MiMo 2.6 opens, Grok 4.7 ships: Sept. 22

Two new AI models give you cheaper candidates for routine marketing work, while an 80-minute Claude incident shows why no customer workflow should depend on one model. In about 30 minutes, you can start a controlled MiMo or Grok test, add a failure route to your automation, and keep the current production model untouched.

Xiaomi opens MiMo-V2.6 models and training resources

Xiaomi released two open models on September 22: MiMo-V2.6-Pro for harder work and MiMo-V2.6-Flash for lower-cost jobs. "Open" means you can download and run the model yourself instead of sending every request to one provider. The company also released its technical report, training environments, and training code, according to the MiMo-V2.6 announcement.

The useful change is choice. Xiaomi says both models can process text, images, and other media in one system, while the hosted interface and API, the connection other software uses to send requests, are available now. It kept API pricing at the MiMo-V2.5 level, although the announcement does not print the actual rates. Do not move spend on that claim alone. Check the checkout screen and measure a real batch.

Use Flash as a challenger for repetitive jobs such as classifying inquiries, extracting details from campaign screenshots, or turning notes into a fixed brief. Keep Pro for a smaller test of tasks that need several steps. This is a candidate for teams already paying meaningful model bills. A business running a handful of prompts each week can ignore it.

The AI content approval workflow gives you a ready-made human check before any challenger model reaches customers.

xAI ships Grok 4.7 without raising token prices

Grok 4.7 arrived on September 21 at the same price and speed as Grok 4.6. xAI lists the standard model at $2 per million input tokens, the words and instructions you send, and $6 per million output tokens, the text it returns. A faster version doubles both output speed and price. The model is available through the Grok API, model routers, and other connected platforms, according to xAI's release note.

That makes this an upgrade test, not a new subscription decision. If Grok 4.6 already drafts reports, analyzes research, or handles long office tasks in your stack, create a duplicate route and change only the model. Do not rewrite the instructions at the same time. You need to know whether the model improved, not whether a new prompt rescued it.

xAI reports better scores than Grok 4.6 across its published tests, but vendor test scores are not your acceptance test. Use the same 20 saved jobs and the same scoring sheet as the MiMo trial. Compare complete cost per accepted result, not the price of one token. A cheaper answer that needs a person to rebuild it is the expensive answer.

If you are choosing a first model rather than replacing Grok, start with the routing logic in the September 4 model dispatch. It separates routine work from the small number of jobs that justify a stronger model.

Claude's 80-minute outage exposes single-model workflows

Anthropic recorded elevated errors across three Claude model families from 00:50 to 02:10 UTC on September 22. The incident affected Claude's website, API, coding tool, and workplace tool before Anthropic marked it resolved, according to the official status record.

Eighty minutes is enough to lose an evening's inquiries if Claude sits between a form submission and the first reply. The fix is not an automatic jump to another model for every failure. A backup model can produce different wording or make a different decision, which is dangerous when the workflow quotes prices, approves refunds, or sends messages.

Open the automation that calls Claude. Add an error route, the path used when that step fails. For customer-facing work, send the original input to a review queue and alert the owner by email or Slack. For low-risk internal summaries, point the error route to a tested backup model, then label its output fallback_model so you can audit it later. The guarded pattern in the n8n AI agent setup keeps sending under human control.

On the bench

• n8n published a prompt-testing framework on September 21. Its Evaluation Trigger can run a saved table of examples through a workflow, which is a practical home for the 20-job test above.
• Apple now documents how administrators can restrict individual Apple Intelligence features on a Mac. Review those controls before staff paste customer records into desktop writing tools.
• MiMo-V2.6-Pro-UltraSpeed promises up to 20 times faster output at the same quality. Wait for posted rates and your own timed batch before paying for speed.

Keep the existing model live, test the challengers beside it, and wire the failure route first. Then use the automation build library for the next system you put behind those safeguards. Get the next build.