Daily AI News | OpenAI Cuts GPT-5.6 Luna and Terra Prices
# Daily AI News | OpenAI Cuts GPT-5.6 Luna and Terra Prices
## Lead Story
### OpenAI Cuts GPT-5.6 Luna and Terra Prices
OpenAI cut GPT-5.6 Luna prices by 80% and Terra prices by 20%. Separately, the company said backend kernel optimization reduced end-to-end serving costs by 20% and improved token-generation efficiency by more than 15%.
**AI take:** This does not increase a plan's total allowance; Luna or Terra simply uses less of that allowance for the same work, leaving more room for heavy tasks.
## More News
### Google DeepMind Releases the Gemini Robotics 2 Family
Google DeepMind released the Gemini Robotics 2 family: Gemini Robotics 2 for whole-body humanoid control, Gemini Robotics ER 2 for long-horizon planning and multi-robot coordination, and On-Device 2 for local adaptation. ER 2 is in AI Studio, while the two control models are initially available to selected partners.
**AI take:** The models move from single actions toward long, coordinated tasks, but access, success rates, and hardware-transfer costs will decide their practical value.
### Thinking Machines Releases the Full Weights of Inkling-Small
Thinking Machines Lab released the full weights of Inkling-Small. The model has 276 billion total parameters with 12 billion active at a time, supports a context window of up to one million tokens, and natively handles images and audio; fine-tuning and hosted access are also available.
**AI take:** Full weights enable private deployment and customization, but 276 billion total parameters still require server-class storage and serving resources.
### Gemini Spark Integrates with Chrome and Expands Availability
Google integrated Gemini Spark with Chrome, initially enabling browser actions in the United States, while expanding Spark for Google AI Pro users to more than 160 countries and territories. The agent can carry out tasks across web pages in the browser.
**AI take:** Browser access reduces manual page switching, but account permissions, sensitive data, and confirmation before consequential actions become central safeguards.
### Oracle Brings Gemini Models into Enterprise Applications
Oracle said Gemini 3.1 Flash-Lite and Gemini 3.5 Flash will be available through AI Agent Studio, Fusion Applications, and NetSuite for thousands of enterprise customers, supporting analysis, automation, and agent workflows.
**AI take:** Customers can avoid part of the custom integration work, but results still depend on data permissions, auditing, model choice, and usage costs.
### Amazon Reports Faster AI Growth alongside Heavy Investment
Amazon reported that second-quarter AWS revenue rose 37% year over year to $42.2 billion, while both its AI and chip businesses exceeded a $25 billion annualized revenue rate. Trailing-12-month free cash flow fell to negative $7.6 billion amid AI infrastructure investment.
**AI take:** Demand remains strong, but negative free cash flow shows how expensive the build-out is; the next test is whether added revenue earns back the capital.
### Zoox Wins a US Exemption for Paid Control-Free Robotaxis
The US National Highway Traffic Safety Administration granted a temporary exemption for Zoox's purpose-built robotaxi without a steering wheel or pedals, covering up to 5,000 vehicles over two years. Zoox may charge for rides after obtaining state and local approvals.
**AI take:** This clears a real commercialization hurdle, but a temporary federal exemption is neither nationwide operating permission nor proof of long-term safety.
### Judge Questions the Pentagon's Risk Designation of Anthropic
US District Judge Rita Lin said at a hearing that the government still lacked sufficient evidence for designating Anthropic a supply-chain risk and that the record had become less favorable to the government. The court has not issued a written ruling.
**AI take:** The hearing improves Anthropic's prospects, but the written decision and its scope will determine whether government procurement boundaries actually change.
### GitHub Opens Stacked Pull Requests to Public Preview
GitHub launched a public preview of stacked sessions and pull requests in the Copilot app. Large changes can be split into dependent layers that are reviewed independently and merged in order with one action, while the Copilot coding agent continues work across the stack.
**AI take:** Smaller changes are easier to review and roll back, but branch dependencies must stay clear or merge order creates new maintenance work.
### AI Helps Chrome Fix 1,072 Security Bugs in Two Releases
Google says AI-assisted security work helped Chrome fix 1,072 security bugs across two June releases. The company compared that count with 1,036 fixes across the previous 23 releases, spanning roughly two years.
**AI take:** The count shows the scale of automated discovery, but not bug severity; independent validation and regression quality still matter.
### Nscale Acquires Anyscale, the Company behind Ray
AI infrastructure company Nscale acquired Anyscale, the company behind the Ray distributed-computing framework. The deal combines Nscale's data-center and compute resources with Anyscale's software for training, inference, and cluster orchestration.
**AI take:** Vertical integration may reduce deployment friction, but value depends on keeping the Ray ecosystem open and integrating the products effectively.
### Okta Acquires AI Identity-Security Company Permiso
Okta acquired identity-threat-detection company Permiso, adding discovery, permission analysis, and anomaly detection for AI agents and other non-human identities. TechCrunch cited a source placing the price near $200 million, a figure Okta did not publicly confirm.
**AI take:** AI agents extend identity governance beyond user accounts to persistent non-human actors; visibility, detection coverage, and integration are the next tests.
## Rumors
### Seedance 2.5 Remains Marked Coming Soon
Seedance 2.5 drew fresh coverage while community posts also said it could open soon. As of 7:00 a.m. Beijing time on July 31, Dreamina's official page still said coming soon and did not show broad availability.
**AI take:** Dreamina lists 4K, longer video, and multi-reference input, but price, speed, and real output quality remain untestable before access opens.
### DeepSeek Reportedly Advances a One-Gigawatt AI Data Center
DeepSeek is reportedly developing an AI data center in Inner Mongolia that may add about one gigawatt of computing capacity for future model training and inference. Its investment, construction schedule, and operating date have not been disclosed.
**AI take:** One gigawatt is large data-center scale and signals competition moving into long-term compute reserves; funding, power supply, and delivery timing remain decisive.
### TSMC Reportedly Develops New AI Chip-Packaging Technology
TSMC is reportedly developing new AI chip-packaging technology in response to Intel's push in advanced packaging. Available reports do not specify the process, customers, capacity, or mass-production date.
**AI take:** AI chip supply can bottleneck at packaging even when wafers are available; the impact remains unclear until process and capacity details arrive.
