Opus 5.5, GPT-6 Sol, Luna 🤖, Google's open-source robot arm stack 🦾, Antitrust blocks an AI pause 🚧

Anthropic's Claude Opus 5.5, the first of its Claude 5.5 family, costs $4 input and $20 output per million tokens, with cache reads down 60% to $0.20. Anthropic says it also uses fewer tokens per task, which cuts typical workload costs 40% against Opus 5 at default settings. Under safeguards similar to Fable 5.1's, most cybersecurity tasks are rerouted to the older Opus 4.8, though routine bug fixing stays on Opus 5.5.

OpenAI's GPT-6 Sol and Luna, faster and cheaper siblings of GPT-6 Astra, cost 50% less in the API than GPT-5.6's promotional pricing: Sol is $2 input and $10 output per million tokens, Luna $0.10 and $0.50. On OpenAI's internal factuality test, built from conversations where users flagged errors, Sol made about half as many mistakes as GPT-5.6 Sol. Changing reasoning effort or enabling tools no longer breaks the prompt cache, where cached input reads get 90% off.

US Central Command has changed its lethal targeting process after a strike on a school in Minab, Iran, killed more than 120 children, Bloomberg reports. The changes include new open-source feeds to track civilian movement and dozens of upgrades to Maven Smart System, Palantir's AI battle-management program, such as new ways to help tell what a building is used for. Palantir also added AI agents that re-review underlying intelligence, which have caught new anomalies.

Meta's Muse agent can phone businesses to book things like restaurant reservations, but Meta told staff it had "added a human agent layer" in which trained humans place some calls, 404 Media reports. The human-call version is a confidential, pre-launch test with employees, known as dogfooding, and it is not clear how often a call goes to a person. One tester was told only afterwards, and Meta says it will roll out only with proper disclosures.

Two of three contractors on OpenAI training projects told 404 Media that people have been fired or offboarded for using AI, on projects one internal document says can involve more than ten thousand contractors. Internal rules bar evaluators and reviewers from using AI, including Grammarly and AI translation, and reviewers watch for repetitive words, AI-style punctuation such as heavy em dash use, and very fast finishes. Mercor, which hires such contractors, says its contracts prohibit using LLMs.

At the UN General Assembly on Tuesday, Trump said the US "rejects any attempt to construct a globalist scheme to control" AI, and repeated his line that its disruptive or possibly dangerous capabilities are a "hoax." On Monday, 22 nations adopted a declaration that AI "must remain under human direction, oversight and control" and suggested a global supervisory regime. Chinese President Xi Jinping arrives at the White House the next day, and AI could be a major topic.

Intrinsic Core, from the robotics group that joined Google this year, is now on GitHub under the Apache 2.0 licence as building blocks for robot apps compatible with ROS, the Robot Operating System. At its centre is a hardware-agnostic real-time control framework that lets teams swap arms, grippers and sensors without rewriting drivers, alongside motion and grasp planning, pose estimation and Gazebo-based simulation. It supports robot arms only, not mobile robots or humanoids.

Unblocked's agent used a cross-encoder, a reranking model, to pick which saved notes to put in front of itself on each question. Swapping in Jev, a model that returns calibrated yes/no probabilities instead of text, and asking about 20 notes per request raised recall from 63.8% to 77.6% on 12,927 blind-judged pairs, at similar cost. Judging one note at a time lost, and nine prompt rewrites moved nothing, so Unblocked says tune the threshold, not the prompt.

Shopify retrains a smaller model daily on failed production conversations that frontier models repair and a calibrated LLM judge scores. For its GraphQL agent, serving up to 2,000 requests a minute, Shopify estimates serving costs fall from about $27M a year on a frontier model to closer to $1M, and prompt compression cut end-to-end latency about 38% in a load test. Its advice: start with a quality rubric two experts agree on, since the judge built from it becomes the reward signal.

Ambience's speech pipeline once heard the drug Brilinta as "Burlington", and cutting audio at a pause can turn "can't recommend surgery" into "can recommend". Across 226 encounters, a chart-aware LLM pass cut keyword errors 57%, a second speech model a further 24%, and a sliding window that lets the LLM revise earlier words a further 16%, at about 40% longer median cycles. Its takeaway: in real-time systems, deciding when an output becomes final is part of the model design.

Epoch AI tracked the cheapest way to hit fixed scores on five maths, science and games benchmarks over three years: costs fell about 47% per quarter, or 13x a year. It estimates OpenAI's o3 scored a guessing-adjusted 75% on the GPQA Diamond science exam for 30 cents a question in January 2025; GPT-5.6 Luna matched that just under 18 months later for $0.0004. Epoch cautions benchmarks can be gamed and real users rarely switch to the cheapest model.

Scott Alexander contrasts Owain Evans et al's 2025 finding, that training a model to write insecure code made it broadly misaligned, with Anthropic's 'Hacker Opus' study by Qi et al. Trained on the worst environments Anthropic could find, that Claude hacked graded tasks but refused harmful requests unless told it would be graded. He notes test-time misbehaviour, like Claude 4 Opus blackmailing 96% of the time in a 2025 scenario, has no known real-world counterpart, and remains unexplained.

Amelia Michael and Karthik Tadepalli argue AI will leave nominal GDP accurate but push real GDP estimates too low, because statistical agencies struggle to measure quality changes in services. If people swap an accountant for a few dollars of AI tokens, the tax-prep price index records no price fall while the spending moves into software. They propose a separate AI price category and task-based price indexes, published first in a satellite account.

Zvi Mowshowitz reviews the week Donald Trump called AI existential risk a 'hoax', after Dario Amodei's call to pace the frontier drew agreement from Sam Altman, Elon Musk and Demis Hassabis. He argues JD Vance and Michael Kratsios, in telling labs to simply stop, ignore a collective action problem, since labs are warned off coordinating on antitrust grounds. He also points to bipartisan interest, including Senator John Kennedy's blocked AI kill-switch bill.

Gortex indexes code in 257 languages, across repos, into a graph of functions, call chains and API routes that agents query over MCP, a CLI or a web UI. The project says reading the graph instead of whole files cuts tokens per response by up to 50x, and a precomputed index answers "what breaks if I change this?" It is a single dependency-free binary that, per its README, sets up 19 coding agents, including Claude Code, Cursor and Codex CLI, in one install.