Gemini 4 Carbon starts a staged rollout as Copilot expands cross-repository work
Gemini 4 Carbon brings new cyber-defense capabilities first to trusted teams, while MiniCPM-OCR separates document parsing from quality assessment. GitHub updates agent development through cross-repository sessions, a shared SDK, and organization-owned authentication.
This edition also covers lazy-loaded SageMaker containers, legal templates for Claude, more Siri languages, Firefox page tasks, development activity data, and open tools for medical research.
01
Models and research
3 stories
2026-10-28Google DeepMind
Gemini 4 Carbon begins with access for trusted cyber defenders
Google DeepMind announced Gemini 4 Carbon with a focus on long-running engineering work and cybersecurity. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for validating exploitability, and 71.0% on CWE-bench for producing patches. The three scores describe different security tasks, rather than one combined accuracy measure.
Access begins this week with selected cyber defenders, followed by additional trusted partners through the Fairwind Program. Broad availability for developers and consumers is still to come. Google lists introductory prices of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 respectively after the introductory period.
2026-10-28OpenBMB
MiniCPM-OCR separates document parsing from quality assessment
OpenBMB released MiniCPM-OCR as two complementary models. The 1.3-billion-parameter MiniCPM-OCR 1.0 parses Chinese and English documents, PDFs, tables, and formulas; the 2-billion-parameter Jev model assesses the results. The team reports scores of 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2 for the parsing model.
Jev takes an image together with its OCR output and assigns quality scores from zero to one to text, formula, and table regions. These scores help identify content that needs human review. Model weights and local deployment instructions are available, and parsing can produce Markdown or JSON, allowing recognition and quality assessment to form one document workflow.
2026-10-28Cartesia
Cartesia tests learned byte grouping in updated H-Net scaling research
Cartesia published new scaling experiments for H-Net, replacing fixed tokenization with learned grouping of bytes. In its one-stage comparison with a BPE Transformer, the backbone architecture is held constant while the outer networks are matched for parameters and compute. H-Net requires about 22% less compute to reach the same held-out pretraining loss.
The advantage changes with the evaluation distribution. Across ten programming languages absent from the training mixture, the fitted compute-efficiency multiplier has a median of 3.10. That result describes particular out-of-distribution evaluations, rather than a universal threefold speedup. The model learns to compress raw bytes into variable-length semantic units before the main network processes them.
02
Development and infrastructure
4 stories
2026-10-28GitHub
GitHub adds cross-repository projects and changes enterprise Copilot access
At Universe 2026, GitHub introduced Copilot projects for managed, parallel agent sessions across repositories within the Copilot app. Copilot Fusion chooses and orchestrates models by task. Citing the Artificial Analysis Coding Agent Index, GitHub says Fusion Max finishes within two points of Opus 5.5 Max at about half the cost and with 40% less time per task.
GitHub Enterprise customers can now access Copilot features without a separate Copilot subscription and pay for the AI credits they consume. The option is available through sales, with self-service to follow; existing Business and Enterprise plans can remain in place. Cloud agent sessions can also continue across the app, VS Code, the CLI, and the web.
2026-10-28GitHub
Copilot SDK moves toward a shared Microsoft agent foundation
GitHub announced a phased transition from GitHub Copilot SDK to Microsoft Copilot SDK. Its role is expanding from the interface to GitHub Copilot’s harness into a programmable interface for the shared agent foundation used across Microsoft and GitHub products. Developers can use that foundation to build and customize their own agent applications.
Control over prompts, tools, policies, and application experiences remains with developers. Repository updates are preparing for a move to the Microsoft organization, alongside branding and documentation changes. Existing package names, namespaces, and some documentation may retain the GitHub name during the transition. The announcement requires no immediate integration changes from existing users, and support remains unchanged.
2026-10-27GitHub
Copilot previews organization-owned authentication for backend agents
GitHub Copilot’s server-to-server authentication is now in public preview. A GitHub App installed on an organization can make Copilot SDK requests on the organization’s behalf without a signed-in employee’s credentials. Developers grant the app the read-only Copilot Requests organization permission and authenticate with a short-lived installation access token.
Usage and charges belong to the organization that owns the installation, with its model policies, budgets, and spending controls still applying. Administrators can inspect authorized apps, identify their requests through the app_client_id field in AI usage CSV reports, and revoke access by suspending or uninstalling an app. This gives persistent services and unattended agents an identity independent of an individual user.
