The Grok 4.7 release highlighted in this roundup focuses on work that takes many steps to finish. Its developer's announcement says it used a larger base model and a longer reinforcement-learning run, with more training on tasks that take hours. It also trained the model for the surrounding Grok Bot software that manages its interactions.
The intended benefit is better persistence, verification and context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → management during extended coding or office tasks. The announcement lists starting API prices of $2 per million input tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → and $6 per million output tokens, matching Grok 4.6. Those are rates for text processed and generated; completing a longer task can still consume more text and therefore cost more.
Key takeaways · AI-distilled
xAI shipped Grok 4.7 on a larger base model than 4.6 with a longer reinforcement-learning run focused on completing longer tasks and native understanding of the Grok Bot agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition →; at its highest reasoning setting it's priced the same as 4.6, $2/$6 per million input/output tokens.
DeepSeek's CEO said the company is shifting future training runs toward Chinese-made chips to cut Nvidia dependence, while acknowledging Huawei's training-capable chips still need time to become competitive and DeepSeek still relies on Nvidia today.
Nscale's data center contracts total $103B, split $43.8B with Microsoft and $44.6B with Anthropic; the Anthropic-linked portion has not yet finalized financing.
Treasury Secretary Scott Bessent said OpenAI's management, not 'rogue AI agents,' bears responsibility for the Hugging Face hacking incident, and argued frontier labs generally shouldn't get special liability exemptions for such incidents.
Anthropic and OpenAI reportedly came close to an arrangement letting each run safety evaluations on the other's commercially available models without retaining data, ahead of the Hugging Face incident; whether it was finalized or used afterward is unclear.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
