Kimi K3: How $0.40/M Tokens Changes the Unit Economics of AI-First Products

Kimi K3: How $0.40/M Tokens Changes the Unit Economics of AI-First Products

Six months ago, I sat down with a spreadsheet to cost out a feature I wanted to build: an AI layer that would sit inside every customer portal page, offering contextual summaries, smart search, and proactive recommendations based on what the user was looking at. The idea was simple. The maths was not.

At GPT-4-level pricing — about $15 to $30 per million output tokens at the time — a single user session generating 20 interactions with 2,000-token responses would cost roughly $0.60 to $1.20. For 1,000 active users per month, that's $600 to $1,200 in inference costs alone, before you've paid for anything else. For 10,000 users, you're looking at a line item that eats your margin before lunch.

I shelved the feature.

Last week, I pulled the spreadsheet back out. The numbers had changed.

The new denominator

Kimi K3 is the headline, but the number that matters for product builders isn't the model size or the benchmark scores. It's the pricing trajectory of the Kimi family on third-party inference providers.

Moonshot AI's K3 launched 16 July at $3 per million input tokens and $15 per million output on the official API, already a meaningful discount against GPT-5.6 Sol. But the real story is what happens when open weights hit Hugging Face. The previous generation, Kimi K2.6, is already served by providers like DeepInfra at $0.75 per million input tokens, with cached input as low as $0.15 per million. The K2 model family bottoms out at $0.40 per million cached input on DeepInfra. Once K3 weights drop and providers absorb them, the effective cost of frontier-level inference converges toward that band.

Let me put that in plain terms. You can now run AI inference at roughly one-fiftieth to one-seventy-fifth the cost of what the same capability would have run you in early 2024. Not "slightly cheaper." Not "budget tier." A genuine collapse in the unit economics of AI-powered features.

This isn't an infrastructure story. It's a product story. And it changes which products you can build.

The feature graveyard: what you couldn't ship at old prices

Every builder has a list of features they designed but didn't build because the maths didn't work. Here are three categories that just crossed the viability threshold.

Real-time AI personalisation. The idea of running an inference call every time a user loads a page — generating personalised content, adapting the interface, suggesting next actions — has been the holy grail of AI-first products since GPT-3. The blocker was never the capability. It was the cost. At $30 per million output tokens, serving personalised responses to 100,000 page views per month would cost thousands of dollars. At $0.40 per million input and $2.00 per million output — the K2 family floor on third-party providers — those same 100,000 personalised interactions drop to somewhere between $50 and $150 per month. That's not a premium feature anymore. That's a default.

Bulk content analysis and enrichment. Processing a library of 50,000 documents, extracting structured data, generating summaries, and tagging content with AI: at GPT-4 prices, that's a project with a five-figure inference bill before you've built the product around it. At open-model inference rates, the same job costs $40 to $200. This changes what's possible for internal tools, knowledge bases, and data products. You can now afford to run AI over your entire document corpus, not just the high-value subset.

AI-powered customer portals. The feature I shelved. A customer portal where every page has an AI layer — contextual summaries of what's on screen, proactive recommendations, natural language search over all your account data, document Q&A — was a premium product concept. At $0.40-per-million rates, it's a feature you can ship to every customer tier, not just enterprise. The per-user cost drops from "noticeable" to "rounding error."

The maths, concretely

Let's walk through a real product scenario. You're building a no-code customer portal for a B2B SaaS company. The portal includes:

• AI-powered search: natural language queries over help docs, invoices, and account history, averaging 5 searches per user per month, each consuming 3,000 input tokens and 500 output tokens
• Smart summaries: every page generates a contextual summary of the data on screen, 10 page loads per user per month, 2,000 input tokens, 300 output tokens each
• Proactive alerts: the AI checks for anomalies in account data and surfaces insights, 3 checks per user per month, 1,000 input tokens, 400 output tokens each

Per user, per month, that's:

• Search: 5 × (3,000 + 500) = 17,500 tokens
• Summaries: 10 × (2,000 + 300) = 23,000 tokens
• Alerts: 3 × (1,000 + 400) = 4,200 tokens
• Total: roughly 44,700 tokens per user per month

At GPT-4 rates (early 2024): roughly $1.30 per user per month. At 1,000 users, that's $1,300/month in inference. At 10,000 users, $13,000/month — a serious line item.

At K2 family rates on third-party inference (mid-2026, cached): roughly $0.04 per user per month. At 1,000 users, $40. At 10,000 users, $400.

The feature didn't get 10% better. The economics inverted by roughly 30x. When your per-user AI cost drops below what you spend on hosting, the decision to ship stops being financial and starts being purely about whether the feature adds value.

Why this changes what no-code builders can offer

The no-code ecosystem has a particular stake in this shift. No-code platforms traditionally compete on speed and accessibility, not on raw AI capability. When the best AI models cost $30 per million tokens, embedding them deeply into a no-code platform's feature set means either passing a visible cost to the user or absorbing it and hoping usage stays low.

At $0.40 per million, that trade-off disappears. A no-code platform can offer AI-powered search, contextual help, smart form filling, and document analysis as platform features — not premium add-ons — because the marginal cost is negligible. The platform's value proposition shifts from "we help you build faster" to "we help you build smarter, and the AI is included."

The builders using those platforms see the same effect downstream. If you're building a client portal, an internal tool, or a workflow app, you can now treat AI as infrastructure — like a database query or a file upload — rather than a metered luxury. That's a category change, not a cost optimisation.

The catch: output tokens still cost

The pricing collapse is real, but it's uneven. Input tokens have cratered — $0.40 to $0.75 per million on open models, with cache hits dropping to $0.15. Output tokens remain more expensive by a factor of 4x to 6x. For K2.6 on DeepInfra, output is $3.50 per million versus $0.75 input. For K3 on Moonshot's own API, output is $15 per million versus $3 input.

This asymmetry rewards product designs that minimise output tokens: structured responses, short summaries, classification rather than generation. A feature that returns a JSON object with five fields costs a fraction of one that returns three paragraphs of natural language. As inference gets cheaper, the discipline of designing for token efficiency doesn't go away. It becomes a competitive advantage.

The takeaway

I reopened my spreadsheet this week and the feature I shelved six months ago now costs about the same per user as a CDN bandwidth bill. I'm building it.

If you have a feature in your backlog that you designed but never costed out — or costed out and abandoned — now is the time to rerun the numbers. Use K2.6's third-party rates as a floor ($0.40 to $0.75 per million input) and K3's official API as a ceiling ($3 input, $0.30 cached). Even at the ceiling, the maths has shifted. At the floor, entire product categories become defaults.

The $0.40-per-million-token world isn't coming. It's here. The question is which builder ships the product that was impossible last quarter.