Users report Kimi K3 tasks can burn 20% of a weekly $20 plan

• Kimi K3 landed near the closed-model coding frontier: LLMJunky's benchmark chart puts it at 67.5 on DeepSWE, 88.3 on Terminal Bench 2.1, 81.2 on FrontierSWE, and 42.0 on SWE Marathon.
• The most convincing creator demos were spatial UI and frontend work, where stevibe's iPhone test found K3 produced a cleaner exploded view than GPT-5.6 Sol.
• The Rocket League-style test exposed the feel gap: LLMJunky called K3 “honestly pretty good” with nice UI, then said the physics and small gameplay details still trailed Fable.
• The subscription math got ugly fast: LLMJunky's Kimi Code run got about 85% through a browser game before hitting the 5-hour quota and burning 20% of the weekly allotment.
• Anthropic moved on access at the same time: ClaudeDevs said Claude Code weekly limits would stay 50% higher through August 19, while Allie K. Miller said Fable would remain in Max and Team Premium seats.

Moonshot's Kimi K3 quickstart frames K3 as a 2.8T-parameter, 1M-context model built on Kimi Delta Attention and Attention Residuals. The archived Kimi tech blog says K3 is live on Kimi.com, Kimi Work, Kimi Code, and the API, with full weights due by July 27. OpenRouter's K3 page warns that upstream capacity is limited and may return frequent 429s. A Kimi Code setup analysis says the $19 Moderato tier gets K3 at 256K context, while Allegretto unlocks the full 1M window.

Moonshot's official docs describe K3 as its most capable flagship model, with 2.8T parameters, native visual understanding, and a 1M-token context window.

The Kimi pricing page lists K3 at $0.30 per 1M cached input tokens, $3 per 1M input tokens, $15 per 1M output tokens, and a 1,048,576-token context window. Kimi's launch blog says K3 uses max thinking effort by default at launch, with low- and high-effort modes coming later.

The Kimi Code console screenshot in LLMJunky's post labels K3 as the flagship model and says it supports up to 1M context tokens, optimized for coding, 3D gaming, and complex knowledge tasks.

The benchmark chart LLMJunky shared does not make K3 a clean winner. It makes it hard to dismiss.

• DeepSWE: GPT-5.6 Sol 73.0, Fable 5 70.0, Kimi K3 67.5.
• Terminal Bench 2.1: GPT-5.6 Sol 88.8, Kimi K3 88.3, Fable 5 84.6.
• FrontierSWE: Fable 5 86.6, Kimi K3 81.2, GPT-5.6 Sol 71.3.
• Program Bench: Kimi K3 77.8, GPT-5.6 Sol 77.6, Fable 5 76.8.
• SWE Marathon: Kimi K3 42.0, Opus 4.8 40.0, GPT-5.6 Sol 39.0, Fable 5 35.0.

LLMJunky called K3 near-parity with existing SOTA models and singled out UI design, spatial awareness, and writing as likely strengths in the same post.

The cleanest K3 examples came from interface-heavy prompts. stevibe tested a single HTML canvas animation of a rotating iPhone that disassembles into an exploded view, pauses, then reassembles, and said K3's output had the cleaner exploded view against GPT-5.6 Sol.

Creators pushed the same pattern into bigger browser builds:

• levelsio showed a macOS-style desktop running in the browser.
• Min Choi collected K3 examples including a Game Boy Advance emulator.
• Min Choi's thread listed a CS:GO x Portal clone made in three shots with 600K tokens.
• the same thread included a macOS-style UI built in 15 minutes from a single prompt.
• the final item pointed to a cyberpunk web-swinging game.

That is Christmas morning for vibe-coders: a cheap-looking API sticker, a visual model picker, and outputs that make a landing page or browser toy feel shippable.

LLMJunky's Rocket League-style benchmark became the useful cold shower. K3 made a polished-looking result, but he said it needed follow-ups to become playable and still lacked the feel Fable got on the first try in his follow-up.

The missing pieces LLMJunky named were specific:

• physics and game feel
• goal explosions
• the circle under the ball
• momentum mechanics
• smoke behind the ball
• small gameplay details Fable generated without nudging

He still called K3 a good model in one reply and “really freaking good” for UI in another. The gap was not screenshot quality. It was control feel.

The first $19-plan complaint was not about model quality. It was quota burn.

LLMJunky said one browser-game prompt got about 85% done, hit the 5-hour quota, and consumed 20% of the weekly allotment in his Kimi Code post. A later screenshot showed a 403 usage-limit error at 101K tokens, 40% of a 256K context window in the terminal capture.

He repeated the same math in replies: one prompt on the $20 plan used the full 5-hour window and 20% of the weekly limit before finishing in one reply, and half of one context window was enough to stop the run in another.

OpenRouter had its own capacity issue. LLMJunky said K3 on OpenRouter rate-limited every two or three tool calls, which matches OpenRouter's warning that upstream capacity was limited and frequent 429 errors could occur.

Pieter Levels posted the working path for using Kimi K3 inside OpenCode:

1. Ask Codex or Claude Code to install OpenCode.
2. Create a Kimi account, pay for membership, and get an API key.
3. Run OpenCode, use /connect, connect to Kimi Code, and paste the key.
4. Switch to BUILD mode with Shift+Tab, with bypass permissions set in /settings if desired.

Levels said he first tried OpenRouter, got “rate limited upstream,” then went direct through Kimi and continued working on a Windows XP browser project in his setup post. The attached OpenCode screenshot shows Kimi K3 inspecting MSN relay code with 103,156 tokens in context and a permission prompt for /opt/escargot/*.

The API sticker says $3 input and $15 output per 1M tokens, but illscience's eval put a different number on real tasks.

In a 10-model critical-thinking test over claims and a roughly 3,000-claim knowledge base, illscience measured Kimi K3 at $440 per 1,000 claims and 204.2 seconds per claim. The same table put Fable 5 at $453 per 1,000 claims and 54.7 seconds per claim, GPT-5.6 Terra Medium at $72 and 12.1 seconds, and GPT-5.6 Luna High at $65 and 18.7 seconds.

The linked eval board is available from illscience's thread. His conclusion was narrow: K3 looked token-inefficient on that workload, while also strong at critical thinking and prose.

Anthropic's Fable move landed in the same fight. Allie K. Miller said Fable would be permanently included in Max plans and Team Premium seats, while arguing the lower limit still pointed to compute pressure.

ClaudeDevs also said Claude Code weekly limits would stay 50% higher through August 19 for Pro, Max, Team, and seat-based Enterprise users in the official post. The linked [Claude Help Center page][link:16:0] says the promotion applies automatically, excludes Free plans and consumption-based Enterprise seats, and does not affect the 5-hour usage limit.

Bilawal Sidhu framed the limit math bluntly: the 20x plan became the 10x plan, and the 5x plan became the 2.5x plan in his post. After the extension, Sidhu's update said the 1.5x promo stopped those plans from effectively becoming 6.67x and 1.67x for another month.

The July 27 weight release still points at provider-class hardware for most creators. stevibe split the landscape into local models, open models, and closed models: local means “I can run them on my hardware,” open means “some people can,” and closed means only the creator can run them.

That distinction matters for K3. LLMJunky's benchmark post joked that he needed 3.5TB of VRAM, and his follow-up said a lower-end path might still be about 1.7TB and several hundred thousand dollars of hardware.