Kimi K3 Sold Out: Why Moonshot Paused New Subscriptions

Last updated: July 20, 2026

Kimi K3 did not just launch as another open model benchmark story. Within days of the model getting attention, Moonshot AI paused new Kimi subscriptions because demand pushed its GPU capacity close to the limit, according to Moonshot’s official X statement, also mirrored on LinkedIn. The visible result for users was simple: Kimi’s paid subscription cards showed Sold out.

Quick answer: Kimi K3 is not discontinued, and existing subscribers are not being removed. Moonshot says on X it is temporarily pausing new subscriptions, prioritizing compute for current members, adding capacity, and planning to reopen subscription spots in batches. The company also says it will split membership into Kimi Membership for web, app, and work use, and Kimi Code Membership for coding workflows.

That matters because Kimi K3 is one of the most important AI launches of the year: Moonshot describes it as a 2.8-trillion-parameter, 1-million-token, native multimodal model that sits near the top proprietary frontier. In benchmark terms, Kimi K3 is not a clean winner over every OpenAI and Anthropic model. But it is close enough to make the market pay attention, and in several coding and agentic tests it beats or matches models developers already treat as premium.

Source check — July 20, 2026: this article checks Moonshot’s official Kimi K3 technical blog, Kimi K3 API quickstart, Kimi API pricing page, Moonshot’s official subscription-pause statement on X, mirrored on LinkedIn, AP’s report on the subscription halt, South China Morning Post, Artificial Analysis, Vals AI, OpenAI’s GPT-5.6 page, and Anthropic’s Claude pricing and model overview. Because model access and AI subscriptions can change quickly, verify the live Kimi pricing page before making a purchase decision.

For related model context, compare GPT-5.6 Sol, GPT-5.6 vs Claude Fable 5, Claude Opus 4.8, Claude Fable 5 vs Opus 4.8, and Claude Code vs Codex.

What Happened to Kimi Subscriptions?

Moonshot’s public message is straightforward: Kimi K3 got more demand than the company expected, and the compute behind the product became the bottleneck, as its official X statement, LinkedIn mirror, and AP’s subscription-halt report all indicate.

In its official X statement, Moonshot said Kimi K3 had received more attention than expected, that demand over the previous 48 hours had pushed close to current capacity, and that it would pause new subscriptions to protect the experience of existing subscribers. It also said existing subscribed users would not be affected and that new subscription spots would reopen in batches as capacity comes online.

The important nuance is that sold out does not mean Kimi K3 is gone. It means the consumer subscription funnel has been gated while Moonshot tries to preserve service quality. That is different from an outage, a model rollback, or a cancellation of paid accounts.

Practically, there are now three different availability states to keep separate:

This is also why Kimi’s next membership split matters. A general web user, a heavy Kimi Work user, and a developer burning through long coding trajectories do not consume compute in the same way. By separating Kimi Membership from Kimi Code Membership, Moonshot can price, queue, and allocate GPU capacity with more precision.

Why Did Kimi K3 Sell Out So Fast?

There are three reasons the sold-out status happened quickly.

First, Kimi K3 is computationally heavy. Moonshot describes it as a 2.8T-parameter model with a 1M-token context window, native visual understanding, and always-on reasoning. Its Mixture-of-Experts design activates only part of the network per token, but this is still a giant model serving long-context, agentic workflows.

Second, the launch narrative was unusually strong. Moonshot framed Kimi K3 as the first open 3T-class model and said its overall performance still trails Claude Fable 5 and GPT-5.6 Sol while outperforming other tested models across its evaluation suite. That is exactly the kind of positioning that attracts developers, benchmark watchers, investors, and AI power users at the same time.

Third, coding demand is expensive demand. A normal chatbot session might be a few prompts. A Kimi Code session can inspect repositories, call tools, run terminal loops, read screenshots, debug, and continue for a long horizon. One serious coding user can consume far more inference than many casual users.

That makes the sold-out page less surprising. Kimi K3’s hype is not only about people asking it a few questions. It is about users trying to push it through long, tool-heavy agent runs where the model’s strengths are supposed to show.

Kimi K3 vs Claude Opus 4.8 and GPT-5.6 Sol

The cleanest way to read the early evidence from Artificial Analysis, Vals AI, OpenAI, and Anthropic is this:

• GPT-5.6 Sol and Claude Fable 5 still look stronger overall across many frontier and reasoning-heavy comparisons.
• Kimi K3 is now in the top group, not the open-model second tier.
• Claude Opus 4.8 is no longer the obvious step above open alternatives for every coding or agentic use case.
• Kimi K3’s cost advantage is real against Opus 4.8, but not as dramatic as earlier cheap Chinese model stories.

Where Kimi K3 looks strongest

Kimi K3’s strongest early story is coding plus agentic work, especially frontend and long-horizon tasks.

On Vals AI’s Vibe Code Bench v1.1, Claude Fable 5 leads at 90.35%, but Kimi K3 takes second place at 84.96%, ahead of Claude Opus 4.8 at 82.72% and GPT-5.6 Sol at 80.50%. That benchmark asks whether models can build working web applications from scratch, which is close to how many people actually use coding agents.

