Open source AI models are 90% cheaper. So why is the bill the same?
Moonshot AI shipped Kimi K3 on July 16: 2.8 trillion parameters, a one-million-token context window, and weights that go public on July 27. The launch post calls it the largest open-weight model ever released.
The number that matters is the API price: $3 per million input tokens, $15 per million output. That is about five times what Kimi K2 cost, and within sight of Claude Opus at $5 and $25. For eighteen months the story of open source AI models was a price war, near-frontier quality at a tenth of frontier price. The new flagship just priced itself next to the frontier. The war at the top is over, and it is worth asking why your AI bill never noticed one was happening.
There are now two price lists
Look past the flagship and the rest of the open source AI models are still standing. These are July 2026 prices per million tokens, OpenRouter weighted averages where the model trades there, vendor list prices otherwise:
| Model | Input | Output | SWE-bench Verified |
|---|---|---|---|
| GPT-5.5 | $5.00 | $30.00 | 88.7% |
| Claude Opus 4.8 | $5.00 | $25.00 | 88.6% |
| Kimi K3 | $3.00 | $15.00 | vendor claims only, for now |
| GLM 5.2 | $0.45 | $3.31 | top open model on Artificial Analysis |
| MiniMax M3 | $0.10 | $1.21 | 80.5% |
| DeepSeek V4 Flash | $0.05 | $0.24 | 79.0% |
The bottom half of that table sits ten benchmark points behind and runs ten to a hundred times cheaper. OpenRouter’s June review of the open-weight market puts the capability gap at a stable three to six months, with no sign of the frontier pulling away.
Ten points is a real gap on hard, multi-hour work. It is not a gap most business software will ever see. Drafting, summarising, classifying, extracting, boilerplate code: on that workload the cheap tier of open source AI models is not almost as good. It is the same product, and the difference shows up on the invoice rather than in the output.
Is DeepSeek safe? The China question is a hosting question
Hundreds of people a month ask Google whether DeepSeek is safe, and the answer is that they are asking the wrong question. The right one: whose machine runs the inference?
Call DeepSeek’s own API and your prompts go to servers in China. Its privacy policy says data is stored there and may be used for training, and the UK’s National Cyber Security Centre has warned that anything you input falls under Chinese law. Do not put client data through that endpoint. That part is simple.
The weights, though, are the model. Download them, or rent them from a Western host like Fireworks, Together or DeepInfra at roughly double the first-party price, and nothing you send crosses a border. The same applies to the other Chinese open source AI models: Kimi, GLM, Qwen. Under GDPR the transfer problem disappears entirely when the weights run on EU servers. This is the thing open source AI models give you that a closed model never can: a choice of landlord.
The carve-out is real. The FY2026 US defence budget restricts DeepSeek across the Department of Defense and its contractors, Australia and Canada have banned it from government systems, and US lawmakers have opened inquiries into companies running Chinese models at all. If your customers are governments, defence, or regulated industries, provenance is a procurement gate and this section answers itself. Everyone else is making a hosting decision, not a geopolitical one.
One footnote on self-hosting. At K3’s 2.8 trillion parameters, Moonshot recommends 64 accelerators or more. “Run it yourself” at that scale means a datacentre. For everyone else the practical move is picking who hosts, not hosting.
Why the bill went up anyway
Token prices fell by about two-thirds over the past year, and enterprise AI bills went up. The FinOps Foundation reports 73% of organisations are over their AI budgets, some running three times over their full-year token budget by April. Uber reportedly burned through its entire 2026 AI coding budget in the first four months of the year. Falling AI token costs did not make AI cheaper, because consumption grew faster than prices fell.
A chatbot answers once. An agent plans, calls tools, retries, and re-reads its own reasoning, burning five to thirty times the tokens to finish one task. Reasoning models add a second multiplier: the thinking you never see is billed as output, three to five times the visible answer. Price per token is the wrong unit. Cost per successful task is the right one, and it is the number almost nobody puts in the spreadsheet before signing.
The expensive half is the checking
Every one of these systems still needs a human in the loop, and the human is the expensive line. Faros AI’s study of more than 10,000 developers found teams using AI merge 98% more pull requests, while review time per PR went up 91%. A 2026 study of legal AI measured 34.8 minutes of attorney review to bring one generated answer up to standard. Generation is nearly free. Verification is not.
That cuts the cheap-model arithmetic two ways. If the task is easy to check, a draft or a summary or a classification you can spot-check, the cheap model wins outright, because the checking costs the same either way. If being wrong is expensive to find, the model that fails less is worth a hundred times the token price, because salary dwarfs inference. Sending hard, hidden-failure work to the cheapest model is how you save $4,000 on tokens and spend $40,000 on review.
Where open source AI models still trail
Long-horizon agent work is the frontier’s actual case. On tasks measured in minutes, everyone scores near-perfect. On tasks measured in hours, open models trail the frontier by six to ten months, and the gap has grown since early 2025, a lag researchers expect to persist on compute and post-training costs alone. Short benchmarks say the race is over. The long ones say it is not. If your product depends on an agent finishing a four-hour job without quietly derailing, the premium models are still the only option, and the premium is the cheap part of that decision.
What we tell clients
Route by task. Most companies already do: multi-model setups are now the default in the majority of enterprise accounts, and organisations routing work across tiers pay a median $2.31 per million tokens against $18.40 for frontier-only. Cheap open source AI models for the bulk of work that is easy to verify. Frontier models for the agent jobs where a silent failure gets found by a client instead of a test. Then spend the difference on the part that was always expensive: defining what correct means for your workflow, and checking against it.
That is also why the pilot failure statistics read the way they do. MIT’s 2025 study found 95% of enterprise AI pilots show no measurable impact on the P&L, and RAND’s post-mortems put 73% of failures down to one cause: nobody agreed what success was before the work started. The model was never the expensive part. The workflow was, and it still is. We made the same argument about what an AI automation agency actually builds: the prompt is the demo, and the system underneath is the product.
We build both tiers into client systems and measure cost per successful task, because what software costs is a question about the whole lifecycle, not the sticker. If your AI budget line keeps growing, the fix is almost never a different logo on the invoice. It is a routing table and an eval suite, and building both is most of what we do.
None of this argues against open source AI models. Most of what we ship runs on them. It argues for reading the price list as two lists, because the labs already do. The labs will keep arguing about who is ahead. Your P&L only cares which tasks pass.