Anthropic's New Anti-Distillation Strategy: Restricting Thinking Block Edits, Fable 5.1 Gets 75% Price Cut

Anthropic's New Security Move: Restricting Thinking Block Edits to Combat Distillation Attacks

On September 2nd, Anthropic dominated the AI headlines with a flurry of moves spanning product upgrades and security strategy changes. First came new model releases — Fable 5.1 and Mythos 5.1 launched simultaneously. The two share the same underlying model weights, differing only in safety policies and access scope. This "same-source, tiered" approach essentially gives enterprises more flexibility to balance capability against compliance.

Both new models show meaningful improvements in coding, long-horizon agentic tasks, and research capabilities. For developers, the most immediate benefit is lower cost: Fable 5.1's cache read pricing has been reduced by 75%, with overall costs on typical tasks dropping by roughly 25%. As API call costs across the LLM space continue to compress, this reduction is significant enough to influence purchasing decisions for many small and mid-sized teams.

The more interesting development is on the security side. Anthropic updated the Messages API to restrict editing of "thinking blocks" in prior Claude context during multi-turn conversations. The stated rationale: making distillation attacks harder. A distillation attack involves an adversary using carefully crafted multi-turn interactions to gradually extract a model's internal reasoning process, effectively "replicating" a smaller model with similar capabilities. By locking thinking blocks in historical context, Anthropic is cutting off a critical pathway for this kind of reverse engineering.

The new rule currently applies only to new accounts using Fable 5.1, with a gradual rollout to all users planned for later. This signals a broader trend: as model capabilities increasingly constitute core assets, leading vendors are treating "distillation prevention" as a standard security engineering concern.

Claude's 270,000-Character System Prompt Allegedly Leaked

On a related note, Claude Fable 5.1's system prompt has reportedly been leaked — at over 270,000 characters, that's a staggering length. The leaked content shows its knowledge cutoff has been updated to late June 2026, confirms shared weights with Mythos 5.1, and reveals adjustments to rules around safety and tool calling.

A system prompt of this scale speaks to how complex modern LLM behavioral constraints have become — requiring enormous volumes of rules to maintain. It also exposes the fragility of "prompt engineering" as a security boundary. When core behavioral logic can be fully reconstructed, the necessity of hardening protections at the API layer (such as restricting thinking block edits) becomes self-evident.

Parallel Progress: On-Device Models and World Models

Beyond the major players, several teams made notable advances in specialized areas.

iFlytek Open-Sources Spark X 2.5 Edge Models

iFlytek open-sourced 4B and 1.7B variants of Spark X 2.5, explicitly targeting on-device agent scenarios. The models use a hybrid attention architecture, natively support context lengths up to 1 million tokens, and are optimized for code, math, and instruction-following. For developers looking to run agents locally on devices, the combination of small parameter count and extremely long context is quite attractive. iFlytek also launched an "Edge Agent Challenge" alongside the release, with registration open until September 6.

Fei-Fei Li's World Labs Releases World Model Atlas

On another front, World Labs — founded by Fei-Fei Li — released Atlas, a full-modality world model set to enter early access within weeks. Atlas supports pixel and camera control, single-image reconstruction of large scenes, native 3D output, and can generate sensor and simulation data for robotics. The value of such world models extends beyond content generation: they provide cheap, scalable simulation training environments for embodied intelligence — potentially an unavoidable piece of the puzzle for robotics at scale.

Apple vs. OpenAI Trade Secret Lawsuit Escalates

This week also brought a major litigation development. Apple has accused OpenAI of destroying evidence in a trade secret lawsuit and has filed for an expedited discovery process.

According to Apple's account, a MacBook belonging to former employee Chang Liu contained records of evidence destruction discussions. Apple also alleges that after leaving the company, Liu downloaded confidential circuit diagrams and used an AI agent to bring this material into OpenAI's workflow. The central question in the case: did OpenAI acquire Apple's technical secrets by poaching three former employees?

This lawsuit reflects the undercurrents running beneath the AI talent war. As top engineers move frequently between giants, the boundaries of technical secrets become extraordinarily blurry. The detail about "using an AI agent to bring materials into a workflow" is particularly noteworthy: it suggests that AI tools themselves may be emerging as new vectors for intellectual property transfer — posing entirely new challenges for corporate information security.

Hardware and Industry: Compute Democratization and Cost Pressures

NVIDIA Jetson Orin Nano Super Drops to $249

On the hardware front, NVIDIA introduced the Jetson Orin Nano Super developer kit at $249. The kit delivers up to 1.7x better generative AI inference performance and 67 TOPS of INT8 compute, suited for RAG, chatbots, vision agents, and robotics development — with a free upgrade path for existing users. Bringing capable edge AI compute under the $250 mark will undoubtedly lower the barrier to entry for local agent development.

AI Industry Investment Remains Intensely Hot

On the investment side, multiple data points reflect the industry's fever pitch. Zhipu AI reported H1 revenue of 954 million yuan, up roughly 400% year-over-year, with its MaaS API annualized revenue surging from $250 million to $2 billion within the half-year while turning profitable — and cloud deployment revenue now exceeding on-premises deployments. Meanwhile, Japan's Ministry of Economy, Trade and Industry has requested a record $49 billion budget to accelerate AI, semiconductor, and robotics development, reinforcing critical technology competitiveness and supply chain resilience.

TSMC Price Hikes Signal Upstream Cost Pressure

The cost side, however, is sending pressure signals. Driven by continued strong AI chip demand, TSMC is reportedly planning to raise wafer prices across all process nodes by 10–15%, with Samsung's 4nm and 5nm pricing expected to follow suit, and the market anticipating advanced process nodes may see further increases before the end of 2027. These upstream cost pressures could ultimately pass through to end products — beneath the surface narrative of ever-cheaper compute lies the tightening reality of manufacturing costs up the chain.