Why won’t the AI distillation debate die? Because the economic motive is too large.

GPT-6 Astra’s output tokens cost 50 times what Kimi charges, and 42 times DeepSeek’s price. Anyone who can reproduce flagship behavior from its outputs can sell for far less and still profit.

My yardstick has always been simple: not who resembles the flagship most, but who finishes the task at the most reasonable total cost. And that is proven by testing, not by charts.

Kimi K3.1 Pro stands accused of distilling GPT-6 Astra, and GLM-5.3 of mimicking the Claude 5 family. The “evidence” is Artificial Analysis personality charts: both models’ speaking styles hug the Western flagships closely.

In my view, a style chart only proves two models talk alike, not who learned from whom. Proving distillation takes API logs and account records, not a radar chart.

An accusation of this weight should be judged on auditable evidence, not on resemblance that feels convincing.

One number from the Claude 5.5 leak made me stop scrolling: Mythos 5.5, a cybersecurity-only edition, is claimed to exploit the Linux kernel and gain root on its own in 53 minutes.

If true, this is no longer ordinary coding benchmarks. This is a model built for offensive work, with access limited to top-tier security institutions.

But precisely because the number is that dramatic, I file it as a leak claim. Even official release numbers need testing, let alone figures pulled from a snippet of source code.

Gemini 4 Carbon has held the number one spot for less than a week, and both of its rivals already “lost” their next flagships in the same weekend.

OpenAI reportedly skipped GPT-6.1 and is quietly testing GPT-6.2 Astra in the Codex backend, handling nearly 20% of traffic. Meanwhile, Claude Mythos 5.5 and Fable 5.5 were spotted in the source code of Anthropic’s own website.

But note: this is a leak, not an official launch. No model card, no release date. Claims and evidence still need to be kept apart.

Kenapa isu distilasi AI tidak akan mati? Karena motif ekonominya terlalu besar.

Token output GPT-6 Astra berharga 50 kali lipat dari Kimi, dan 42 kali lipat dari DeepSeek. Siapa pun yang bisa mereproduksi perilaku flagship lewat outputnya bisa jualan jauh lebih murah dan tetap untung.

Buatku, dari dulu ukurannya sederhana: bukan siapa yang paling mirip flagship, tapi siapa yang menyelesaikan tugas dengan biaya total paling masuk akal. Itu dibuktikan lewat pengujian, bukan grafik.

Kimi K3.1 Pro dituding hasil distilasi GPT-6 Astra, GLM-5.3 dituding meniru keluarga Claude 5. “Buktinya” adalah grafik kepribadian Artificial Analysis: gaya bicara keduanya menempel rapat ke flagship Barat.

Menurutku grafik gaya hanya membuktikan dua model berbicara mirip, bukan siapa belajar dari siapa. Membuktikan distilasi butuh log API dan catatan akun, bukan radar chart.

Tuduhan sekelas ini dinilai dari bukti yang bisa diaudit, bukan kemiripan yang terasa meyakinkan.

Satu angka dari bocoran Claude 5.5 yang bikin aku berhenti scroll: Mythos 5.5, edisi khusus keamanan siber, diklaim bisa mengeksploitasi kernel Linux dan dapat akses root mandiri dalam 53 menit.

Kalau benar, ini bukan lagi benchmark coding biasa. Ini model untuk pekerjaan ofensif, dan aksesnya cuma untuk institusi keamanan papan atas.

Tapi justru karena angkanya sedramatis itu, aku menandainya sebagai klaim bocoran. Angka rilis resmi saja harus diuji, apalagi yang dari potongan source code.

Belum seminggu Gemini 4 Carbon duduk di peringkat satu, dua rivalnya sudah “kecolongan” flagship berikutnya di akhir pekan yang sama.

OpenAI dilaporkan melewati GPT-6.1 dan langsung menguji GPT-6.2 Astra diam-diam di backend Codex, menangani hampir 20% traffic. Di sisi lain, Claude Mythos 5.5 dan Fable 5.5 ketahuan dari source code situs Anthropic sendiri.

Tapi catat: ini bocoran, bukan pengumuman resmi. Belum ada model card, belum ada tanggal rilis. Klaim dan bukti masih harus dipisahkan.