Mistral AI released Leanstral 1.5 as a fully open-source model designed to verify mathematical proofs and check whether software behaves correctly. The model targets formal verification workflows, giving developers an open alternative for proof-checking and code-correctness tasks.
OpenAI's GPT-5.6 Sol tied for first place in Code Arena: Frontend and placed second on the Artificial Analysis Intelligence Index, just one point behind Anthropic's Claude Fable 5. The model completed tasks 61% faster and at roughly half the cost compared to its closest rival, while the GPT-5.6 family introduced three tiers: Sol (flagship), Terra (lower-cost), and Luna (fastest, lowest-cost).
Databricks benchmarked coding agents on its multi-million-line production codebase, finding that open-weight models like China's GLM-5.2 (priced at $1.40/1M tokens) can match expensive frontier models on everyday coding tasks at roughly two-thirds the cost. The result signals that enterprise AI buying is shifting from public leaderboards to private tests on real workloads.
Meta removed its Muse Image AI model from Instagram just days after debuting it as a "creative partner" feature, following immediate public privacy backlash. The reversal highlights growing user resistance to AI features baked into social platforms.
DevRev open-sourced Enterprise-Bench, a vendor-neutral benchmark for evaluating whether AI agents can operate across fragmented data, siloed systems, and permission boundaries. Developed with the Laude Institute and validated by UC Berkeley professor Alexandros Dimakis, the framework includes shared datasets, methods, and public results so anyone can reproduce and compare agent performance.
Both DeepSeek and Zhipu AI (maker of GLM-5.2) were revealed to be developing their own in-house inference chips, part of an all-out Chinese push to break Nvidia and Huawei dependence and own the full AI stack from model architecture to silicon. The move extends DeepSeek's efficiency-first philosophy from software into physical infrastructure.
Meta will begin manufacturing its in-house AI chip, code-named Iris and designed with Broadcom, in September, with TSMC building it. The company plans 7 gigawatts of computing capacity in 2026, growing to 14 GW by 2027, and projects up to $145 billion in AI infrastructure spending as it tries to reduce reliance on Nvidia and AMD.