OpenAI cuts GPT-5.6 prices: Luna down 80%, Terra down 20%
Three weeks after launch, OpenAI cut GPT-5.6 Luna’s API price by 80% (to $0.20 per million input tokens, $1.20 output) and Terra’s by 20% ($2 / $12); Sol stays where it was (see also CNBC). OpenAI credits efficiency gains, including using its own models to rewrite and optimize production inference code and speed up token generation. The competitive context is plain: Google and the Chinese labs keep shipping cheap models, and API price is where that pressure lands first. The more durable fact is the cost engineering itself: models optimizing model inference is now a production practice that can fund an 80% price cut, not a demo.
Judge: the government still has no evidence for Anthropic’s “supply-chain risk” label
Background: in February, President Trump ordered federal agencies to stop using Anthropic products, and the Department of Defense separately designated the company a “supply-chain risk,” cutting it off from military contractors (TechCrunch); Anthropic sued in March, and U.S. District Judge Rita Lin blocked the ban temporarily. At this week’s hearing she said the government still hasn’t produced evidence for the designation, and she is now weighing whether to make the block permanent. She flagged two specifics: the government citing Anthropic’s public criticism of the administration as part of its rationale, which she called “really troubling” for the precedent of punishing dissenting contractors, and the technical claim that Anthropic could remotely disable delivered models, for which no proof exists. Worth following: if a security label can be applied without surviving evidentiary review, it stops working as a security tool and starts working as a political one.
DeepMind’s Gemini Robotics ER 2 is an orchestration brain for robots
The new model sits at the top of the robot stack: it watches continuous video to track task progress (57.4% on five-level progress classification, 91.3% on moment-finding), responds at sub-second latency, orchestrates low-level vision-language-action models and APIs as callable tools over the Gemini Live API’s bidirectional streaming, and coordinates dissimilar robots through shared semantic understanding. In the demos, a Boston Dynamics Spot fetches objects by orchestrating navigation and manipulation APIs, and Apptronik’s Apollo 2 humanoid hands off tasks to a Franka arm; the model is live in the Gemini API and AI Studio. Architecturally this is the agent tool-use pattern moved into embodied AI: an LLM plans and delegates, specialized models execute, and multi-robot collaboration becomes a multi-agent problem.
Google: two Chrome releases fixed more security bugs than the previous two years combined
Chrome milestones 149 and 150, both shipped in June, fixed 1,072 security bugs, more than the 1,036 across the previous 23 milestones (TechCrunch). LLMs now run through Google’s whole vulnerability pipeline: finding bugs (including a sandbox escape that sat in the codebase for 13 years), reproducing and triaging reports, generating candidate patches that critic agents review and test-writing agents cover, with human developers keeping final approval. The guardrails are concrete too: models scan source code at rest on isolated machines with no internet access. This is some of the hardest quantitative evidence yet for AI in defensive security, and since the same capability pointed the other way is automated offense, this pipeline is the best public gauge of whether defense keeps pace.
Constitutional midtraining: alignment inserted early survives longer, but not everywhere
A new paper tests a live dispute head-on: is alignment just a shallow coat applied in post-training? The authors built a 394M-token corpus from Anthropic’s published constitution and inserted it during midtraining (the training phase between pretraining and post-training) at 120B scale. The results cut both ways: gains in blackmail resistance survived later benign fine-tuning (a 17.5-percentage-point advantage), while resistance to in-context pressure and value conflicts faded after SFT. The paper also compared ways of organizing the same constitutional content — varying its curriculum ordering within midtraining, and including or omitting deliberative-reasoning text built around the constitutional principles, a 2×2 grid of four variants — and found little difference between them: what mattered was whether the content entered the training data at all, not in what order or form. Capability benchmarks showed no cost. For open-weight models this is directly useful: downstream fine-tuning eroding alignment is a known failure mode, and moving alignment earlier in the pipeline now has data behind it as a mitigation, even if it doesn’t protect every dimension.
Tested: distilling DeepSeek into GPT-OSS doesn’t carry the censorship along
The AI startup CTGT distilled DeepSeek V4 Flash’s outputs into GPT-OSS-120B to improve financial reasoning, then measured whether political censorship transferred with it. The teacher scored 45.45 points more censored on China-sensitive prompts than on structurally matched controls (152 prompt pairs, four AI judges cross-checked against human raters); the distilled student’s own gap was just 2.58 points, essentially matching the untouched base model’s 3.94, a difference that isn’t statistically significant. Read the scope carefully: the training corpus contained zero China-sensitive content, so this shows censorship doesn’t transfer “subliminally” through unrelated domain data; distill on a corpus that touches those topics and the answer could differ. It’s also a single vendor-published study, not peer-reviewed.
Okta buys Permiso to watch what AI agents do after they log in
Okta signed a definitive agreement to acquire cloud identity security company Permiso; terms weren’t disclosed, though TechCrunch reports just under $200 million, nearly all cash, with closing expected in Okta’s fiscal Q3 2027 (August to October 2026). Permiso monitors identity behavior after login across multi-cloud environments, watching human accounts, service accounts, and AI agents for stolen credentials, privilege misuse, and lateral movement. The signal here: non-human identity management is converting into real acquisition dollars, because an agent acting on valid credentials sails past login-time checks, so the monitoring has to move to the behavior layer.
LinkedIn ships a “seems like AI slop” report button and retires its own AI writer
LinkedIn added a “seems like AI slop” option to every post’s menu: a report doesn’t take the post down or trigger enforcement on its own; it feeds a signal LinkedIn uses to tune its detection models. The company also removed its “enhance your post” AI writing feature and replaced it with a proofreading tool, and says it blocks hundreds of thousands of automated comments daily. I work in content moderation, and two choices here look right to me. First, user reports feed the models instead of triggering enforcement directly, which matters because “looks AI-written” is a judgment annotators often disagree on, and direct enforcement would take down plenty of mediocre human writing. Second, on the old tension of policing AI content while selling AI writing tools, LinkedIn picked a side: kill generation, keep proofreading.
GCC adopts an AI policy: no substantial LLM-generated contributions
The GCC steering committee accepted its AI working group’s policy: contributions that are “legally significant” (roughly 15+ lines under GNU’s copyright guidelines) and contain or derive from LLM-generated content will be declined. Maintainers may still accept LLM-generated test cases, and using LLMs for research, analysis, bug hunting, and patch review stays unrestricted as long as the output stays out of contributions; a periodic review is planned (details in LWN’s report). Debian, for comparison, is still mid-vote on its own LLM policy, with proposals ranging from a complete ban to acceptance with disclosure. GCC’s answer, restricting what enters the code over copyright and provenance while leaving tool use itself open, is now the position other projects will measure theirs against.
One line for today: CTGT’s distillation test and the constitutional-midtraining paper probe the same question from opposite ends, and both point the same way: where a behavior enters the training pipeline shapes whether it transfers and how much later training it survives. On this evidence, alignment and censorship look less like intrinsic traits of a model and more like artifacts of specific training stages.