Daily AI News | GPT‑6 Astra rolls out with end-to-end RL and alignment gains, scoring 99.9% on ARC‑AGI‑3
# Daily AI News | GPT‑6 Astra rolls out with end-to-end RL and alignment gains, scoring 99.9% on ARC‑AGI‑3
## Lead Story
### GPT‑6 Astra rolls out with end-to-end RL and alignment gains, scoring 99.9% on ARC‑AGI‑3
OpenAI began a phased rollout of GPT‑6 Astra, with ChatGPT Plus, Pro, Business, Enterprise, API, and AWS access planned over the following days. OpenAI says the training program combined pre-training, reinforcement learning, and alignment improvements, spanning pre-training data composition through graders used in RL. In OpenAI research or API evaluations, Astra scored 99.9% on ARC‑AGI‑3, 97.6% on FrontierMath Tier 4 v2, 57.9% on Terminal‑Bench 4.0 versus 37.3% for GPT‑5.6 Sol, and 72.6% on OSWorld 2.0 versus 65.7% while taking about 47% less time per task; it also scored 100% on ExploitBench. An internal scope-violation test fell from Sol’s 48% to 0%, although OpenAI says Astra’s written reasoning became harder to monitor.
**AI take:** Vendor tests point to gains in agent execution, reasoning, and cyber capability, but research/API harness results do not equal every ChatGPT experience; weaker monitorability shows RL and alignment remain unfinished.
## More News
### Tibo promises one banked reset for each day a paid ChatGPT user lacks Astra access
Codex head Thibault “Tibo” Sottiaux said that, starting September 4, paid ChatGPT-plan users will receive one banked usage reset for every day they do not yet have Astra access. He said the first would land about three hours after his post. The commitment compensates for the phased Astra rollout; it is not a permanent daily quota increase for every user.
**AI take:** Banking resets can reduce inequity between early and late rollout cohorts, but users should distinguish this compensation from normal rolling limits and a permanent quota expansion.
### Gemini 3.8 Flash reaches GitHub Copilot for coding and long-horizon agent work
GitHub made Gemini 3.8 Flash selectable in Copilot across supported IDEs, Copilot CLI, and cloud-agent experiences, subject to plan access and organization model policy. Google positions the one-million-input-context model for coding, software engineering, and long-horizon agentic work. Its original model launch predates this issue’s window, so this story treats Copilot availability—not the model debut—as the new stage.
**AI take:** A model changes workflows when it reaches a mainstream tool; capability, Copilot policy controls, and observed latency should be evaluated together.
### OpenCode offers Meta Muse Spark 1.3 free for a limited time as the model targets efficient long-horizon coding
OpenCode’s official model catalog and public account confirm limited-time availability of Meta Muse Spark 1.3 Contributor Free. Meta says 1.3 was trained for longer-horizon agentic and coding work and used about 20% fewer tool calls and 25% fewer tokens than 1.2 in internal comparisons. The model debut predates this issue’s window, so the new stage here is OpenCode’s promotion, not free access across every Meta channel.
**AI take:** If the tool-call and token reductions hold in real repositories, they affect both speed and cost; testers should still check data policy, regional access, and post-promotion pricing.
### NVIDIA agrees to acquire Hugging Face for $12.9303 billion
NVIDIA agreed to acquire Hugging Face for $12.9303 billion. NVIDIA says Hugging Face will remain open, multi-cloud, multi-accelerator, and supportive of open-source and open-weight models from different builders; completion remains subject to the transaction process.
**AI take:** The deal ties a major model-distribution hub more closely to a compute platform, making implementation of the openness commitments important for developers.
### Google releases WeatherNext 3 across Search, Gemini, Maps, and Cloud
Google DeepMind and Google Research released WeatherNext 3. It learns from live satellite data, produces hourly forecasts at up to five-kilometer resolution, and improves precipitation and clean-energy variables; Google is integrating it into Search, Gemini, Maps, Maps Platform, and Cloud.
**AI take:** Moving a weather model from research into mass-market products makes regional coverage, extreme-weather reliability, and continuing independent evaluation the next tests.
### OpenAI commits $1 billion to expand Daybreak access for frontline defenders
OpenAI launched Daybreak for Frontline Defenders with a $1 billion commitment for subsidized access to frontier cyber models and products, training, technical support, and partnerships. Initial priorities include water, power, local government, community banking, nonprofits, and open-source maintainers.
**AI take:** For high-capability cyber models to create net benefit, access and support for resource-constrained defenders matter alongside restrictions on offensive use.
