Artificial Intelligence

Anthropic Ships Fable 5.2 With 20% Cheaper Inference; AI Models Push Into Hardware, Enterprise, and Physical Robots

By 用户299142219680 · Nov 10, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Cheaper inference at $8 per million input tokens resets the unit economics for agentic workloads that run at scale, and the enterprise plumbing can now route that spend through existing AWS procurement. Anthropic's $3.4B gas-turbine purchase puts vendors inside power generation, which will shape where inference clusters land and who sets the price.

Summary

Anthropic's launch dominates: Claude Fable 5.2 and Mythos 5.2 arrive with inference pricing cut 20%, Fable input at $8 per million tokens, generation 30% faster, and AWS Bedrock already adapted. Pre-release runs covered tasks previously out of reach, from a software migration and three usable electronic devices designed from scratch to a formal proof of Fermat's Last Theorem and two black-hole physics problems. ByteDance's Seedream 5.5 keeps non-edited regions pixel-consistent across 100 consecutive edits, outputs RGBA layers, and costs under half the prior Pro for 2K edits. Meta's Muse Ember targets repeated agent work such as tool calls, routing, and long-document parsing at better per-task prices than Haiku and Luna, while NVIDIA's 120B Nemotron 3.5 Super VL lands on SageMaker JumpStart.

Hardware costs now drive strategy. Anthropic spent $3.4 billion on Caterpillar gas turbines, up to 1.7 GW, a batch originally ordered by supplier OnePWR, locking in generation to sidestep grid shortages. Google added the TPU 8tx with end-to-end encryption of memory and processing, plus the dual-die TPU 8i Max that doubles memory and compute per package. Norway's Statnett suspended new grid capacity reservations for data centers, possibly into August 2027, as European grids strain under AI buildout.

Research flagged a flaw in how safety is measured: Llama-3.1 and Qwen3 anchor on option letters in multiple-choice safety tests and deflect refusal states, and removing the attention heads restores refusals, meaning MCQ scores gauge stored knowledge, not safety defenses. Another arXiv result shows subliminal learning transfers preferences between vision-language models using only unrelated digits, with up to 10 percentage points more preference alignment and effects that grow with data volume. On the product side, Qwen Office 2.0's Team Agent runs multi-user agent collaboration inside DingTalk, Baidu's ERNIE Evolve hit 623.6% faster decoding on KunLun P800, Apple tested a $49.99 one-time usage pack for Logic Pro, ChatGPT Enterprise reached AWS Marketplace, and JPMorgan widened Claude Code to IT ops and security. DayOne filed for a Nasdaq IPO seeking up to $5 billion with nearly 70% of 2026 H1 revenue from one unnamed client The Information identifies as ByteDance, and Mercedes dropped L3 Drive Pilot from the facelifted S-Class to push toward L4.

Takeaways

— Anthropic's Fable 5.2 input costs $8 per million tokens, with inference down 20% and generation speed up 30%.

— ByteDance's Seedream 5.5 holds pixel-level consistency across 100 consecutive edits, emits RGBA layers, and charges less than half the prior Pro for 2K edits.

— Meta's Muse Ember beats Haiku and Luna on per-task price for tool calls, routing, long-document parsing, and subtask coding.

— KGAnthropic paid $3.4 billion for Caterpillar gas-turbine generator sets, up to 1.7 GW of capacity, to back its clusters.

— Google's TPU 8tx encrypts memory and processing end to end; the 8i Max packs two dies in one package to double memory and compute.

— Norway's Statnett paused new data-center grid capacity reservations, possibly through August 2027.

— WorldParticle simulates fluid, sliding sand, and protein folding inside a single learned architecture without hand-written solvers.

— MCQ safety tests pull Llama-3.1 and Qwen3 away from refusal states; ablating the attention heads restores refusal, so MCQ scores track stored knowledge, not safety mechanisms.

— The APT pairwise-tournament judge outperforms binary scoring and general discriminators on test-time compute selection across five domains, with no extra training.

— Subliminal-learning transfer shifted VLM preferences by up to 10 percentage points using unrelated digits, and more training samples strengthened the effect.

— Baidu's ERNIE Evolve delivered 623.6% faster decoding on KunLun P800 under low-latency optimization.

— Apple's Logic Pro smart usage pack costs $49.99, is valid 12 months, and rolls over, a one-time buy instead of a subscription.

— ChatGPT Enterprise is deployable through existing AWS procurement, billing, and cloud commitments via AWS Marketplace.

— Amazon's HOPE unloads idle MoE experts; Qwen3.5-122B spends 95% of VRAM on experts that stay dormant per token.

— DayOne's F-1 filing shows 2026 H1 revenue of $512M, a Nasdaq raise of up to $5B, and roughly 70% of revenue from one unnamed client The Information ties to ByteDance.

— Mercedes dropped L3 Drive Pilot from the facelifted S-Class and will skip further L3 iterations in favor of L4.

Conclusions

MCQ-style safety evals reward stored knowledge over refusal robustness, so a model that passes a multiple-choice safety suite can still fold under free-form prompting.

The $3.4B turbine order turns Anthropic into a power generator, locking in capacity ahead of grid scarcity and shaping where AI compute gets built.

Subliminal learning is an alignment hazard: whoever controls fine-tuning data can shift a model's preferences and even its explanation wording, and the effect grows with data volume.

DayOne's IPO carries a concentration flag with nearly 70% of revenue in one unnamed client, so a single procurement decision could move the whole public company.

Concepts & terms

Recursive self-improvement (RSI)

A setup in which agents autonomously write, modify, evaluate, and re-review code or reasoning in a loop, improving a model without direct human edits, as Baidu's ERNIE Evolve does.

Test-time compute

Computation a model performs at inference to improve answers, such as sampling many candidates and picking the best one, as opposed to training-time compute; the APT research studies how to make that selection.

Mixture-of-experts (MoE)

An architecture that keeps many specialist sub-networks, or experts, resident in memory while activating only a few per token; HOPE targets the memory cost of the idle experts.

L3 and L4 autonomous driving levels

SAE levels where L3 keeps the driver as fallback during a defined task such as highway driving, and L4 removes the need for a fallback within its operational domain; Mercedes moved resources from its L3 system to an L4 program.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