AI products are moving in two directions at once: cheaper model access and more ambitious delegation. Meta introduced its first paid model API while OpenAI turned ChatGPT into a long-running work agent. A copyright fight shifted toward evidence handling, and two benchmarks made causal abstention and expert proof review part of the evaluation target.
1. Muse Spark 1.1 gives Meta its first paid model API
Meta placed Muse Spark 1.1 into public preview as its strongest model for coding and agentic work. Access begins through a developer waitlist, with $20 in initial credits and rates of $1.25 per million input tokens and $4.25 per million output tokens. The API currently runs only on Meta's own infrastructure, with no third-party marketplace access.
Charging for a proprietary model breaks from the open distribution strategy associated with Llama. Meta says an open-source Muse variant remains in development but has provided no release date. Its performance claims also come from company testing, so the immediate evidence is a new commercial channel and price point, not independent parity with frontier rivals.
Sources: CNBC on Muse Spark 1.1 access and pricing · The New York Times on Meta's paid-model shift
2. ChatGPT Work turns connected apps and files into an execution surface
OpenAI launched ChatGPT Work, an agent that can gather information across connected applications, operate for hours, and produce finished spreadsheets, slides, documents, and web apps. It can continue recurring workflows through Scheduled Tasks, while the desktop version can use local files, applications, a built-in browser, and computer control.
The launch expands the security boundary far beyond conversational prompts. Enterprise controls cover plugins, network access, connected actions, and spending, and an auto-review layer inspects important tool calls before execution. Those controls reduce exposure only when permissions, approvals, data retention, and audit records are configured around each delegated workflow.
Sources: OpenAI's ChatGPT Work announcement
3. Publishers ask a court to sanction OpenAI over disputed evidence
The New York Times, New York Daily News, and other publishers asked a Manhattan federal court to sanction OpenAI in their copyright litigation. They allege OpenAI misrepresented its ability to search models and logs for their works, then deleted billions of relevant ChatGPT conversations or made them unsearchable. The requested remedies include fees and adverse findings about alleged copying.
OpenAI called the allegations false and said the publishers are pursuing private conversations unrelated to the case. No judge has decided the motion, and the underlying infringement claims remain contested. The dispute makes preservation architecture, searchability, and representations about internal data systems part of the legal risk surrounding model training.
Sources: Reuters on the publishers' sanctions motion and OpenAI's response
4. CausalDS rewards agents that know when data cannot answer a question
CausalDS is a new benchmark for causal reasoning in agentic data-science workflows. It generates synthetic structural causal models, observational data, and realistic natural-language settings, then derives tasks across all three rungs of Pearl's causal hierarchy. Many tasks require coding and multiple tools under imperfect observations.
Abstention is scored as a first-class outcome when the available evidence does not warrant an answer. That design separates causal judgment from fluent statistical analysis and reduces simple memorization through newly generated structures. The 55-page paper is a preprint, so independent runs and released evaluation artifacts will determine how reproducible its difficulty is.
Sources: The CausalDS benchmark preprint
5. IMProofBench doubles its private set of research-level proofs
Version two of IMProofBench expands the private benchmark from 39 to 77 peer-reviewed mathematical problems and evaluates 14 models. Each problem requires a detailed proof, includes final-answer subproblems for automated scoring, and runs inside an agent framework with web search and mathematical software such as SageMath.
The revision adds human expert assessment, inter-rater analysis, tool and token measurements, and a non-agentic ablation. Keeping problems private limits contamination, but it also prevents outsiders from fully inspecting task construction and grading. The benchmark is strongest as a curated measurement program, not a standalone claim that current models can conduct reliable mathematical research.
Sources: The revised IMProofBench preprint