Kimi K2.6 and K2.7 Code Are Live on TulexAI — the Moonshot Family Is Complete

TL;DR: Two new Moonshot AI models go live on TulexAI today: Kimi K2.6, a cheap multimodal general-purpose model at 2 PT per message, and Kimi K2.7 Code, an agentic coding specialist at 4 PT. They join Kimi K3 — the frontier reasoner that landed here on July 16 at 7 PT — to complete the Moonshot family. All three are available on every TulexAI plan, from $11/mo. Here's what each one is actually good at, with honest prices and the caveats included.

The Moonshot family, complete

Moonshot AI has quietly become the most interesting lab outside the US frontier three. K3 launched on July 16 and immediately posted an Artificial Analysis Intelligence Index score of 57 — #7 of 190 models, within three points of Claude Fable 5's leading 60 at roughly a third of the blended API cost. Today the supporting cast arrives:

Model	Best at	Input / 1M	Output / 1M	Context	On TulexAI
Kimi K2.6 (new today)	Everyday chat, multimodal input	$0.95	$4.00	262K	2 PT / message
Kimi K2.7 Code (new today)	Agentic coding	$1.90	$8.00	262K	4 PT / message
Kimi K3 (live since Jul 16)	Deep reasoning	$3.00	$15.00	1M	7 PT / message

All API prices verified against Moonshot's platform docs (July 2026). Cached input is aggressive across the family — $0.30 per million on K3 (90% off list) and $0.38 on K2.7 Code. Every one of the three is on every TulexAI plan, Basic ($11) included — no premium-tier gate.

Kimi K2.6: the new default for everyday work

K2.6 is the general-purpose workhorse of the family, and at 2 PT per message it's one of the cheapest capable models on TulexAI. Three things make it more than a budget filler:

• Thinking and non-thinking modes. Ask it something simple and it answers fast; flip on thinking mode for problems that deserve a longer pass. You choose per conversation instead of paying reasoning-model prices for "what's a good subject line."
• Text, image and video input. Multimodal input including video is still rare at this price point — you can drop a clip in and ask questions about it, at 2 PT.
• A 262K context window. Not K3's million, but comfortably enough for long documents, full codebases-worth of context, or a day-long chat thread.

If you've been using a budget model for drafting, summarizing and Q&A, K2.6 is the new one to try first.

Kimi K2.7 Code: agentic coding at the missing price point

K2.7 Code (API id kimi-k2.7-code-highspeed) is a specialist: an agentic coding model built for multi-step tool use — read the repo, plan, edit, run, fix, repeat. At 4 PT per message it lands in a gap in our coding lineup that users kept pointing out: Grok Build sits at 1.2 PT, Claude Opus Code at 12 PT, and until today there was nothing in between. K2.7 Code is the middle option — meaningfully more capable than the budget tier on long agentic chains, at a third of the premium coder's PT price.

The honest framing: we don't yet have independent coding-benchmark numbers for K2.7 Code, so we won't invent any. Artificial Analysis' Coding Agent Index is still led by GPT-5.6 Sol at 80 — and Sol remains on TulexAI at 12 PT if leaderboard-topping agentic coding is the job. K2.7 Code's case is price-per-iteration: coding agents burn a lot of messages, and 4 PT versus 12 PT compounds fast over a long session.

Kimi K3: the frontier reasoner, one week in

A quick recap for anyone who missed the July 16 launch. Per Artificial Analysis, K3 scores 57 on the Intelligence Index (#7 of 190) — three points behind Claude Fable 5, the leader at 60. The blended-cost math is the story: at the standard chat mix, K3 works out to about $12 per million tokens against Fable 5's $40 — a 3.3× gap for a 3-point score difference. On TulexAI that translates to 7 PT versus 25 PT, and K3 is on every plan while Fable 5 requires Pro ($23/mo) or above.

The equally honest catch: K3 is a deep-reasoning model, and it is slow. Artificial Analysis measured 32.6 output tokens per second (#147 of 190) and roughly 161 seconds to first token on its test. Brilliant, not snappy. Use it for the problems worth waiting on — and that's exactly why K2.6 exists for everything else.

