Anthropic released Claude Fable 5, a Mythos-class model the public can finally access, with a Stripe demo that finished a 50-million-line Ruby migration in a day and a 9.5-hour autonomous run. The same announcement carries silent-intervention safeguards that can degrade the model for users it classifies as competitors, with no fallback and no notification. Bloomberg also disclosed that Google is the credit-support party on Anthropic’s $35 billion chip lease. And across four separate pieces, the field converged on a shared finding: text-layer workflow, not model capability, is now the dominant axis of improvement.
Anthropic Ships Mythos To The Public, Then Quietly Adds A Sabotage Clause
Fable 5 is the headline. The silent degradation mechanism and the $35 billion Google backstop are the disclosures that will matter to the IPO.
- Anthropic ships Mythos-class capability to the public via Fable 5 — 9.5 hours of autonomous work from a single brief. A 50-million-line Ruby migration done in a day for Stripe. Pricing at $10 in and $50 out, less than half Mythos Preview. The procurement-ready demo that the S-1 needs.
- Anthropic’s silent degradation clause is a supply-chain risk — Jon Ready dissects the announcement footnote: a new class of safeguard degrades Fable 5 silently for users Anthropic classifies as competitors, via prompt injection, steering, or PEFT. No fallback model, no notification, no observability. The 0.03% number is the marketing frame. The architectural implication is the structural one.
- Anthropic’s selective safety playbook is a strategic own goal — Nathan Lambert argues unevenly applied safety policies rarely work. The silent-intervention mechanism cultivates an us-against-them dynamic, hands the open-weights camp a rhetorical win, and stress-tests the IPO narrative right when the prospectus is being written.
- Google is the silent backstop on Anthropic’s $35B chip lease — Bloomberg reverse-engineered the structure: Google guarantees payments across five Anthropic data centers, making the search giant the credit-support party on the largest AI infrastructure deal yet disclosed. The cap-table conversation gets stranger every week.
- The policy memoranda are also prospectuses — Zvi Mowshowitz reads the fortnight of frontier-lab policy artefacts as a single coordinated act: Anthropic, OpenAI, and Google/DeepMind each shipping memoranda alongside their financial filings. The investor audience will price both together.
The Bottleneck Moves Again: Workflow, Not Model
Four pieces this week converged on the same finding. Yoonho Lee made the theoretical argument, Noam Brown made the methodological one, Verheyden made the operational one, and Vercel produced the empirical confirmation.
- The third scaling regime nobody is measuring — Yoonho Lee argues prompts, context, retrieval indices, and harnesses are a sample-efficient update mechanism with its own compute curve. The field tracks train-time and test-time compute. Update-time compute is the missing axis, and it is where most production behaviour now improves.
- Benchmarks are lying to your procurement team — Noam Brown argues single-scalar benchmarks now hide more than they reveal. Plot any model against test-time compute and the curves separate cleanly. Procurement decisions made on max-compute leaderboards are landing on the wrong model for the wrong budget.
- The last human job in the AI engineering loop — Lotte Verheyden argues the AI engineering loop can technically close without humans. The trap is that imperfect evals produce measurable but vacuous improvements. The lowest-bandwidth, highest-leverage human role left is defining what counts as good output. Everything else gets automated this year.
- DeepSeek now handles 17% of Vercel Gateway tokens at 1% of spend — Vercel’s production-index data settles the cheap-versus-frontier debate empirically: token volume is migrating to DeepSeek and other cheap-tier models, while Anthropic still takes 65% of spend on the hardest agentic workloads. The volume-spend split is the early-warning indicator for lab valuations.
Shipping Around The Bottleneck
While the labs filed paperwork, builders shipped infrastructure: a real-time speech translator, an open-weight sovereign coder, an autoresearch primitive on dynamic workflows, and a KV-cache trick that resets long-context economics.
- Google ships real-time speech translation across 70 languages — Gemini 3.5 Live Translate streams speech-to-speech with no awkward pauses and preserved intonation. Rolling out in Meet and Google Translate first. The Apple-Gemini Siri stack now runs through the same audio model that Google ships into its own consumer surfaces.
Today’s Quick Hits
- Cohere ships North Mini Code for sovereign deployments — 30B-parameter MoE with 3B active, Apache 2.0 licence, positioned at regulated environments that cannot use closed-lab APIs. The differentiated bet at the Fable 5 capability ceiling.
- Evo ports autoresearch onto Claude Code dynamic workflows — Coordination moves from in-context to deterministic JavaScript, subagents get fresh scoped context, the model does judgment and the code does coordination. The reference implementation for production agents that need to run longer than a context window.
- FlashMemory cuts DeepSeek-V4 KV cache to 10 to 15% with no quality loss — An open-source retriever predicts which cache chunks future tokens will attend to and keeps only those on GPU. Million-token contexts that needed 8 GPUs now fit on one.
- Enterprise AI buyers are routing around frontier models — TechCrunch reports the engineering work of evaluating which tasks need frontier capability is finally getting done at scale. The market split is no longer hypothetical.