Mistral AI released Leanstral 1.5 as an open-source model under the Apache 2.0 licence, designed to verify mathematical proofs and check software correctness. The release expands access to advanced formal verification tools that were previously limited to specialized academic use.
OpenAI launched GPT-5.6 for general availability with three models: Sol (flagship), Terra (balanced), and Luna (fastest, lowest-cost). Pricing ranges from $1 to $5 per million tokens with a 90% cache-read discount. Sol scored 88.8% on TerminalBench 2.1, beating Claude Opus 4.8's 78.9%. The release was delayed by US government cybersecurity restrictions.
Databricks adopted Zhipu AI's open-weight GLM-5.2 as its default coding engine after the model matched Claude Opus 4.8 quality at roughly half the cost. In Databricks' Pi harness tests, GLM-5.2 was 2.08x cheaper than Opus 4.8 at comparable pass rates (85% vs 87%). GLM-5.2 API runs at $1.40/M input and $4.40/M output, about 82% cheaper than Opus 4.8 ($5/$25).
SpaceXAI released Grok 4.5, a Cursor-trained model for coding, agentic tasks, and knowledge work priced at $2 per million input tokens. Artificial Analysis ranks it close to frontier performance, gaining 16 points over Grok 4.3 and beating all open-weights models. A single task costs just $0.31 on the Intelligence Index, five times cheaper than Claude Sonnet 5 (max).
Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning model for coding, computer use, and complex tool-based tasks. Alongside the model, Meta opened a public preview of the Meta Model API, its first paid AI model API. ZDNet reports Muse Spark 1.1 scored above both Anthropic's Opus 4.8 and OpenAI's GPT-5.5 on human-evaluated job tasks.
Amazon CTO Werner Vogels said companies are increasingly shifting toward cheaper open-source AI models to rein in costs, moving away from premium proprietary options. The trend aligns with the broader industry shift from bigger models to cheaper, smarter systems selected by task and cost rather than leaderboard rank.
Reuters reports both DeepSeek and Zhipu AI (maker of GLM-5.2) are developing their own inference chips, accelerating a shift away from Nvidia and Huawei dependence. The move signals that AI competition has escalated from model capability to a full-stack war encompassing models, chips, and cloud infrastructure. Meta also announced its "Iris" AI chip enters production in September, with 7 gigawatts of compute capacity planned for 2026.