wednesday, july 1, 2026
top 4 trending●previous 24h●generated 35d ago
Anthropic has restored access to Claude Fable 5 after export restrictions were lifted, making the model available again following a period of unavailability. The company implemented new safety measures that flag and route harmless queries to Claude Opus to address previous security concerns.
- Fable 5 access restored as of July 1, 2026 after export control restrictions were lifted
- New safety system flags harmless queries and routes them to Claude Opus model
- Early access users reported security guardrails were frequently triggered for non-obvious reasons
- Mythos 5 access also being restored alongside Fable 5
- Unavailability had delayed some research and development work for early access users
ZCode, a code execution harness built by the GLM team, has launched for GLM-5.2, enabling the model to generate and run code directly. The tool allows GLM-5.2 to execute complex tasks like building hypervisors within its own environment.
- ZCode is an official harness developed by GLM's creators specifically for GLM-5.2
- Enables GLM-5.2 to generate, execute, and test code autonomously
- Launched July 1, 2026
- Supports complex system-level tasks such as hypervisor development
- Positions GLM-5.2 as a code-execution-capable model comparable to Claude's code tools
Meta is selling excess GPU compute capacity to external customers, similar to SpaceX's approach with spare launch capacity, while simultaneously capping internal AI token spending to manage costs. The move reflects Meta's strategy to monetize underutilized infrastructure and control its own AI compute expenses.
- Meta offering spare GPU capacity for sale to external customers
- Business model mirrors SpaceX's excess launch capacity sales approach
- Company implementing internal caps on AI token spending
- Strategy aims to offset infrastructure costs and improve compute utilization
- Reflects broader industry trend of AI infrastructure providers monetizing excess capacity
Claude Sonnet 5 introduces a new tokenizer that increases pricing for English and Spanish inputs but keeps Simplified Mandarin costs roughly flat. The tokenizer makes English inputs approximately 1.4x more expensive and Spanish inputs 1.33x more expensive compared to the previous version.
- New tokenizer increases English token costs by ~1.4x
- Spanish token costs rise ~1.33x under new tokenizer
- Simplified Mandarin pricing remains approximately unchanged
- Claude Sonnet 5 is positioned as a capable but non-frontier model