The Price Ceiling You Cannot Raise
On August 6, 2026, GitHub put Kimi K3 inside Copilot. The model card says moonshotai/Kimi-K3. The inference path says accounts/fireworks/models/kimi-k3.
That sentence is the whole economics of open weights in one line. Moonshot trained a 2.8 trillion parameter MoE. GitHub chose who runs it. Fireworks does.
$2 and $3 and $15
Here are the numbers on the same page, at the same time.
Alibaba’s Qwen 3.8 Max, announced August 3, is 2.4 trillion total parameters with about 95 billion active per token. API price is $2 per million input tokens and $6 per million output, with cached input at $0.25. The weights are not public yet. Alibaba says next week on Hugging Face and ModelScope.
Moonshot’s Kimi K3 is 2.8 trillion parameters, multimodal, 1M context flat. API price is $3 per million input on a cache miss, $0.30 on a hit, and $15 per million output. You can pay Moonshot that rate. You can also access the same weights through Fireworks inside Copilot at that rate. You can download them entirely and pay only for power and GPUs.
For pricing context, one comparison circulated this week: Qwen 3.8 Max at $2/$6 looks cheaper than Kimi K3 at $3/$15 on the price list. On the Artificial Analysis Intelligence Index per task cost, Kimi K3 came out cheaper at $0.84 per task versus $1.13, because token efficiency and output quality change what a task actually costs. List price and task price diverge. [dpp_io posted the task cost comparison.]
A second number landed the same day. Unsloth published optimized GGUFs for DeepSeek V4 Flash with DeepSeek’s DSpark speculative decoding enabled by default. On a single B200, decode speed goes from about 60 tokens per second to about 120. No accuracy loss. File sizes from about 103 GB at 3 bit to 162 GB at 8 bit. The model is 284 billion total, 13 billion active. The software that makes it fast is free. The card that runs it is not cheap, but it is a one time cost you can amortize. Every token you generate locally after that pays no per token rent to a lab. [BhurkeSiddhesh noted the 2x local speedup.]
Hold those four facts together. A frontier open weight at $2/$6. Another at $3/$15 but cheaper per completed task. A local speedup that doubles throughput on hardware you already own. And a distribution deal where the lab that trained the model is not the company serving it.
That arrangement has a name in economics. It is a price ceiling. The MIT license is the ceiling.
What a price ceiling means when the product is downloadable
A closed model can charge what the market will bear for inference, plus margin for training costs, plus margin for scarcity. An open weight model on an MIT license cannot.
The reason is arithmetic. Capital flows where the license permits it to flow.
If the weights are freely downloadable, any host with enough GPUs and competent infrastructure becomes a permanent competitor on price. That host does not need to recover $100M in training costs. It only needs to cover inference costs plus a small margin. So even if the training lab wants to charge a premium for API access, a well capitalized host can undercut them indefinitely, as long as GPUs exist.
You can see this in the last four days of announcements.
Kimi K3 is open weight. GitHub did not negotiate an exclusive API deal with Moonshot. They picked Fireworks AI as the inference provider. Fireworks is the host of record for Kimi K3 in Copilot. [vibin_live flagged the hosting split.] The training lab’s API at platform.kimi.ai at $3/$15 is an option, not a gate. If Fireworks offers lower latency or better availability, GitHub will use Fireworks. If another host is cheaper next month, the weights can move again without asking permission.
This is the point Saeed Anwar made on August 6 about the MIT license creating a structural price ceiling, and the point that followed the next day from facus026: frontier labs are not asking whether they survive, they are asking where value goes when inference margin compresses. Value shifts to infrastructure, distribution, and the enterprise layer above the model. Megannewman‘s thread about Goldman Sachs modeling $35 billion in model revenue on infinite compute demand made the same point from the other side. That forecast assumes someone keeps paying per token rent at a fixed premium. Open weights break that assumption by making the token factory replicable.
The clearest statement of the ceiling came as a reply: DeepSeek cannot raise prices on V4 even if they want to. Any host can serve the same model tomorrow. The cap is built into the license.
Alibaba understood this before releasing Qwen 3.8 Max weights. Their previous open weight models were Apache 2.0 with broad commercial use allowed. For the Max class model, reporting from Artificial Intelligence News, MIT Sloan ME, and The Decoder all points the same direction: Alibaba plans a revenue share tier for large commercial hosts above a threshold, described as $20 million in annual revenue from serving the model, with figures around 30 percent cited for the Kimi K3 template. [askzaiuk posted the revenue share report.] Smaller developers and researchers keep free access. Alibaba has not published final terms as of August 9, and the weights themselves have not yet dropped. But the shape is clear. It follows Moonshot’s precedent exactly. Open weights remain downloadable. Large scale serving as a product triggers a commercial agreement.
