Podcast··41m

Episode 12 — Fable's Price Tag and China's Token Takeover

Anthropic's Fable model brings Mephos-grade intelligence at $50 per million output tokens, while Chinese labs led by DeepSeek dominate OpenRouter's token consumption charts.

Listen onYouTubeSpotify

Episode notes

Post-hackathon, the team digs into Spotify's personal DJ voice agent, Anthropic's Fable model (the public sibling of the locked-down Mephos), and the OpenRouter charts showing Chinese models eating the token market.

Chapters

  • 01:00 — Spotify's personal DJ: a voice agent that actually works
  • 13:00 — Hackathon recap: builders, keyboards, and putting Haarlem on the map
  • 20:00 — Fable and Mephos: Anthropic's security-grade model goes public
  • 30:00 — OpenRouter token wars: Chinese models eating the market

Episode 12 — Fable's Price Tag and China's Token Takeover

Description: Anthropic's Fable model brings Mephos-grade intelligence at $50 per million output tokens, while Chinese labs led by DeepSeek dominate OpenRouter's token consumption charts.

Two forces are pulling in opposite directions right now. Anthropic is shipping its most expensive, most capable model yet under a tight time-limited window. Meanwhile, Chinese labs are quietly taking over the token consumption charts by being almost unimaginably cheap. The tension between those two strategies is the story of this episode.

Spotify's Personal DJ Is the Voice Agent That Stuck

Spotify shipped a personal DJ that behaves like a usable Siri or Alexa for music. You speak to it, it controls playback, and it reads your years of listening history to surface exactly the right track.

  • Context-aware sequencing: it uses what you already listen to, then proactively suggests new music in the same direction, switching vibes mid-session ("let's go a little more Latin now").
  • Behavioral data done right: the team's takeaway is that when a company uses long-horizon behavioral data this well, giving up more of it feels worth it. Three days in, the feature had already changed daily routines.
  • The uncomfortable corollary: music is a strong proxy for mood. Once Spotify's emotional profile gets joined with a search profile or a shopping profile, the targeting implications are significant, and probably already happening.

Hackathon Recap: Haarlem on the Map

The first Digilize hackathon pulled builders from across the Netherlands and one attendee from Italy, with unlimited food, six hours of build slots, and two challenges.

  • Reviews were strong for a first event, with clear room to improve the format next time.
  • Community signal: people brought their own keyboards, stayed locked in for the full build window, and the energy validated the bet on putting Haarlem on the map as a place where things get built.
  • What's next: weekly episodes resume, with more community building planned. Recap and photos live on the hackathon site.

Fable and Mephos: Anthropic's Security-Grade Model Goes Public

Anthropic's Mephos is the cybersecurity and biology-capable model originally released under Project Glasswick to a select group of US companies. Fable is its public sibling, now available in Claude Code and the web UI, with much tighter guardrails.

  • Guardrails by routing: ask Fable about cybersecurity or biology and it routes the query through Opus 4.8 instead of answering with the Mephos architecture. It is a huge model, likely above one trillion parameters, with intelligence well beyond Opus on complex tasks.
  • Pricing: $10 per million input tokens and $50 per million output tokens. Compared to DeepSeek V4 Pro at roughly $0.43 input and $0.87 output per million, the gap is staggering.
  • Time-limited access: Fable is available in Pro and Max subscriptions only until June 22, after which Anthropic pulls it and pushes users toward the API platform. They have hinted at extension if they secure more compute.
  • Subsidized subscriptions: independent researchers found the Max plan ($200/month) effectively yields up to 8.8k of API usage, a deliberate loss-leader to pull users into the ecosystem.
  • Already jailbroken: Pliny cracked Fable using out-of-distribution tokens (tokens from languages outside Anthropic's training data) and multi-step distillation. Anthropic now routes any prompt containing "Pliny" away from Fable, a backhanded acknowledgement.

The practical playbook emerging in the community: use Fable for planning, auditing, and the big-picture reasoning, then hand the implementation off to a cheaper model like DeepSeek V4 Pro.

OpenRouter Token Wars: Chinese Models Are Eating the Market

A weekly OpenRouter chart tracking token consumption across the top nine models tells a clear geopolitical story.

  • Head-to-head, then a rupture: US models ran roughly neck-and-neck until a November drop (likely a mix of model-release drought and year-end budget cycles), then token consumption exploded upward, driven by Chinese models.
  • Top of the chart is Chinese: DeepSeek, Minimax M3, Tencent, Xiaomi, and Step 3.7 dominate the weekly top models. The team's reaction: "Chinese, Chinese, Chinese."
  • Cost is the wedge, but intelligence has to be there: cheap and stupid does not win. Chinese models win because they are cheap and genuinely good, especially Qwen 3.6 as a local default.
  • Open weights drive adoption: the local-to-cloud ladder matters. Run Qwen 3.6 locally, then scale up within the same model family. That continuity is a real adoption moat that closed American models cannot replicate.
  • New local coder: a community distilled Opus 4.8 into Qwen 3.6 and released "Quopus 35B Coder," the first specialized local coding model on the Qwen 3.5 architecture, with MTP for speed.

The open question is how durable the Chinese subsidy is. The EV parallel looms: aggressive government funding can inflate numbers until the funding pulls back. The New York Times reported the hedge fund backing DeepSeek is connected to the CCP, ostensibly to capture training data, mirroring how American labs operate with their own funding sources.

Conclusion

The model market is bifurcating in real time. Anthropic is betting that frontier intelligence at premium prices will hold, and Fable is the proof-of-concept for that bet. Chinese labs are betting that open weights and rock-bottom pricing will capture the volume. Both can be right simultaneously, and the builders who win will be the ones who know which model to use for which job.

Listen on Spotify