Hari ini Anthropic rilis Claude Fable 5. Mythos-class model pertama yang available buat publik.

Dua bulan lalu model sekelas ini cuma bisa diakses segelintir partner lewat Project Glasswing. Too dangerous, katanya. Cybersecurity capabilities-nya bisa dipake buat nemu zero-day, bikin hacking tools, hal-hal yang bikin pemerintah dan enterprise parno.

Sekarang? Anyone with a Claude subscription can use it.

Caranya adalah mereka bikin versi yang "safe." Underlying model-nya identical sama Mythos 5. Bedanya, ada safety classifier yang ngeblok response di area high-risk. Cybersecurity. Biology. Chemistry. Distillation attempts.

Someone asks how to make ricin?
Redirect ke weaker model.
Someone tries to extract capabilities?
Redirect juga.

Otak Mythos, pagar safety.

Tapi ini bukan cerita soal safety. Ini cerita soal threshold.

Kita udah denger "AI is getting smarter" probably 50 kali dalam 3 tahun terakhir. But this time it's different. Karena yang dua bulan lalu masih dianggap "too dangerous for public release" tiba-tiba jadi produk berlangganan.

Pertanyaannya bukan lagi "what can this AI do?"
Tapi... what becomes possible now?

Jenny Wen, Design Lead Claude di Anthropic, recently ngomong di Lenny's Podcast. Di Anthropic, one engineer + one designer bisa shipping fitur yang dulu butuh 5-10 orang. Scheduled tasks diluncurkan cuma oleh dia dan satu engineer. One designer. One engineer.

That's not a 20% efficiency gain. That's the collapse of traditional team structure.

Dulu pembagiannya ada researcher, PM, designer, frontend, backend, DevOps. Sekarang? One person + Claude handles most of that. Not saying specialists are obsolete. Tapi threshold buat "bisa build something functional" drops dramatically.

Dan Fable 5 is 10%+ more capable than Opus 4.8 on engineering benchmarks. This isn't prediction. It's already happening.

Ada kategori ide yang biasanya gak dikerjain bukan karena nobody thought of it. Tapi karena cost/effort-nya gak masuk akal. Research yang butuh 3 orang 2 minggu. Analisis 500 halaman dokumen legal. Software prototype yang nunggu sprint berikutnya.

With a model this capable, the calculus changes. 3 orang 2 minggu bisa jadi 1 orang 2 hari. Dokumen 500 halaman bisa jadi initial pass sebelum lawyer review. Prototype bisa jadi dalam satu sore breee.

Not always perfect. But the threshold for "worth trying" drops so much. And the more you try, the more you find.

Yang paling gampang kelewat adalah Anthropic didn't say "we built safe AI so we can't release it." They said "we built safe AI SO we can release it." The safety classifier isn't a brake. It's a distribution enabler.

Kayak mobil. Nobody says "rem bikin mobil lambat." Rem justru yang bikin mobil bisa dipake sehari-hari. Same logic.

Ini template buat future releases. Build the most capable model, put guardrails on it, ship to public. Keep the unrestricted version for trusted partners. Two parallel tracks. Buat developer, you can build on this API for production workloads without worrying the model gets misused through your account.

Anthropic sekarang valuasi $965 miliar. Revenue run rate $47 miliar, naik dari sekitar $10 miliar tahun lalu. IPO confidential filing udah jalan. OpenAI juga baru filed, June 8. SpaceX/xAI merger IPO besok.

Three AI IPOs in one month.

Ini bukan 2023 hype cycle. Pasar modal, yang notoriously skeptical terhadap tech valuation, is starting to see AI as a revenue engine. Not just a research project burning cash.

Why does this matter? Because once AI companies go public, their incentives shift. They need to grow. Need to chase revenue. Need to build products people pay for. Dan ekosistem developer, tools, dan startup yang dibangun di atas model-model ini ikut terdorong naik.

Yang masih belum jelas.

Safety classifier-nya seberapa akurat? WIRED noted some benign queries get redirected to the weaker model. Anthropic says they'll iterate. But for now, defaultnya "err on caution." For production workloads, if the false positive rate is high, that's a dealbreaker.

Harga. $10 per million input tokens, $50 per million output. Twice the price of Opus 4.8. For startups and indie developers, that's significant. Anthropic says "higher ROI because better accuracy." Tapi itu vendor claim. Real-world usage yang bakal ngebuktiin.

Dan yang paling mengganjal adalah we don't really know what models like this will be used for. Not in the "danger" sense. But use cases.

Internet took years before people found genuinely useful things beyond email and web browsing. Smartphone took the App Store before becoming the ecosystem we know. AI kelas Mythos mungkin sama. The capability is here. The transformative application? Not yet.

Atau mungkin udah mulai muncul. We just haven't noticed.

Lucunya lagi... ini semua terjadi pas kompetisi udah bukan lagi soal research paper, tapi revenue dan market share. And suddenly the most capable model is the one opened to the public, not locked in a lab.

Karena di dunia dengan 3-4 big players yang roughly setara, the competitive advantage isn't "smartest model" anymore. It's distribution. Ecosystem. Who developers build on. And you can't build an ecosystem if only 10 companies can access your model.

So Fable 5 isn't just about "more capable AI." It's a strategy shift.. from "keep the strongest one locked up" to "distribute the strongest one, safely."

Buat kita yang bukan AI researcher atau VC, artinya... The most advanced tool ever built is now accessible to anyone willing to pay $20 a month.

The question isn't how smart it is. The question is what we do with it.

We probably don't know half the answers yet.
Probably just getting started. Hoohoohoo