Pricing, security tooling and agent performance moved together over the weekend. OpenAI cut GPT-5.6 Sol pricing while asking California for broader frontier-model safeguards. Anthropic put Mythos 5 into enterprise security scans, Nvidia reported a perfect public-set result for its AVO agent, and Chinese companies pushed multimodal models, video generation and robotics funding. The common thread is deployment: frontier systems are becoming cheaper, more specialized and more closely tied to physical infrastructure.
OpenAI cuts GPT-5.6 Sol pricing and presses for wider safeguards
OpenAI said it is reducing GPT-5.6 Sol API and credit pricing by more than 20% for the next three months, putting API rates at $4 per million input tokens and $20 per million output tokens. The temporary cut lowers the cost of trying a frontier model in production, but it also tests whether lower prices can broaden demand without permanently resetting customer expectations. Separately, Politico reported that OpenAI wants California to expand SB 53 safeguards, including monitoring frontier models while they are still in training. Pricing and oversight are moving in parallel as powerful systems reach more users.
Anthropic moves Mythos 5 into enterprise security scans
Anthropic said Claude Security scans now run on Claude Mythos 5. The feature is in public beta for Claude Enterprise customers, scans codebases for vulnerabilities and proposes patches for human review. Anthropic says scans are billed as standard token usage under the customer's existing plan, with no separate add-on. This is a focused deployment rather than a general model release. It shows how frontier labs are packaging specialized capabilities around high-value workflows where human approval, auditability and false-positive control matter as much as raw model performance.
Nvidia reports a perfect ARC-AGI-3 public-set result for AVO
Nvidia said its Agentic Variation Operators system achieved a 100.00 RHAE score across all 25 ARC-AGI-3 public environments, completing all 183 levels in 6,624 environment actions. AVO used Claude Opus 5 with persistent memory, supervision and a text-only representation of each environment. Nvidia cautioned that comparisons with other systems are not controlled ablations because their models, reasoning settings, memory and interfaces differ. The result is still important because it reinforces a practical lesson: agent benchmarks measure the surrounding harness as well as the underlying model. Memory and recovery loops can materially change what a frontier model accomplishes.
DeepSeek adds image understanding to V4 Flash
DeepSeek released the experimental V4 Flash Vision model on its API platform. The model accepts images alongside text, allowing developers to analyze screenshots, charts, documents and other visual inputs through the same API workflow. DeepSeek is presenting it as an experimental extension rather than a finished flagship. That distinction matters for buyers evaluating reliability, support and long-term compatibility. The release nevertheless closes a practical gap for developers who want multimodal input without moving to a separate model family, and it increases competitive pressure on providers that charge a premium for vision-enabled frontier systems.
Nvidia raises the stakes around models and infrastructure
The Wall Street Journal reported that Nvidia plans to use its $6 billion Poolside deal to build an open-weight US model aimed at competing with Chinese releases such as DeepSeek and Kimi. The reported arrangement includes access to Poolside's model-development technology and staff. In a separate development, Bloomberg reported that some Nvidia customers were warned of price increases above 15% for new AI systems, including Vera Rubin and Grace Blackwell configurations expected in early 2027. Both reports point to the same strategic shift: Nvidia is trying to influence the model layer while retaining pricing power in the infrastructure beneath it.
Alibaba turns work documents into longer generated video
Alibaba launched Wan3.0, a video model that can produce clips up to 30 seconds from documents, spreadsheets, slides and web pages, Reuters reported. The input formats are notable because they shift video generation away from isolated prompts and toward ordinary business material. A presentation, report or product page can become the source for a generated sequence without first being rewritten as a shot list. That makes the technology more useful for marketing and internal communication, while raising familiar questions about factual fidelity, rights clearance and disclosure when source material is automatically transformed.
XPeng robotics raises more than $900 million
XPeng said its robotics business raised more than $900 million in its first external funding round, according to Reuters. IDG and Gaorong Ventures led the round at a reported $6.3 billion valuation, with Tencent and Alibaba among the participants. The size of the financing shows that capital for physical AI is moving well beyond laboratory prototypes. Robotics companies need models, sensors, manufacturing capacity and service networks at the same time. Investors are therefore underwriting an industrial buildout, not just software development, with longer timelines and more execution risk than a typical application startup.
What matters today
The weekend showed that frontier AI competition is no longer one race. OpenAI is testing lower prices and wider rules, Anthropic is specializing models for security, Nvidia is proving the value of agent architecture while expanding its reach across models and hardware, and Chinese companies are funding multimodal software and robotics at scale. For teams adopting these systems, the important questions are shifting from whether a model can perform a task to how reliably it fits a workflow, what controls surround it and whether its economics remain stable after introductory pricing and benchmark demonstrations.