[AINews] The Field Guide to Fable
Latent Space ● Covered by 3 sources
Everyone's stress-testing Anthropic's Claude Fable 5 before its subsidy ends, while Tencent's new open-weight Hy3 model drops hard. Both show the same trend: open models are catching up fast, and the real fight now is who can serve them cheapest and fastest.
Based on reporting by Latent Space — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
There's a particular kind of chaos that happens right before a subscription discount expires, and this week it hit Fable 5. Thariq's "Field Guide to Fable" talk, rushed together the night before the relaunch, ended up being the most useful thing anyone published about the model: strip away the harnesses and prompts you built for older models, he argues, because they're quietly hobbling the new one. His four-part breakdown — unhobbling old habits, finding blind spots you didn't know you had, processing the weird grief of watching weeks of work compress into hours, and refusing to accept tradeoffs just because that's how it's always been — reads less like a product pitch and more like a survival manual for a step-change in capability.
Meanwhile the actual leaderboard fight got a lot more interesting. Tencent's Hy3, a 295B-parameter mixture-of-experts model with only 21B active parameters, 192 experts, and a 256K context window, shipped under Apache 2.0 with vLLM support ready on day zero — tool-calling parsers, speculative decoding, validated NVIDIA and AMD paths, the works. Tencent even upstreamed its own production kernels, claiming throughput gains near 3x on mixed-length decoding. That's not a research toy; that's a lab shipping infrastructure alongside weights, and it's forcing comparisons against GLM-5.2 that didn't feel plausible a few months ago.
The benchmarks paint a messier picture than any single number suggests. On Artificial Analysis's new AutomationBench-AA, which throws agents at 657 tasks across 40 simulated SaaS apps, Claude Fable 5 edged out Opus 4.8 by a hair — 48.6% versus 48.5% — while every single model, regardless of vendor, still broke business rules it was supposed to respect. Gemini 3.5 Flash trailed on raw score but won on cost efficiency and guardrail discipline. Open models, led by GLM-5.2 max at 27.8%, are still meaningfully behind on this particular test, even as they close the gap elsewhere. Nobody is winning cleanly anymore; they're winning on different axes, and cost-per-task is becoming as important as the leaderboard position itself.
Then there's Anthropic's J-space research, which is the kind of finding that makes interpretability people sit up and philosophers start arguing. The company claims to have found a small, privileged cluster of activations inside Claude that behaves like a global workspace — available for report, modulation, flexible reasoning, the sort of thing cognitive scientists have theorized about for decades in biological brains. Researchers like Neel Nanda called it the strongest evidence yet for something like working memory in an LLM. But Anthropic's own framing, which flirted with consciousness language, drew immediate pushback from people who think
My take — AI-written commentary, not fact-checked reporting
Watching Anthropic tiptoe toward consciousness language while shipping actual working products is the most Anthropic thing imaginable — brilliant interpretability work buried under a framing choice that guarantees a week of Twitter arguments instead of sober discussion. Meanwhile Hy3 quietly does the more important thing: ships an open, Apache-licensed model with production-grade serving on day one, which does more for the field than any workspace metaphor ever will.
Read more about this at: Latent Space
