The top things worth knowing about in AI today.
OpenAI made its GPT-5.6 family generally available today after a limited preview and a US government safety review. The line has three tiers: Sol, its strongest model, at $5 per million input tokens; Terra at $2.50; and Luna at $1, with output priced at $30, $15 and $6 respectively. Terra matches last generation's GPT-5.5 at roughly half the cost, so most teams can cut their inference bill without losing capability.
Read moreIllinois Governor JB Pritzker signed the Artificial Intelligence Safety Measures Act, requiring developers earning over $500 million a year to publish how their models could cause catastrophic harm and to report critical safety incidents to the state within 72 hours. Annual third-party audits become mandatory when the law takes effect on 1 January 2027. With New York and California, roughly 40% of the US market will sit under comparable rules, making state law the effective national standard while federal action stalls.
Read moreThe UN held its first Global Dialogue on AI Governance in Geneva on 6-7 July, bringing member states together to work out how the technology should be managed across borders. The Secretary-General warned against leaving AI oversight to a handful of companies and countries. For businesses it signals that international rules, not just national ones, are starting to form, though concrete commitments remain some way off.
Read moreAnthropic extended its Cowork agent from the desktop to web and mobile, so a task can start on a laptop, keep running in the background, and be checked from a phone after the app is closed. It is in beta for Max subscribers first, with wider access over the coming weeks. Anthropic also noted most Cowork use is not coding, a sign these agents are moving into general office work rather than just software teams.
Read moreOpenAI, Microsoft, Anthropic and Amazon have now committed around $8 billion combined to units that embed engineers directly inside customer companies to get AI working in real workflows. OpenAI is putting in more than $4 billion, Microsoft $2.5 billion, Anthropic $1.5 billion and AWS $1 billion. The move follows survey data showing 71% of large-company executives blame organisational readiness, not the technology, for stalled AI projects, so the bottleneck is now integration rather than model quality.
Read moreAnthropic published interpretability research identifying a small internal region, which it calls J-space, that Claude appears to use to hold and work through ideas before putting them into words. The team found a split resembling the human divide between deliberate reasoning and automatic processing. Anthropic stopped short of claiming the model is conscious; the value is a clearer view of how these systems represent information, which matters for safety and debugging.
Read morexAI released a beta of Voice Agent Builder, which lets people create a working Grok voice agent from a plain-language description in about two minutes, with telephony, knowledge retrieval and guardrails bundled in. It runs at $0.05 per minute and supports 25+ languages with mid-call switching, plus voice cloning from two minutes of audio. It lowers the bar for building phone-based customer service and sales agents without stitching together separate speech and language services.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.