Anthropic's export ban is speedrunning an Asian Mythos clone market

What happened

When Anthropic announced in December 2025 that it would restrict Mythos access for customers in China, Singapore, South Korea, and a handful of other APAC markets — citing US Commerce Department guidance on advanced AI exports — the company framed it as a temporary compliance posture. Six months in, "temporary" looks structural. And the market has responded the way markets always respond to artificial scarcity: with substitutes.

TechCrunch's report yesterday cataloged four Mythos-class models shipped by Asian labs in Q2 2026 alone. Sakana AI's Kotodama-7 (Japan), Alibaba's Qwen3-Coder-Max, Naver's HyperCLOVA-X Reasoner (South Korea), and Zhipu AI's GLM-5-Code all target the same workload Mythos was designed for: long-context coding agents with tool use, structured output, and 200K+ token windows. Three of the four shipped with native MCP support — the protocol Anthropic itself published — and all four have OpenAI-compatible endpoints out of the box.

The benchmark gap is smaller than the export-ban thesis assumed it would be. On SWE-Bench Verified, GLM-5-Code lands at 58.2% versus Mythos-3's 64.1%. Qwen3-Coder-Max hits 56.8%. HumanEval Plus shows similar compression: 4-8 points behind Mythos-3, ahead of GPT-5-mini, and well ahead of any open-weights model from six months ago. Two of the four — Qwen3 and GLM-5 — are priced at roughly 35% of Mythos-3's per-million-token rate.

Why it matters

The export-control playbook assumes a durable capability moat: if you cut off access to the frontier, the gap widens because the cut-off party can't train against the latest. That's the chip-export logic, and for hardware it has mostly held. For models, the same logic is failing in real time, and the reason is that the inputs to a frontier model — papers, code, evals, distillation traces — leak much faster than fab equipment.

The Sakana paper accompanying Kotodama-7 is unusually candid about this. Their training mix included synthetic traces generated against Mythos-2 (which remained accessible through Q3 2025), GPT-5, and Gemini 2.5 Pro, plus a curated SWE-Bench-style task set built from public GitHub PRs. The architecture is a mixture-of-experts with 7B active parameters and 56B total — published, reproducible, no exotic kernels. If you can rent H200s and read arXiv, you can now train a 90%-of-Mythos model in a quarter. That's the actual story.

The community reaction on HN is split along predictable lines. The top comment (412 points) is a Japanese ML engineer pointing out that Kotodama-7's tool-use reliability beats Mythos-3 on Japanese-language function-calling tasks — a workload Anthropic has never optimized for. A counter-thread argues the gap reopens on agentic benchmarks longer than 50 steps, where Mythos-3's RL training still shows. Both are right. The honest read: frontier capability is now a spectrum with a long, flat shoulder, not a cliff.

There's also a coordination story. The four labs are not coordinating, but they're converging — same architecture family (sparse MoE, 50-80B total params), same context length targets (200K-1M), same protocol surface (MCP + OpenAI-compatible). This is what happens when an industry has a clear reference design and the reference vendor stops shipping to half the market. The substitutes don't innovate on shape; they innovate on locale, price, and latency. Naver's HyperCLOVA-X serves Korean enterprises from Seoul data centers with sub-30ms time-to-first-token. Anthropic, routing through us-east-1, can't match that even when it's allowed to serve the customer.

What this means for your stack

If you're a senior engineer at an APAC company that's been waiting out the ban, the calculus has flipped. The migration cost from Mythos to GLM-5 or Qwen3 is now mostly prompt-tuning and eval-rebuilding, not architectural. MCP servers port directly. Tool schemas port directly. Your structured-output Pydantic models port directly. The work is in the evals: your golden set was implicitly tuned to Mythos's quirks, and the new models fail differently. Budget two engineer-weeks for eval rework, not two engineer-months for re-architecture.

If you're at a US or EU company with no ban exposure, the secondary effect is pricing pressure. Anthropic has not cut Mythos-3 prices in eight months despite a 35%-cheaper competitor shipping in their primary cut-off markets — but that gap is now public, and procurement teams have started citing it. Two readers of this site reported that their renewal negotiations in May included "Qwen3 is 35% of your price for our Singapore workload" as a concrete lever. It worked for one of them (12% discount), not the other.

The more interesting structural question: should you build provider-portable from day one? The answer used to be "no, the abstraction tax isn't worth it." That was when there was one frontier vendor that mattered for coding agents. With four credible Mythos-class alternatives, the LiteLLM / OpenRouter / Vercel AI SDK abstraction layer is no longer a hedge — it's table stakes. If your agent code still has `import anthropic` at the top, you have an unhedged dependency on a vendor that has demonstrated, twice now, that it will turn off service to entire countries on 30 days notice.

Looking ahead

The export ban was supposed to slow Asian AI capability. It accelerated it, by creating a guaranteed market with no incumbent and a published reference design. Expect Q3 to bring at least two more entrants — Tencent's Hunyuan team has been quiet for too long, and ByteDance's Doubao roadmap leaked last week showed a coding-specialized variant landing in August. The interesting fight isn't capability anymore; it's distribution, locale, and price. Anthropic still wins on the long-horizon agentic workloads that justify its premium, but the addressable market for that premium just got smaller, and the floor under it just got higher. The ban will probably be relaxed in 2027. By then it won't matter.