Picking the right model for the job

10 August to 16 August 2026

Three model events landed in a single week: Qwen3.8-Max open weights finally shipped (after a multi-week wait), DeepSeek took V4-Pro out of preview and published peak/off-peak API pricing, and Google launched Gemini 3.7 Flash as a cheap closed workhorse. Each alone is a minor release note. Together they change the question operators ask.

Quick terms, so the rest is clear:

• Open weights means the model’s trained files are downloadable. You can run it yourself rather than being locked to one company’s API. It is not a full open-source build, but it gives portability.
• Licence controls what you may do with the downloaded model. MIT is a permissive boilerplate: inspect it, modify it, run it commercially. A custom licence, like Qwen’s, means terms written for that model.
• Self-hosting means running the model on infrastructure you control instead of paying per API call. It removes per-use fees, but pushes the cost to hardware, power and upkeep.
• Mixture of experts (MoE) activates only a slice of the model per token, which is why a 2.4-trillion-parameter model is not that expensive to run.

It used to be: “which is the smartest model?” Now it is: “how do I route work across a menu where cost, licence and schedule differ?”

Choosing the model was never the hard part

Qwen3.8-Max is a 2.4-trillion-parameter model with a mixture-of-experts (MoE) architecture, activating only a slice per token. Self-hosting it still takes multiple machines, so it is not something you spin up on a spare box. DeepSeek V4-Pro now has a real MIT path and agent-tuned capabilities, which is more useful for most teams. Gemini 3.7 Flash undercuts on price. No single model won this week. The menu just got deeper.

Peak and off-peak pricing is the new curveball

DeepSeek’s published rates are the bigger story because they introduce something model APIs have not really done before: time-of-use pricing. From 16 August, V4 Pro and Flash cost roughly twice as much during peak hours as off-peak, so load becomes schedulable like electricity.

Trezor: the second wallet trust break in two weeks

The hardware-wallet story picked up again. A shipping-provider breach exposed around 13,700 Trezor customers’ personal data. A shipping provider is the third-party company that packs and delivers an order, not the wallet maker. So this was at ShipMonk, Trezor’s fulfillment partner, and what leaked was real: full names, shipping addresses, phone numbers and email addresses.

The damage is not that someone can take your coins. It is that an attacker now has your name, address and phone tied to the fact that you hold a hardware wallet, which makes phishing much more convincing. And phishing is how wallets get drained.

It is a different fault from the Coldcard seed-generation flaw we covered last brief, and the second one in two weeks.

These need different fixes, but they fail the same way. The problem is never the wallet itself, it is the edges around it, the firmware, the shipping, the company that holds your address.

A small correction in my own stack

This week the routing point hit home in my own setup. I did not realise until I took a look at my OpenRouter spending, which was higher than expected. Hermes agent had been pinned on an older DeepSeek snapshot at $0.084 per million input tokens, while the same model family’s `-latest` alias was running $0.06, about 28% cheaper and just as good. The fix was not cleverer AI. It was a pricing alert that runs every morning, so I see what each model costs before I lean on it, and a habit of checking the price page when rates change. It is unglamorous work, and it is where the savings actually live.

One question I am carrying into next week: models keep getting cheaper and the menu keeps growing, which means the model itself is becoming a commodity. Stripe just underlined that in the biggest way: it spent over $7 billion buying OpenRouter, and OpenRouter does not make models at all. It runs the gateway that lets you choose between hundreds of them, by task and budget. The market put its money on the plumbing that lets you switch between models, not on any single model.

It changed how I look at models. I used to ask which one was the smartest. Now I ask what the job actually needs, whether that is price, capability, or value for money, and pick from there. The cheapest model is usually not the right one, and the most expensive is rarely necessary.

Sources

• Qwen3.8-Max open checkpoint on Hugging Face
• DeepSeek Models & Pricing
• Google, “Introducing Gemini 3.7 Flash”
• Coinkite, Current COLDCARD Security Status
• Trezor, customer data exposed in shipping-provider incident
• TechCrunch, “Stripe will reportedly acquire AI gateway startup OpenRouter for $7B”