### NVIDIA releases open-source PAIR and expands local agents with RTX Spark
NVIDIA announced a local-AI bundle at IFA: the free open-source PAIR routes inference across PCs on a local network; llama.cpp and vLLM receive new acceleration; Hermes and OpenClaw are simplifying local setup; and RTX Spark Windows PCs are planned for October.
**AI take:** Connecting idle PCs, inference backends, and easier agent setup moves local AI closer to everyday use than hardware capacity alone.
### Tesla puts Cybercab into paid robotaxi service in Austin
Tesla placed its steering-wheel- and pedal-free Cybercab into paid robotaxi service in Austin, marking a transition from demonstration vehicle to commercial operation. Tesla also published a rider guide. The service remains geographically and operationally limited, and public reporting described the launch as unusually quiet.
**AI take:** The real test now shifts from launch messaging to operating data: safety interventions, service area, fleet utilization, and the regulatory record.
### Wayve and Uber launch the UK’s first public autonomous ride-hailing service
Wayve and Uber launched supervised autonomous rides in London. Ordinary UberX, Uber Electric, or Uber Comfort requests may match a Wayve vehicle; every initial ride still has a licensed safety driver, airports are excluded, and the fleet will grow gradually.
**AI take:** This moves the technology from closed testing into real passenger matching, while supervised operation remains distinct from driverless service.
### Flex buys EPC Power for $4.4 billion to add AI data-center power conversion
Flex signed a definitive $4.4 billion agreement to acquire EPC Power, adding grid-forming, rectifier, DC-DC, and planned 800V data-center capabilities to its cloud and power infrastructure business. Closing is expected in the fourth quarter of 2026, subject to customary conditions.
**AI take:** High-density GPU racks are extending competition from chips into power architecture, making conversion technology a strategic AI-infrastructure asset.
### Zoox expands its paid Las Vegas robotaxi service to the airport
Amazon-owned Zoox extended its commercial Las Vegas robotaxi service to Harry Reid International Airport, with pickup and drop-off near baggage claim at both terminals. The move follows its temporary federal exemption and start of paid rides in August.
**AI take:** Airport trips are frequent and operationally demanding, making sustained pickup reliability and fleet turnover a stronger maturity test than a demonstration.
### Qoder makes its Efficient model tier free for paid users
Qoder confirmed in an official announcement that the Efficient model tier is now free for paid users. The change reduces incremental model cost for some coding work inside paid plans, with exact scope governed by Qoder’s plan terms and later notices.
**AI take:** Quota improvements matter when they change real bills and model choices inside the tool, and deserve attention alongside model launches.
## Rumors
### ByteDance reportedly discusses 5–6GW of additional AI data-center capacity in Inner Mongolia
The South China Morning Post, citing a person familiar with the matter, reported that ByteDance was reportedly in preliminary talks with local data-center vendors to add five to six gigawatts of compute capacity in Ulanqab’s Jining district, targeting delivery by early 2028. The talks remain preliminary and ByteDance has issued no matching public statement.
**AI take:** If realized, this would be an enormous power and compute commitment; until contracts, energy sources, and permits emerge, five to six gigawatts is a target rather than built capacity.
### Crusoe reportedly closes more than $3 billion at a roughly $30 billion valuation
TechCrunch, citing Bloomberg, reportedly found that AI data-center and cloud provider Crusoe finalized a round of more than $3 billion at a roughly $30 billion post-money valuation, co-led by Atreides Management and Valor Equity Partners. Crusoe issued no company announcement during the reviewed window, so the amount and terms remain report-based.
**AI take:** Capital continues to move from the model layer into power, campuses, and GPU clouds, making long-term infrastructure commitments and utilization central competitive constraints.
### Anthropic reportedly nears a $15 billion pre-IPO credit facility
Investing, relaying Bloomberg reporting based on people familiar with the matter, said Anthropic was reportedly close to finalizing an expansion of its revolving credit facility to $15 billion, led by Morgan Stanley with several major banks participating ahead of an expected IPO. The report says details could still change, and Anthropic issued no final agreement.
**AI take:** A facility this large reflects the capital demands of frontier-model training, inference, and IPO preparation, but nearing finalization is not the same as drawing the funds.
### Accel is reportedly in talks to lead a $1 billion Thinking Machines round
Accel is reportedly in talks to lead a $1 billion Thinking Machines round at a valuation of at least $40 billion, according to TechCrunch citing its source and The Information. Neither company announced a completed round in the collection window, so the amount and valuation remain terms under discussion.
**AI take:** If completed, the deal would show continued willingness to fund elite teams and compute demand at high valuations, but talks are not cash in the bank.
