Muse Spark 1.3 landed September 2 in Muse Code and the Meta Model API, per Meta AI Research, at $1.25 per million input tokens and $4.25 per million output, with a 1M context window and an 88 percent prompt-cache discount, per Artificial Analysis, which scores it 48 on its Intelligence Index (13th of 202 models). Meta engineers report roughly 20 percent fewer tool calls and 25 percent fewer tokens than 1.2 on the same coding work. Worth a routing test if you run high-volume repeated-context jobs, because the cache discount is where the money actually shows up.

Enterprise Frontier Safeguards, announced September 1, keeps monitoring data in your own Amazon S3, Azure Blob, or Google Cloud Storage account under your encryption keys and audit logging, and routes detection signals to your reviewers rather than Anthropic's. Anthropic says more than 100 customers shaped it, including the CISO group behind the largest US banks plus Comcast, KPMG, Mastercard, Salesforce, and Visa. That answers the 30-day retention policy that has kept legal, healthcare, and finance teams off Fable-class models since Fable 5. Rollout is phased and starts later this fall, so today this is a procurement conversation, not a toggle you flip. Anthropic's writeup.

Three changes shipped September 3, per the changelog: a new /diff command opens a side-by-side panel while Claude is still mid-edit, /cost now names the probable cause of a cache miss (changed tool definitions, edited system prompt, idle past TTL), and org policy state surfaces in /status and claude doctor. If your agent-loop spend swings between otherwise identical runs, the cache-miss line is the first place to look. The policy line kills a whole category of "why is this blocked" tickets.

Justice filed a statement of interest on September 1 in the Southern District of New York multidistrict litigation against OpenAI, arguing that a contrary ruling would distort copyright law and weaken US competitiveness and national security, as reported by IPWatchdog on September 3. First time the federal government has taken a side in training-data litigation. It binds no judge, but if you are building on a model trained on scraped text, the downside scenario just got less likely and your indemnity clause just got less load-bearing.