AI Masterclass Podcast Week 28: from model launches to governed workflows

Week 28 did not give us a simple answer to the question of which AI model is best. It gave us something more useful: a clearer view of how providers are turning models into work layers that can be governed, measured, and stopped when something goes wrong.

This week's podcast covers Claude, Gemini, Grok, Manus, Mistral, OpenAI, and Perplexity. We separate practical progress from feature noise and ask what an organization can test without starting a large technology project.

New models do not solve the governance question

OpenAI introduced GPT-5.6 in three tiers: Sol, Terra, and Luna. The idea is to choose between frontier capability, a balanced tier, and a lower-cost option for high-volume work. At the same time, ChatGPT Work moves Codex technology beyond coding into documents, spreadsheets, presentations, browser tasks, and scheduled work.

xAI launched Grok 4.5 for coding, agentic tasks, and knowledge work. Its 500,000-token context window is smaller than Grok 4.3's listed one million. High reasoning effort is the default, and output costs more than with its predecessor. This is not a model to adopt by changing one name in a configuration file. Teams need to test cost, latency, and long inputs against their own workload.

The shared lesson is straightforward: route by task. An expensive frontier model may make sense for a difficult analysis and be wasteful for classification, extraction, or routine synthesis.

Reusable instructions are becoming operational assets

Google and Mistral made a less spectacular but more practical move. Gemini Enterprise gained reusable skills: packages of instructions, files, and sometimes scripts for recurring tasks. The launch has meaningful constraints, including an organizational allowlist and incomplete central administrator controls.

Mistral Studio goes further in the governance model. Prompts and skills can be versioned, owned, tested, labeled, traced in logs, and rolled back. That sounds administrative, but it solves a real problem. When an important instruction lives only in someone's chat, nobody can reliably answer which version produced a customer document or why the output suddenly changed.

A smaller organization does not need to build an elaborate agent platform to use this lesson. Start with one recurring instruction and treat it as a procedure: who owns it, what context may it use, how is the answer checked, and how do you restore the previous version?

Permissions and fallbacks are part of the feature

Manus launched Branch, which lets a reviewed conversation split into separate paths. One project record can produce an executive summary, a project plan, and a presentation draft without filling the main thread with competing edits. Manus also expanded its Google connection across Drive, Docs, Sheets, Slides, and Forms with Off, Read, and Full access modes. The important part is not the number of integrations. It is that read and write authority becomes an explicit choice.

Claude showed the same need from another direction. Claude Code received fixes around approvals, remote control, structured output, and safer file handling. The same research window also contained several incidents affecting models, MCP authorization, and agent surfaces. A workflow that works only while one model and connector are perfectly stable is not yet robust.

A useful fallback can be simple: save intermediate results, distinguish model failures from connector failures, and define when work should pause instead of retrying automatically.

Orchestration may matter more than one winning model

Perplexity is testing an adapted GLM 5.2 model that can escalate difficult parts to a stronger adviser model. Its argument is that lower-cost capability should handle the baseline while expensive intelligence is used selectively. The published results are vendor-run and need reproduction, but the principle matters beyond Perplexity: the quality and cost of the complete workflow matter more than the name of the model that started the task.

Perplexity also made GPT-5.6 and Grok 4.5 available through its product and API surfaces during the week. That shows how quickly models are becoming interchangeable components inside a larger system.

A sensible test after listening

Choose one recurring deliverable that currently takes 30–90 minutes, such as a weekly summary or meeting brief. Before selecting a tool, write down four things:

• Read boundary: which folders, documents, or systems may the AI use?
• Write boundary: may it only suggest changes, or update a document?
• Approval: who checks facts and external actions?
• Fallback: what happens after an outage, a cost increase, or a poor answer?

Run the same task three times and measure time saved, review effort, and errors. If the result holds up, the procedure can be formalized, for example through Tool Forge. If it does not, you have learned that cheaply before the workflow became critical.

This week's provider map

• Claude: user reflection, governance, and safer agent operations, alongside a clear reliability reminder.
• Gemini: the AlphaEvolve optimization agent, reusable skills, identity lifecycle through SCIM, and more detailed logging.
• Grok: a new coding and agent model that requires fresh tests for context, price, and reasoning effort.
• Manus: branching from reviewed context and clearer permission modes for Google Workspace.
• Mistral: governed prompts and skills, plus Robostral Navigate as a narrow but notable step toward physical AI.
• OpenAI: GPT-5.6 tiers, ChatGPT Work, and more built-in ways to coordinate tools and subagents.
• Perplexity: selective escalation between models and rapid distribution of new provider models.

This episode is an AI-generated masterclass built from Hammer Automation's deep weekly research into AI-provider updates and features, processed with NotebookLM. No Whisper transcription was used in the automated workflow.