The token got cheaper again | Tuesday, July 28, 2026

Microsoft released MAI-Cyber-1-Flash on July 27, its first model built for security work: a sparse mixture-of-experts design with 137 billion total parameters and 5 billion active per token. Inside Microsoft's MDASH vulnerability harness, paired with OpenAI's GPT-5.4, it pushes the CyberGym benchmark to about 96%, roughly 12 points above Anthropic's Mythos, at half the cost of the prior configuration. The model shoulders about 90% of routine security tasks and leaves GPT-5.4 for the hardest cases. The enterprise read: two things at once. Microsoft is building its own models to cut its OpenAI bill on routine work, and even so, its best score still needs a frontier model for the last mile. Cheaper for the bulk, dependent at the top. It is in Azure AI Foundry private preview for approved MDASH customers.

AMD used its Advancing AI event on July 22 and 23 to ship Helios, its first fully integrated rack-scale system: 72 Instinct MI455X GPUs in a single liquid-cooled rack, with 432 gigabytes of HBM4 per chip. CEO Lisa Su claims about 15% more compute than Nvidia's Vera Rubin, 50% more memory, and 30% more tokens per dollar, on AMD's own benchmarks. A configured rack lists near $5 million and ships through HPE, Lenovo, and Supermicro, and AMD paired it with Cerebras for low-latency inference. The enterprise read: the advice to stay portable across two providers now has a real second name attached. The catch is that the lock-in was never the chip, it is the decade of CUDA software on top of it, and the 30% figure is AMD's own. A second source is leverage only if your stack can actually run on it.

Block released Buzz on July 21, an open-source workspace where humans and AI agents work side by side, each with its own cryptographic identity tied back to a human owner. It folds team chat, code hosting, and automation into one place, and is built on Nostr, an open messaging protocol, rather than centralized accounts. Block already runs Buzz internally in place of Slack and GitHub, and its own coding agent handles roughly 15% of the company's production code changes at more than 200,000 operations a day. The enterprise read: as agents start doing real work, who did what becomes an audit question, and Block's answer is to give every agent a signed, attributable identity. That is the unglamorous plumbing most agent pilots skip and then regret.

Gartner projects worldwide spending on AI models and platforms will reach $64 billion in 2026, up 63% from $39 billion in 2025. The split is the story: generative-model spend is forecast to grow 117% and domain-specific, specialized models 210%, while broader platform spend rises a comparatively modest 37%. The enterprise read: budgets are still expanding, but scrutiny is expanding faster, and the money is moving toward smaller, task-specific models and toward vendors that build in cost transparency, evaluation, and usage tracking. For a CIO, that is permission to stop over-buying frontier capacity for problems a cheaper specialized model solves, and a reminder to make usage metering a procurement requirement, not an afterthought.

OpenAI expanded ChatGPT Health to all US adults on July 27, a dedicated experience that connects to Apple Health and supported medical records to answer questions in context. OpenAI says people already send it around 300 million health-related queries a week, up from 230 million in January. The expansion arrived a day after a Florida plaintiff sued to block it over a near-fatal medical suggestion. The enterprise read: this is a consumer launch, but it drags a question onto every enterprise roadmap that touches health-adjacent data. Where does the model's output sit in your liability chain, and what is your evidence trail when it gives advice? If a 300-million-query consumer product is already facing that suit, a regulated enterprise deploying the same class of model needs the logging and human-review layer in place first.