This observation covers changes that were formally released, took effect or could be verified over approximately the day to 31 October 2026. Four items are included.
Microsoft releases the Mage-VL-Cowork computer-use model
The Microsoft Mage team released the Mage-VL-Cowork model weights, code and technical report on 30 October. The roughly four-billion-parameter model uses one policy for screenshots, graphical interface actions and a persistent Bash and Python environment. The published model files carry the Apache 2.0 licence. Compared with systems that call an external vision model or separate graphical actions from terminal execution, this delivery provides a downloadable unified policy and a documented training recipe.
Microsoft reports a 66.0% result on 361 OSWorld-Verified tasks under a 50-step budget, averaged over two runs and excluding Google Drive tasks. The released package includes Transformers loading instructions and two safetensors weight shards totalling about 9.48 GB. The training process used a five-stage curriculum, 293,806 synthetic desktop and web tasks, and 318,837 retained trajectories. These performance, training-cost and efficiency figures are developer-reported and have not received a public independent reproduction. Loading the model requires trusting repository custom code; desktop permissions, terminal access and isolation remain deployment responsibilities.
OpenAI's Evals platform becomes read-only
OpenAI announced the deprecation of its older Evals platform on 3 June. On 31 October, existing Evals content became read-only under the published schedule. Developers can still inspect existing material during the transition, but can no longer treat the platform as a writable evaluation workspace. OpenAI plans to close the Evals dashboard and API on 30 November and identifies Datasets and Promptfoo as migration paths. The change applies to the older Evals platform; it does not remove evaluation methods or the evaluation concepts used elsewhere in OpenAI's documentation.
Mistral API retires Z.ai GLM 5.2
Mistral's model lifecycle records show that the third-party model identifier zai-glm-5-2 entered deprecation on 29 September and retired on 31 October. This change applies only to that model identifier on Mistral's hosted API; it does not establish that Z.ai has withdrawn GLM 5.2 weights or services from other platforms. Mistral names zai-glm-5-3 as the replacement and states that it has the same price. Integrators still need to recheck prompts, function calling, structured output and long-context behaviour rather than assume byte-for-byte compatibility.
Databricks closes the older Genie MCP beta endpoint
Databricks made Genie One MCP generally available through Unity Gateway on 25 September under the service name system.ai.genie_one_mcp, with Unity Catalog permissions enforced on each request. On 31 October, the older beta endpoint at /api/2.0/mcp/genie reached its scheduled sunset. Existing agents and MCP clients must use the Unity Gateway address; the replacement address includes the workspace identifier and uses the ai-gateway OAuth scope. This is an interface migration and retirement of the old path, rather than a new Genie model-capability release.
Sources
Microsoft Mage Team, Mage-VL-Cowork model card, 30 October 2026, https://huggingface.co/microsoft/Mage-VL-Cowork
Microsoft Mage Team, Mage-VL-Cowork technical project page, 30 October 2026, https://microsoft.github.io/Mage/cowork/
OpenAI, Deprecations, verified 31 October 2026, https://developers.openai.com/api/docs/deprecations
OpenAI, Working with evals, verified 31 October 2026, https://developers.openai.com/api/docs/guides/evals
Mistral AI, Changelog, verified 31 October 2026, https://docs.mistral.ai/resources/changelogs
Mistral AI, Z.ai GLM 5.2 model page, verified 31 October 2026, https://docs.mistral.ai/models/zai-glm-5-2
Azure Databricks, Genie One MCP server, verified 31 October 2026, https://learn.microsoft.com/en-us/azure/databricks/agents/mcp-tools/genie-mcp
Azure Databricks, September 2026 product release notes, 25 September 2026, https://learn.microsoft.com/en-us/azure/databricks/release-notes/product/2026/september