A Sirrius Office publication
Sirrius Atlas
AI builder updates
Important changes to the AI tools people use to build—explained plainly and linked to the original source. No rumours and no leaderboard drama.
We publish a brief only when it points to a vendor’s own product announcement, technical post, or release note.
Source ledger
Choose a company or tool, read what changed, and open the original source if it matters to your work. We update this page carefully rather than pretending it is a live feed.
Anthropic’s announcement positions Claude Opus 5 as its current model, with emphasis on coding, verification, debugging, and extended agent work.
Treat availability, model choice, and permissions as project decisions. Verify the exact capability surface in your own Claude Code environment before standardising a workflow.
Secure MCP Tunnel is customer-run software that keeps an outbound HTTPS connection to OpenAI so private MCP servers can remain behind existing network controls.
The interesting shift is architectural: agent tooling is moving closer to controlled private data access. Review authentication, data boundaries, and network policy before adopting it.
Google’s Gemini API release notes describe Gemini 3.6 Flash as improving token efficiency and coding or agentic-planning capability, with 3.5 Flash-Lite aimed at lower-latency subagent work.
Model routing can become a product-design lever: reserve higher-capability models for planning or difficult work, and use a smaller subagent where latency and cost matter.
Kimi’s technical blog introduces K3 and says the model is available through the company’s product and API surfaces; it also described a planned model-weights release by 27 Jul 2026.
Follow the rollout documentation rather than social comparison posts. Model availability, effort controls, and open-weight timing are all implementation details that can move quickly.
Qwen Code’s weekly update says `/learn` can take a directory, URL, conversation, or text source and organise it into a reusable `SKILL.md`, alongside code-review and Web Shell changes.
The durable idea travels beyond one vendor: capture repeatable work as a reviewed procedure, then keep its source material and permissions visible.
GitHub says BYOK is in public preview for supported models in Copilot Chat, Copilot CLI, and VS Code. The provider assumes the token bill while GitHub maintains the developer-tooling harness.
This is a governance choice as much as a pricing one. Teams should weigh provider contracts, policy control, visibility, and developer experience together.
SpaceXAI’s first-party announcement describes Grok 4.5 as its current model for coding, agentic tasks, and knowledge work, available through Grok and its API.
A new model claim is a reason to test, not a reason to migrate blindly. Compare the work that matters to you, including tool use, reliability, latency, and cost.
Manus says Plan Mode pauses before execution, expands a request into a structured Markdown plan, and waits for the user to edit, confirm, or dismiss it before building.
The useful pattern is broader than one product: review the approach before an agent changes a site or application, especially when requirements are incomplete.
DeepSeek’s API documentation announces open-weight V4 Pro and Flash models, a one-million-token context window, API availability, and integrations with agent coding tools.
Open weights and API compatibility broaden deployment choices, but teams still need to test data handling, hosting, tool behavior, and the exact model retirement timeline.
How to read this page
The useful question is not “which model won this week?” It is whether a release changes the delivery surface, governance posture, integration design, or practical workflow for your team. Check the source, test in your environment, and treat every inference as provisional.