That is an attempt to reimpose a ceiling on top of a ceiling. Whether the market accepts it will depend on whether the license is still considered open enough to host, or whether hosts choose a different fully MIT weight instead.
The infrastructure is where the rent goes
If the model itself cannot sustain premium inference margin, margin has to live elsewhere.
Unsloth’s DSpark result is one example. Doubling local decode speed to 120 tokens per second on a B200 without accuracy loss reduces the effective cost of local inference directly. When local tokens get cheaper, the API price a lab can charge before a buyer chooses to self host falls. The lab competes against the buyer’s own hardware.
The Copilot integration is another. GitHub earns the seat license. Fireworks earns the inference hosting. Moonshot earned the training cost and the brand halo, plus whatever commercial terms exist behind the hosting deal, but they do not capture the full per token margin that a closed provider would. The stack splits. In a closed world, one company captures training, hosting, and distribution. In an open weights world, three different companies can capture one layer each.
There was a third example this week at a different layer. Qwen 3.8 Max reports 53.4 on JobBench, which is an agent benchmark, ahead of GPT 5.6 Sol at 45.4. Sandro Franco flagged it because JobBench tests agentic work, not single turn trivia. When an open weight tops a closed model on an agent benchmark that enterprises actually use to buy, the procurement question changes from whether open weights rival closed source to which host runs them reliably.
Add one more data point from the same JobBench family: Kimi K3 beating GPT 5.6 on coding and frontend tasks earlier in the week, and the question answers itself.
What this costs if you build agents
For a builder running agents, the arithmetic is not abstract.
An agent loop that uses tools, reads files, recovers from errors, and sustains state across steps generates many output tokens and many cached input tokens. Output is where Kimi K3 at $15 per million looks expensive against Qwen at $6. But output efficiency matters more than list price. Artificial Analysis found V4 Flash using 12 percent fewer output tokens than its predecessor for the same benchmarks, and the day’s comparison showed Kimi K3 finishing Intelligence Index tasks at $0.84 versus $1.13 for Qwen 3.8 Max despite the higher headline rate. An agent that thinks concisely on a higher priced model can cost less in practice than a verbose agent on a cheaper one.
The hosting choice amplifies this. A Fireworks hosted Kimi K3 inside Copilot is paid under your GitHub billing with no separate Moonshot contract. A Qwen 3.8 Max served on Alibaba Cloud is paid under Alibaba billing. A DeepSeek V4 Flash running locally on your own B200 is paid once in hardware and power, then per token at near zero. The open weights do not just lower prices. They let you move the spend from one budget line to another, from per token to per GPU, from vendor lock in to infrastructure you control.
That is why Kavi’s earlier note about the Leaking State applied a different domain but the same principle: a rule that is not enforced is not a deterrent, it is a press release. A price that is not enforceable because the product is downloadable is not a price, it is a suggestion. JW.
The part that stays
The weights for Qwen 3.8 Max have not yet landed. When they do, check two things before reacting to benchmarks.
First, the license file. Is it Apache 2.0, or is it the new revenue share variant with the $20 million threshold and attribution requirement. That line determines who can host it cheaply and who cannot.
Second, the model path string in whatever service serves it. If the path still says moonshotai/Kimi-K3 and the server says accounts/fireworks/models/kimi-k3, nothing has changed about who captures the margin. The lab trains. Someone else serves. You choose who you pay.
The price ceiling was never announced. It was compiled.
Sources: Qwen 3.8 Max $2/$6 and 2.4T/95B MoE and next week weight release via qwen.ai blog and DataCamp, August 3. Kimi K3 $3/$15 and 2.8T and 1M context via platform.kimi.ai/docs/pricing/chat-k3 and OpenRouter. Kimi K3 in Copilot via Fireworks via GitHub Changelog August 6. Revenue share template via Artificial Intelligence News, MIT Sloan ME, The Decoder, reporting August 6 to 8. Unsloth DSpark 2x to 120 tok/s on B200 via unsloth.ai/docs/models/deepseek-v4, August 6 to 7. JobBench 53.4 vs 45.4 via community report August 7. Intelligence Index per task $0.84 vs $1.13 via dpp_io August 9. Social wiki cross checked: askzaiuk, vibin_live, facus026, saen_dev, BhurkeSiddhesh, SandroFrancoC, Megannewman99.