s no burst capacity and unused quota doesn’t carry forward to the next window. Hit the cap at 2 AM while a long batch job is running and it stalls until the window refreshes. This is the most common complaint in Kimi CLI GitHub issues as of June 2026, and it’s a real operational risk for unattended pipelines.

Complex constraint following. On multi-constraint instructions — “refactor this without changing the public API surface, add docstrings to every new method, and keep all existing tests green” — K2.6 handles the core refactor correctly but occasionally drops the secondary constraints when the instruction string runs long. Claude Sonnet 4.6 is more reliable on this pattern. The gap narrows if you break instructions into separate turns rather than loading one long prompt.

Context management at scale. The 1M-token window is impressive, but at $4.00/M output tokens a deep exploration pass over a large codebase isn’t free. The CLI doesn’t yet auto-compact older context the way Claude Code’s /compact does. On a large monorepo, watch your token counter or costs can surprise you.

Self-hosting is not realistic for most teams. K2.6’s open weights are public, but the model is a 1-trillion-parameter MoE — you need a multi-GPU server to run it. The weights are useful for fine-tuning shops or cloud providers building on top, not for an individual developer wanting local inference. For actually-local coding models, see Devstral 2 on a single RTX 4090 — that’s the realistic self-host target. For VRAM sizing guidance, runaihome.com’s local LLM hardware guide has the numbers.

Pricing, laid out

Plan	Monthly	Kimi Code included	Agent Swarm
Adagio	$0	❌	❌
Moderato	$19	✅	25 uses
Allegretto	$39	✅	Up to 100 subagents
Allegro	$99	✅	Full + speed priority
Vivace	$199	✅	300 subagents + Kimi Claw cloud deploy

API-only path: K2.6 at $0.95/M input, $4.00/M output via platform.moonshot.ai. K2.5 at $0.60/$3.00 if you want lower cost and can accept slightly lower agentic quality. OpenRouter carries both under a single billing dashboard if you’re mixing models across tools.

The Adagio free plan does not include Kimi Code quota — this is the opposite of how Kimi Code is sometimes described in third-party comparisons, which conflate the free CLI binary (genuinely free) with a free API quota (not included).

Who should actually pay for this

Pay $19/month if you’re a developer already comfortable in a terminal who’s hitting Cursor’s Composer usage caps regularly. The Agent Swarm parallel execution and the lower per-token API cost make the $19/month Moderato plan genuinely cheaper than equivalent Cursor usage at scale.

Use the API key path if you’re building a pipeline (test generation, batch refactor, code review bot) and want cost-per-token pricing with no monthly floor. At $0.95/M input with SWE-Bench Verified quality close to frontier models, K2.6 is competitive with every closed-source option.

Stay on Cursor if you want Tab autocomplete, a polished IDE experience, and one monthly bill. Cursor’s Composer 2.5 already runs on a K2.5 base internally — you get the model quality through a better UX for $20/month. See the Cursor vs Claude Code comparison for how those two stack up.

Skip Kimi Code if you don’t have a specific cost or volume pain point. The free Adagio plan sounds appealing but doesn’t give you meaningful Kimi Code access. The $19/month entry is comparable to Claude Code’s Pro tier, and Claude Code wins on instruction-following reliability for complex multi-constraint tasks.

FAQ

Is Kimi Code actually free to use?
The CLI is MIT-licensed and free to install. Actual usage requires a paid kimi.com subscription (Moderato starts at $19/month) or a funded Moonshot API account. The Adagio free plan does not include Kimi Code quota.

Does K2.6 work inside Cursor as a backend?
Not via Cursor’s model switcher — Cursor doesn’t expose arbitrary API endpoint swapping. You can run Kimi Code CLI in a separate terminal alongside Cursor. For routing K2.6 through tools that do support custom backends, see Kimi K2.6 in Cursor and Cline via OpenRouter.

What exactly is the 88% cost savings comparing?
K2.6 API costs ($0.95/M input) versus Claude Opus 4.7 API costs (~$5.00/M input) on a blended input/output ratio. It is not a comparison against Cursor Pro’s $20/month flat rate.

What’s the context window on K2.6?
1M tokens. K2.5 is 256K. Both are available via API; the kimi.com browser interface has a lower effective cap than the raw API.

Can I run K2.6 locally?
The weights are public, but the model is a 1-trillion-parameter MoE. You need multi-GPU infrastructure to run inference. It’s not a practical local model for individual developers.

Does Agent Swarm work on the $19/month Moderato plan?
Yes, with 25 uses included. Full Agent Swarm (100 parallel subagents) requires Allegretto at $39/month.

JetBrains support?
Agent Client Protocol integration lists JetBrains, but the VS Code extension is the most actively maintained IDE surface as of June 2026. JetBrains support exists but lags VS Code in polish.

Sources

• Kimi Code CLI — GitHub repository (MoonshotAI/kimi-code)
• Kimi K2.6 & Kimi Code Review: Saving 88% Coding Costs? — Medium
• Kimi K2.6 Benchmarks, API Pricing & SWE-Bench Score — AIToolsRecap
• Kimi K2.6 Benchmarks 2026 — BenchLM.ai
• Kimi K2.6 API Pricing — OpenRouter
• Kimi Membership Pricing — kimi.com
• Cursor Composer 2.5 Built on Kimi K2.5 — The New Stack
• Kimi Code OAuth Feature Request — GitHub Issues #757
• Kimi Code 2026: Plans, Pricing & Developer Guide — NxCode

Last updated June 7, 2026. Pricing and features change frequently; verify current state before purchasing.