Claude text gets a watermark, Pangram spots AI images

Pangram released Pangram Image 1.0 on 4 November, a detector that marks with a heatmap where AI-made content sits in a picture, and Pangram Image Flash 1.0, a faster model for AI video. In Pangram's own tests, Image 1.0 scored 99.96% on 1,200 AI images made by 12 generators to mirror 100 real photos, and Flash scored 98.7% on the DF26 benchmark. It now checks images as small as 256x256, with lower confidence below 512x512. You can use it on Pangram's site, in its Chrome extension or through its API, which now takes images and video.

OpenAI released 2 image models to its API on 4 November, GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare, for generating and editing on /v1/images/generations and /v1/images/edits. They add a complexity setting for how much the model plans, SVG output and animated GIFs; Flare can return a low-quality image in as little as 6 seconds, and Sunburst gives higher quality and more consistent edits when quality is set to medium or high. Both list $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. In ChatGPT and Codex, Images 2.6 is on every plan.

OpenAI wrote on 4 November that a research prototype scanned about 7% of the latest Linux kernel in mid-October and reproduced 1,724 flaws, 948 of them distinct ways for a local user without privileges to become root. The run cost about $1.21 million in API credits. When the scan and OpenAI's own analysis finished 10 days later, 61 of them were patched upstream, 43 had patches in staging and 299 were being reported and discussed in public; OpenAI is working with its Patch the Planet partners to send patches for the rest gradually. No fixed kernel version is named, so apply your distribution's kernel updates as they arrive.

Since 4 November, a Databricks job can have an AI Function task that runs ai_parse_document, ai_extract or ai_classify on files in a volume or rows in a table, with no SQL to write. Each run handles only the inputs that are new or changed since the last one, and the results land in a Unity Catalog table that later tasks can read. It is in Beta.

Ricursive Intelligence published Ripple-65B Preview on Hugging Face on 4 November, a 65-billion-parameter dense model for reasoning and agentic coding under the Apache 2.0 licence. It takes up to 1 million tokens of context, and Ricursive says it fits on hardware such as NVIDIA's DGX Spark, with an NVFP4 checkpoint for running it locally. By Ricursive's own figures it scores 80.5 on Terminal-Bench 2.1 and trails some models on MMLU-Pro and GPQA Diamond.

Since 3 November, text from Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries Anthropic's text watermark, in Claude's own products and on AWS, Google Cloud and Microsoft Foundry; on Amazon Bedrock the rollout for these 3 completes by 10 November. Anthropic's list of covered models also includes Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5. The mark adds no tokens or hidden characters, does not change the meaning or quality of the answer and carries no information about you. Checking text for it is in private preview, for organizations such as regulators, media, fact-checkers and researchers, and for companies that must verify it under EU law; the free Claude Content Checker reads Content Credentials in files such as PNG and JPEG, and a mark found means only that the content may have been generated or processed by Claude. On Sonnet 4.6, a request that sets temperature, top_p or top_k to a non-default value may come back unmarked, so leave all 3 out if you rely on it.

Claude Code 2.1.317, from 3 November, asks before a web or shell tool call sends your account email address over the network. /schedule now warns when a new cloud scheduled task gets connectors you did not ask for, and says how to remove them. Version 2.1.318 the next day fixes claude -p and Agent SDK sessions that reported success when an API error cut off a reply while a mod hooked turn.step.

Since 3 November, Cortex AI Gateway is generally available for inference, observability and cost management: one governed endpoint for an organization's AI clients, which takes requests in the OpenAI Chat Completions or Anthropic Messages format. Each request is traced, with its timing, tokens and errors in AGENT_TRACE_TABLE, and spend is tied to the user who made it in AI_GATEWAY_USAGE_HISTORY. Shared budgets track spend by team, and per-user quotas block a user at a limit. The release also adds trace analytics in Snowsight, in preview, and tools for the gateway remain in preview.

Since 2 November, the Messages API can keep a conversation on Anthropic's side, in beta with the message-threads-2026-08-12 header, so each later request sends only the new messages. Anthropic says this gives long conversations and agent loops smaller requests and a lower time to first token. A thread expires 5 minutes after its last request and is deleted within 1 hour of its last use. It works on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Claude in Microsoft Foundry.

Google deprecated gemini-2.5-pro and gemini-2.5-flash on 2 November and took both out of AI Studio's model picker. From 1 December 2026, API requests to gemini-2.5-pro go to gemini-3.8-flash and requests to gemini-2.5-flash go to gemini-3.5-flash-lite, billed at the new model's price. Google lists breaking changes: function calling and thought-signature validation, thinking_level replacing thinking_budget, and removed sampling parameters and prefilled model turns. Change the model string yourself and test before that date.