Which Kimi — and when to reach for Claude or GPT instead

• Pick Kimi K2.6 when: it's everyday drafting, summarizing, Q&A, or anything involving image/video input. At 2 PT it's the volume workhorse.
• Pick Kimi K2.7 Code when: you're running agentic coding sessions and want more capability than the 1.2 PT tier without paying 12 PT per iteration.
• Pick Kimi K3 when: the problem is genuinely hard — math, analysis, multi-step reasoning — and you can wait for a deep-thinking pass. Best intelligence-per-PT on the platform.
• Reach for Claude Fable 5 when: you need the absolute ceiling. It leads the Artificial Analysis Intelligence Index at 60 and wins long-horizon agent work (56% on AA-Briefcase). It's 25 PT, on Pro ($23/mo) and up.
• Reach for GPT-5.6 Sol when: coding-agent benchmarks are the priority — Sol is #1 on Artificial Analysis' Coding Agent Index (80) at 12 PT, on every plan.

Three new models, zero new subscriptions

This is the third model-family launch on TulexAI in July alone — Grok 4.5, then GPT-5.6, now the completed Kimi lineup. That cadence is the argument for the platform: subscribing to one vendor means betting on who's best this month. On TulexAI, all three Kimi models sit next to Claude, GPT-5.6, Gemini and Grok — 50 models on one login — with a live token meter showing what each message costs before you send it. Run the same prompt through K3, Fable 5 and Sol, and pick the winner yourself.

Try Kimi K2.6, K2.7 Code and K3 free on TulexAI — no card, first prompts on us. Compare what one subscription replaces on the pricing page, or see how TulexAI stacks up as a ChatGPT alternative.

Frequently Asked Questions

What is the difference between Kimi K2.6, K2.7 Code and K3?

K2.6 is the cheap general-purpose model (2 PT) with thinking/non-thinking modes and text, image and video input. K2.7 Code (4 PT) is a specialist tuned for agentic coding workflows. K3 (7 PT) is the deep-reasoning flagship — Artificial Analysis Intelligence Index 57, #7 of 190 — with a 1M-token context window. All three are on every TulexAI plan.

How much do the Kimi models cost?

Via the Moonshot API: K2.6 is $0.95/$4.00 per million input/output tokens, K2.7 Code $1.90/$8.00 (cached input $0.38), and K3 $3.00/$15.00 (cached input $0.30). On TulexAI: 2 PT, 4 PT and 7 PT per message respectively, included in every plan from $11/mo.

Is Kimi K3 as good as Claude Fable 5?

Close, not equal. Artificial Analysis scores Fable 5 at 60 (the leader) and K3 at 57 (#7 of 190) — a 3-point gap at roughly 3.3× the blended API cost ($40 vs $12 per million at a standard chat mix). On TulexAI that's 25 PT (Pro plan and up) versus 7 PT (every plan). Fable also still leads long-horizon agent evals. For most hard problems, K3 is the value pick; for the hardest, Fable keeps the crown.

Should I use Kimi K2.7 Code or GPT-5.6 Sol for coding?

On neutral scoreboards, Sol — it leads Artificial Analysis' Coding Agent Index at 80, and costs 12 PT on TulexAI. K2.7 Code's case is economics: at 4 PT, long agentic sessions cost a third as much per message. A sensible pattern is K2.7 Code for iteration and Sol (or Claude Opus Code at 12 PT) when a task keeps failing.

Is Kimi K3 slow?

Yes, by design. Artificial Analysis measured 32.6 output tokens/sec (#147 of 190) and around 161 seconds to first token — it's a deep-reasoning model that thinks before it answers. Use K2.6 for quick turnarounds and save K3 for questions worth the wait.

Which TulexAI plans include the Kimi models?

All of them. K2.6, K2.7 Code and K3 are available on Basic ($11), Pro ($23), VIP ($46) and Elite ($99) — no tier gate, alongside 50 models in total. Your first prompts are free, no card required.