On Artificial Analysis, Kimi K3 reaches 57 on the Intelligence Index and is described as comparable to Opus 4.8 and GPT-5.5 while still behind Fable 5 and GPT-5.6 Sol. Its GDPval-AA v2 score of 1668 also beats Claude Opus 4.8’s 1600 in that report, while remaining behind Claude Fable 5.

For developers, that is the point: Kimi K3 does not need to win every benchmark to matter. If it can beat Opus 4.8 on some coding and agentic workflows while costing less per output token, it becomes a serious routing option.

Where GPT-5.6 Sol still looks safer

GPT-5.6 Sol still has the cleaner case for teams that want a top proprietary frontier model with broad product integration. OpenAI positions GPT-5.6 Sol as the flagship GPT-5.6 model for complex professional work, and public benchmark snapshots still show it leading Kimi K3 in several coding and terminal categories.

For example, Vals AI’s benchmark overview lists GPT-5.6 Sol first on SWE-bench Verified at 96.2%, with Kimi K3 third at 93.4%. It also lists GPT-5.6 Sol first on Terminal-Bench 2.1 at 85.8%, with Kimi K3 second at 80.9%.

The tradeoff is price. OpenAI lists GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens. Kimi K3’s first-party API output price is $15 per million tokens, which is half of Sol’s listed output price. But task-level cost can narrow because reasoning models may use tokens differently. Artificial Analysis estimates Kimi K3 at $0.94 per Intelligence Index task versus $1.04 for GPT-5.6 Sol, which is close, not a 10x difference.

Where Claude Opus 4.8 still fits

Claude Opus 4.8 still has a role. Anthropic describes Claude Opus 4.8 as a model for complex agentic coding and enterprise work, with 1M context, 128k max output, adaptive thinking, and broad availability through Claude API, AWS, Google Cloud, and Microsoft Foundry.

The case for Opus 4.8 is less about being the cheapest or winning every new benchmark. It is about enterprise deployment, Claude ecosystem maturity, governance, and predictable integration. If your team already depends on Claude Code, Anthropic contracts, or Anthropic’s cloud routes, Opus remains a practical default.

Kimi K3 pressures that default because it is close in intelligence and sometimes ahead in frontend or agentic work. But the subscription sold-out moment is also a reminder: a model can be impressive and still not have enough consumer capacity for every new user at once.

Is Kimi K3 Actually Open Source?

Moonshot uses open-source/open 3T-class language in its launch materials, but the careful term right now is open-weight roadmap.

As of this July 20 check, Moonshot says the full Kimi K3 model weights will be released by July 27, 2026. Until those weights are public and independently tested, many of the strongest Kimi K3 claims should be treated as early evidence rather than settled fact.

That does not make the launch unimportant. It simply changes how to evaluate it:

• The hosted Kimi K3 product is live but capacity constrained.
• The API is priced and documented.
• The weights are planned, not fully available at the time of writing.
• Independent benchmarkers have early results, but full community testing will expand after the weight release.

If Moonshot delivers the weights as planned, Kimi K3 could become the leading open-weight model by capability. If the release is delayed, heavily restricted, or difficult to run in practice, the open-model narrative will need updating.

What Kimi K3 Sold Out Means for Users

For users, the sold-out status is mostly a planning issue.

If you already have a Kimi subscription, Moonshot says you should not lose access because of the pause. You should still watch for quota changes, speed changes, or plan migration details as the company splits membership categories.

If you do not have a subscription, the checkout page may not let you buy a plan until Moonshot reopens spots. That does not necessarily block all evaluation. You can still check API access, wait for batch reopenings, monitor Kimi Code Membership details, or compare alternatives.

If you are choosing a model for production work, do not treat a sold-out subscription launch as proof that Kimi is either doomed or unbeatable. Treat it as evidence of demand plus compute scarcity.

What to Do If Kimi Is Sold Out

If you wanted to buy Kimi and now see Sold out, the practical next step depends on your use case.

Do not overfit to the subscription page alone. A consumer membership sold out can coexist with API availability, enterprise allocation, future batch reopenings, and open-weight downloads. The question is not only “can I click subscribe?” It is “can this model serve my workload reliably at the cost and governance level I need?”

Final Verdict: Kimi K3 Is Real, but Capacity Is the Story Now

Kimi K3 is a serious model launch. It has the size, context window, coding benchmarks, and open-weight roadmap to make developers and AI teams pay attention. The strongest evidence so far suggests it is competitive with Claude Opus 4.8 and GPT-5.6 Sol in several important coding and agentic categories, while still trailing the very top proprietary systems overall.

The sold-out subscription page adds a second story: frontier capability is only valuable if users can actually access it. Moonshot appears to be choosing a conservative route by protecting existing subscribers rather than selling unlimited new plans into limited GPU capacity.

That is frustrating for new users, but strategically understandable. If Kimi K3 keeps its benchmark momentum and Moonshot adds enough compute, the sold-out moment may become part of the launch story: the point where an open-weight Chinese frontier model became popular enough to run into the same capacity problems as the largest U.S. AI products.

For now, the right answer is balanced: Kimi K3 is worth testing, not blindly replacing everything with. Wait for subscription batches if you want the consumer product, test the API if you are a developer, and revisit the benchmark picture after the planned July 27 weight release.

FAQ