>_TheQuery
← All Articles

The Cheap Model Is Winning

By Addy · April 15, 2026

On April 13, Moonshot AI confirmed via email to Kimi Code beta testers that the model they had been using for the previous week was k2.6-code-preview.

The confirmation was quiet. No benchmark deck. No launch keynote. No theatrical reveal. Just an email telling developers that what they had already noticed - that Kimi Code had quietly gotten better - was real.

Six days earlier, on April 7, Anthropic announced Claude Mythos Preview alongside Project Glasswing, a coalition of twelve major technology organizations, $100 million in usage credits, and a price of $25 per million input tokens and $125 per million output tokens.

That announcement was loud.

And almost nobody can use it.

These two events describe the same market from opposite ends.

One model is cheap enough to become infrastructure. The other is strong enough to become policy.

The Numbers, Because They Explain Everything

Anthropic's own pricing for Claude Opus 4.6 remains $5 per million input tokens and $25 per million output tokens. That is already premium pricing by current market standards.

Claude Mythos Preview, the model Anthropic says is its most capable yet for coding and agentic work, is priced at $25 input and $125 output. It is not generally available. Anthropic says it is available only to Project Glasswing participants and additional research-preview partners - roughly fifty organizations in total once you combine the twelve named launch partners with the forty critical-software groups Anthropic says it is also granting access.

Moonshot AI's official API pricing is the opposite kind of signal. On the Kimi API platform, Kimi K2 is listed at $0.60 per million input tokens and $2.50 per million output tokens. Kimi K2.5, the newer multimodal model, is listed at $0.60 input and $3.00 output. The weights for K2.5 are publicly available under Moonshot's Modified MIT license.

That means the open model you can actually deploy today is roughly 8x cheaper than Opus 4.6 on input and between 8x and 10x cheaper on output, depending on which Kimi model you compare. Against Mythos Preview, the gap is much wider: about 42x cheaper on input and roughly 42x to 50x cheaper on output.

If you process 100 million input tokens and 100 million output tokens in a month, the difference is not subtle. Kimi K2 comes out around $310. Mythos Preview comes out around $15,000.

That is not a price difference.

That is a different category of buyer.

What K2.6 Actually Is

K2.6 is not a normal launch.

Moonshot has not published weights, a technical report, or an official benchmark card for k2.6-code-preview. What exists right now is a reported rollout email to beta users, plus the observable fact that users inside Kimi Code started noticing better behavior before Moonshot named the model publicly.

That matters because it tells you what Moonshot is optimizing for.

Not spectacle. Iteration.

The official Kimi material we do have gives the baseline. Kimi K2 is a 1 trillion parameter Mixture of Experts foundation model with 32 billion active parameters per token. K2.5 extended that line into a multimodal system and posted a 76.8% score on SWE-bench Verified in Moonshot's own technical report, versus Anthropic Opus 4.6 at 80.8% on Anthropic's published Glasswing comparison.

That gap is real. But it is not 10x real. It is not 40x real. And it is not "you must apply for access" real.

K2.6 appears to be Moonshot's attempt to narrow the coding-agent gap further before publishing the formal benchmark story. Until official evals land, the honest statement is narrower: Kimi Code users got a better model before they got a launch post.

That alone tells you a lot about where the company thinks competition actually happens.

The Frontier That Nobody Can Reach

Mythos Preview may well be the strongest coding and cybersecurity model currently in existence.

Anthropic's own Glasswing materials say Mythos Preview found thousands of high-severity zero-day vulnerabilities across critical infrastructure, including every major operating system and web browser. Anthropic's published benchmark table puts it at 93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro, and 82.0% on Terminal-Bench 2.0 - all above Opus 4.6.

Those are extraordinary numbers.

But the market does not buy benchmark tables. It buys access.

And access, in this case, is tightly controlled. Anthropic's explanation is coherent: a model this capable at vulnerability discovery and exploitation should go to defenders first, not the general public. That is a serious argument, and Anthropic has backed it with unusually detailed red-team documentation.

But the economics are sitting in the room with the safety story. Mythos is not just gated. It is expensive enough that even if it were opened up tomorrow, its natural customer base would still be concentrated in enterprise and government buyers with specific budgets and narrow use cases.

This is the part worth naming clearly.

The model that is strongest is also the model that is hardest to reach, hardest to budget for, and hardest to build a mass-market product around.

That is not irrelevant. That is the whole market question.

Is It Safety, or Just Strategy?

The skeptical reading of all this is easy: every frontier lab eventually describes its best model as too dangerous to release right around the time the pricing becomes enterprise-only.

That reading is too cynical to be fully fair.

Anthropic's cybersecurity case for gating Mythos is more specific than the usual hand-waving. The company has published exploit-chain examples, benchmark methodology, and an explicit rationale for trying to get defenders ahead of attackers by a few months. Those concerns are not invented.

But the strategic convenience is real too. A model priced at $125 per million output tokens was never going to be the mass-market default. Gating it narrows the market in a way that is simultaneously safer and commercially legible.

These two things can both be true.

Mythos can be genuinely dangerous. Mythos can also be naturally aligned with a high-end enterprise business model.

The open question is not whether Anthropic is lying. It is whether the market cares enough about capability at the very top to tolerate that combination for long.

So far, software history suggests the answer is: only for a while.

The Cheap Model Usually Wins for a Reason

There is a pattern here that frontier labs do not love talking about.

The open alternative does not need to be better in absolute terms. It needs to be good enough, cheap enough, and available enough that developers can build around it without asking permission.

That is how Linux beat commercial Unix. That is how PostgreSQL outlasted a long list of proprietary database assumptions. That is how Chromium became the browser substrate of the modern web, including for companies that spent years trying to protect a closed moat around the client.

AI is not identical to any of those markets. Training a frontier model is more capital-intensive than spinning up an open-source database project. But deployment is a different layer of the stack.

Once the weights exist, the market stops asking only who trained the best model. It starts asking who can actually afford to use one, fine-tune one, self-host one, and build product around one without waiting for access approval.

That is where Kimi's line matters.

The weights exist. The APIs are cheap. The licensing is permissive enough for broad use. The model is already integrated into real developer tooling. And the company is iterating in public fast enough that the benchmark gap is narrowing before the press cycle has time to catch up.

That is what winning looks like in software before everyone admits it is happening.

Where This Goes

Mythos Preview is a real capability jump. But capability that only a few dozen organizations can access does not define the mass market.

Kimi K2.6 may or may not close the remaining coding gap with Anthropic's top line. The benchmark answer is still pending. The market answer is already visible.

Developers are choosing the best model they can actually use, at a cost they can actually sustain, with deployment options they can actually control. On those terms, the cheap model is not a consolation prize. It is the default.

This is why the expensive model being locked in a room matters so much.

A closed frontier model can still win the research race. It can still win specific enterprise accounts. It can still be the best option for the handful of use cases where the last five or ten benchmark points justify the budget and the gating.

But markets are usually set by what is available, not by what is theoretically superior.

Right now, one side of this market is saying: here is the strongest thing we have, and almost nobody can touch it.

The other side is saying: here is something close enough, cheap enough, and open enough to build on immediately.

Software has seen this movie before.

The expensive model gets the headlines. The cheap model gets the users.

And in the long run, those are usually not the same winner.

Sources:

Previously on TheQuery: Anthropic Gave Its Dangerous Model to Defenders and Open Source AI Is No Longer a Side Project - the safety and open-model context this pricing split sits inside.