2026-10-27AWS
SageMaker starts jobs before the full container image is downloaded
AWS added SOCI image indexes to SageMaker AI training and processing jobs. Containers fetch the files needed to start first and load the remainder as code accesses them. Across eight compressed images ranging from 1.8 to 35.4 GB, AWS measured startup waits of 3–11 seconds with lazy loading. Jobs fall back to a full download when an index is missing or unavailable.
Users can select indexed AWS Deep Learning Containers or index their own images without changing entry points or data channels. In the reported image-classification and language-model fine-tuning tests, the first epoch took an additional 3–16 seconds, with subsequent epochs within normal run-to-run variation. After that initial overhead, each complete job finished roughly one to six minutes earlier.
03
Applications and workflows
3 stories
2026-10-28Anthropic
Claude adds reusable legal artifacts for review and intake workflows
Anthropic introduced a gallery of ready-to-use, adaptable artifacts for Claude for Legal. They cover tabular contract review, citation checking, regulatory tracking, and intake of legal requests from business teams. Each opens with an example before a team adds its own documents and connectors. Outputs link back to sources, with a lawyer reviewing Claude’s first pass.
Teams can tailor tools to internal playbooks and share them within the organization or with clients and partners, using sharing controls to maintain information boundaries. Anthropic says it has added more than 60 legal connectors since launching the product in May. Partner examples include Harvey’s intake interface and tools that organize contractual obligations and regulatory changes into collaborative boards.
2026-10-26Apple
Siri AI adds five languages with the 27.2 updates
Apple added French, Japanese, Korean, Portuguese, and Spanish to Siri AI with its 27.2 software updates. The assistant can draw on personal messages, email, and photos and act on onscreen content. This update expands language availability beyond English, rather than launching the entire Siri AI experience again.
Siri AI remains in beta and requires an Apple Intelligence-capable device configured to a supported language. Traditional Chinese, Norwegian, Turkish, and Vietnamese are planned by year-end. Availability still excludes iOS, iPadOS, and watchOS in the EU and users under 13. Some capabilities that depend on server-side models are subject to daily usage limits.
2026-10-27Mozilla
Firefox Smart Window adds form suggestions and scheduled page watching
Mozilla added form-filling suggestions based on tabs selected by the user in Smart Window. On forms with at least four supported text fields, users can choose source pages and have relevant details, such as work history, extracted. Gemini 3.1 Flash Lite processes the selected page information. Users review the suggestions, and Smart Window does not submit the form itself.
A second feature uses /watch to check for requested changes daily or weekly across up to five pages per task, such as sales or new job listings. Checks run only while the browser is open. Both features begin in the English-language desktop beta in the US and Canada. Form suggestions currently support text fields; ordinary saved autofill information remains on the device.
04
Industry and research practice
2 stories
2026-10-28GitHub
Octoverse reports agent-assisted starts for over 40% of September pull requests
GitHub’s 2026 Octoverse report records 17.9 billion pushes, 34.3 billion commits, and 1.3 billion merged pull requests over the previous twelve months. Its telemetry shows developers starting more than 40% of pull requests with an agent in September, compared with 17% in March. This describes how work begins, not a claim that AI independently completes all of the code.
In September, 61% of Copilot users who selected a model themselves worked with more than one model. Among repositories created during the year, those with an agent configuration file averaged eight times as many commits as new repositories without one. These are observed differences in platform activity, rather than an isolated measurement of an equivalent productivity gain caused by agents.
2026-10-27NVIDIA
C-BRAIN and NVIDIA develop open tools for Alzheimer’s research
C-BRAIN, a consortium of 18 institutions, is working with NVIDIA on open AI tools for neurodegenerative-disease research. Its centralized and federated infrastructure can keep some institutional data on site, with BioNeMo and MONAI planned for research pipelines. Current computing resources include 64 on-premises NVIDIA data-center GPUs and cloud nodes on Azure.
The existing OpenScientist tool develops hypotheses, runs analyses, and produces auditable results from a research question and dataset. InsightEngine checks evidence using literature, user-supplied data, and a biomedical knowledge graph. The consortium next aims to train a foundation model combining imaging, omics, and clinical information. These are research infrastructure and analysis tools, rather than validated treatments.