1300 166 138 Tell us what you need
News / Open Source Models

Kimi K3 rivals Claude Opus 4.8 for a fraction of the price

By Monika Tantau • Published 4 months ago
Kimi K3 rivals Claude Opus 4.8 for a fraction of the price

Key takeaways

• Moonshot released Kimi K3 as open weights: 2.8 trillion parameters, 104 billion active, one-million-token context
• It scores 88.3 on Terminal-Bench 2.1 and 91.2 on BrowseComp, close to Claude Opus 4.8
• Telnyx lists it at $2.70 per million input tokens and $13.50 per million output
• Anthropic says it has never sought a ban and wants mandatory safety testing instead

Most Read

Moonshot AI has published the weights for Kimi K3, a 2.8 trillion parameter model that lands within striking distance of the closed frontier on coding and agentic benchmarks. It reached the top of Hacker News inside a day. Anyone with the hardware, or a provider willing to host it, can now run it.

The timing is the interesting part. A day earlier, Anthropic published its position on open-weights models, stating it has never advocated banning them and describing capable models without dangerous capabilities as a public good. Two announcements, one question for Australian operators: has a frontier-class model you control just become a realistic option?

What Moonshot actually released

Kimi K3 is a mixture-of-experts model with 2.8 trillion total parameters spread across 896 experts, of which roughly 104 billion activate per token. It carries a one-million-token context window, built on what Moonshot calls Kimi Delta Attention and Attention Residuals. The specifications and benchmark tables are published on the Hugging Face model card.

The scores are the reason it travelled. Kimi K3 records 88.3 on Terminal-Bench 2.1, 91.2 on BrowseComp and 84.8 on OSWorld-Verified, with 67.5 on DeepSWE and 77.8 on ProgramBench. Across most of those it sits near or above Claude Opus 4.8 and GPT-5.5, though GPT-5.6 Sol still leads on several coding tasks. AutomationBench, at 30.8, is a reminder that agentic reliability remains the hard part for every model in this class.

One caveat matters for procurement. The weights ship under a custom Kimi K3 License, not MIT or Apache. This is an open-weights release, not open source in the strict sense, and the licence terms need reading before anything touches a commercial workload.

The cost picture and its limits

Telnyx, one of the inference providers that listed the model on 28 July, prices it at $2.70 per million input tokens and $13.50 per million output, with cached input at $0.27. That is materially below frontier closed-model pricing for work of comparable quality, and it is the number that will get the model onto shortlists.

The self-hosting story is more complicated than the licence suggests. A 2.8 trillion parameter model is not something a mid-market firm stands up in a server cupboard, even with only 104 billion parameters active per token. The active count governs how fast it runs; the total count governs how much memory has to hold the weights before it runs at all. Those are different bills, and the second one is the one that rules out most on-premise plans.

For most Australian businesses the practical route is therefore a hosted provider, which means the sovereignty argument depends entirely on where that provider runs its hardware. An open licence on a model served from a US region gives you a different legal position to the same model served from Sydney, and the difference is not visible in the pricing table. Teams weighing this against the incumbents will find the frontier context in our breakdown of Claude Opus 5 versus Fable 5.

Anthropic draws its own line

On 27 July, Anthropic set out its position on open-weights models, opening with the statement that it has never advocated for a ban on them. Models released openly without dangerous capabilities, the company argues, are a public good.

What it asks for instead is narrower and more specific. Restrict sales of chips and chipmaking equipment to China and pursue smuggling. Target industrial-scale distillation through legal and commercial means, including identifying and banning the accounts doing it on Anthropic’s own platform. And apply mandatory safety testing to every sufficiently capable model, whether the weights are open or closed.

That third ask is the one with commercial consequences. It draws the line at capability rather than at licensing, which means a test applied to Kimi K3 would apply to Claude on the same terms. For a business choosing between them, the practical question stops being open versus closed and becomes which vendor can evidence what it tested.

Anthropic draws its own line

Why Australian businesses should care

Open weights remain the only path that lets an organisation hold both the model and the data without a vendor contract sitting between them. For workloads touching health records, legal files or anything the Privacy Act 1988 makes you responsible for, that is not an ideological preference but a compliance one. The trade-offs are the same ones we covered in our guide to AI data residency and the CLOUD Act.

The risk sits in the timing. US legislators are already drafting rules for capable models, and Anthropic’s mandatory-testing proposal points the same way. A model selected this quarter on price and benchmark scores may be operating under a different compliance regime by the time it reaches production. Australian firms building on any frontier model, open or closed, should be treating regulatory movement as a live variable in the business case rather than a footnote, which is where AI strategy work earns its keep.

What to watch next

Two signals are worth tracking over the coming weeks. The first is whether Moonshot publishes safety evaluations to match its benchmark tables, because that is the gap Anthropic’s proposal is aimed at. The second is whether Australia’s voluntary AI safety standard starts moving toward the mandatory testing being discussed in Washington.

For teams tempted to act now, the sensible sequence is unglamorous. Run Kimi K3 against a real workload as a second opinion or a batch job, measure it on your own data rather than on published benchmarks, and keep the core process on a model whose support and legal position you can already account for.

The gap between open and closed models has narrowed to the point where the decision is no longer technical. It is a question of what your organisation can defend if someone asks.

Scroll to Top