Anthropic extended plan-inclusive access to Claude Fable 5 through July 12, postponing the original July 7 cutover, after which the model shifts fully to usage-credit billing at $10 per million input tokens and $50 per million output tokens (fable5.app). That makes Fable 5 the most expensive entry on Anthropic’s current price list - double Opus 4.8’s rate of $5/$25 per million (TechTimes). Through July 12, Pro, Max, Team, and select Enterprise subscribers retain access capped at 50% of their weekly plan limit; after that date every Fable 5 request draws from usage credits (fable5.app). A session with roughly 50,000 input tokens and 15,000 output tokens costs approximately $1.25 at those rates, placing sustained daily heavy use well above $1,000 per month (Digital Applied). Prompt-caching discounts cut repeated-context reads to $1 per million - a 90% reduction from the headline rate - partially offsetting costs for workloads with large shared context (fable5.app).
A CNBC report published July 7 documented accelerating adoption of Chinese open-weight models by U.S. enterprises seeking to cut inference costs (CNBC). Z.ai’s GLM-5.2 and Xiaomi’s MiMo series are priced 60% to 90% below comparable proprietary API offerings from OpenAI and Anthropic, with GLM-5.2’s first-party API at $4.40 per million output tokens against Fable 5’s $50 (VentureBeat). Vercel data cited by CNBC showed GLM-5.2 daily token volume grew roughly 27x and its customer count approximately 80x in the model’s first full week after its June 16 release, the fastest adoption Vercel has tracked in 2026 (CNBC). Xiaomi’s MiMo series separately crossed 4 trillion weekly tokens on OpenRouter, ranking as the platform’s highest-volume model family by token count (OpenRouter).
Ollama released a release candidate on July 7 addressing CUDA and GPU compatibility, server cleanup, and UTF-8 file handling (GitHub). The build refreshes the bundled llama.cpp and mlx components, removes support for deprecated ROCm devices, and includes CUDA fixes covering toolkit lookup, parallel processing, JIT packaging, and a fallback to standard CUDA when the JetPack runner is absent (Releasebot). The update also improves agent harness core behavior and includes stability fixes across Windows, Vulkan, and generation handling (releases.sh).