TL;DR Google rolled out video generation and lower pricing for Gemini Omni 1.2 Flash, while Thinking Machines introduced Inkling 2 Small with Apache 2.0 weights planned. Meanwhile, OpenAI launched GPT-6 Cyber for security auditing, and Anthropic published fresh agent memory research.
What actually moved today
Good morning, everyone! Grab your coffee and let's dive into the latest updates hitting the AI engineering ecosystem. Today brings a fascinating mix of concrete model drops, practical pricing shifts in media generation, and some heavy-hitting security releases. We are seeing real movement in both proprietary APIs and open-weight architectures, giving us plenty of new endpoints to test out in our local environments.
The model watch
Gemini Omni 1.2 Flash
Google has officially rolled out video generation up to 15 seconds for Gemini Omni 1.2 Flash, replacing older endpoints like gemini-omni-1.1-flash. Alongside this capability, they dropped 720p API pricing by 30 percent to about $0.07 per second and introduced a new Interactions API for in-dialogue edits. You will want to re-run your multimodal evals to check if the cheaper generation holds up for your specific video iteration workflows.
Inkling 2 Small
Thinking Machines reported a 594B parameter MoE model with 17B active parameters under the name Inkling 2 Small. It is currently available in Tinker for inference and fine-tuning, with Apache 2.0 weights planned for the near future alongside demonstrations of automated fine-tuning for Qwen3.5-4B. It sits alongside models like DeepSeek V4.1 Flash and GLM-5.3-Flash as a strong baseline for open weights performance and advanced agentic tool usage.
GPT-6 Cyber
OpenAI officially launched GPT-6 Cyber for approved Daybreak Red users. During early testing, it successfully found 948 vulnerabilities in the Linux kernel. This kind of automated security discovery at kernel scale shifts how vulnerability research and automated code auditing are conducted in production systems.
Anthropic Agent Memory and R&D Speed
Anthropic published research on cross-session agent memory and dreaming mechanisms using Claude 3.7 Sonnet and Opus 5, following earlier iterations of Opus 4.7. Additionally, METR assessed that Anthropic internal model development speed doubled, driven 80 percent by AI use in R&D. Better session memory directly improves long-running autonomous agent reliability.
Rumours, clearly labelled
As always, keep in mind that unverified claims and leaks should be taken with a grain of salt until official documentation drops. Here is where things currently stand regarding unconfirmed reports:
| Item | Status | Basis | Notes |
|---|---|---|---|
| Collaborative mode for Claude Code Projects | Unconfirmed | Partner screenshot and social media hint | Team members noted multiplayer is not yet available, though interface updates hint at future work. |
| US AI market regulation shifts | Unconfirmed | Recent political election results | Market observers are speculating on policy changes, but no concrete legislative text has been finalized. |
Odd finds and small tools
Sometimes the smaller, lesser-known repositories provide the most immediate utility for our day-to-day work. An open-source face recognition SDK was recently released on GitHub, giving Python developers a straightforward way to add facial recognition without relying on external cloud APIs. It is a neat self-hosted alternative if you are building offline security or attendance logging tools.
What I would try this weekend
If you have some free time over your coffee this weekend, spin up a quick script to test the new Interactions API on Gemini Omni 1.2 Flash to see how smoothly you can perform in-dialogue video edits. Alternatively, grab the new Python face recognition SDK and spin up a local container to test its offline processing speed against your webcam feed.
Where this leaves us
We are continuing to see rapid iteration in multimodal generation costs and specialized security agents, but managing long-term session state remains an ongoing engineering challenge. While models are getting better at remembering context within a dialogue, bridging that across multiple days and sessions still requires custom database plumbing on our end.
How do you currently handle long-term session memory in your production AI agents?
Top comments (0)