Here’s what’s shaping AI right now - without the noise:

OpenAI released the full GPT-5.6 family this week and paired it with something that signals a bigger strategic shift: ChatGPT Work, a direct answer to Anthropic's Claude Cowork, and a redesigned desktop app that merges Codex into the main ChatGPT experience.

The model lineup:

The platform moves are as significant as the model. ChatGPT Work brings Codex's engine to everyday tasks in a more accessible format. The revamped desktop app now includes a built-in browser and computer control alongside the new Work agent. Sam Altman noted that every enterprise is now actively thinking about AI spend and the pricing is clearly designed with that conversation in mind.

Read the full GPT-5.6 announcement here.

The bigger signal: 👉 OpenAI is no longer just a model company. With Work, a desktop agent, and browser control in one app, the superapp strategy has moved from talk to product.

If you’re a decision maker at your company, you need to be on the bleeding edge of, well, everything. But before you go signing up for seminars, conferences, lunch ‘n learns, and all that jazz, just know there’s a far better (and simpler) way: Subscribing to The Deep View.

This daily newsletter condenses everything you need to know about the latest and greatest AI developments into a 5-minute read. Squeeze it into your morning coffee break and before you know it, you’ll be an expert too.

Subscribe right here. It’s totally free, wildly informative, and trusted by 600,000+ readers at Google, Meta, Microsoft, and beyond.

Most teams that hit unexpected Claude bills are not using the wrong model. They are using the right model the wrong way. The architecture around your prompts matters as much as the prompts themselves.

Here are the techniques that make the biggest practical difference:

Stop sending entire conversation histories

After 100 turns of conversation, you could be sending 500,000 tokens with every single request. Instead, store a rolling summary of the conversation and attach only the last five messages. Typical reduction: 80%.

Stop sending entire documents

Pasting a full PDF into a prompt is one of the most expensive habits in AI workflows. Use retrieval-augmented generation instead - chunk the document, index it, and send only the relevant sections. A 100,000-token prompt can become 2,000 tokens without losing meaningful information.

Set explicit output limits

Asking Claude to "explain everything about X" can produce 4,000+ token responses. Asking for ten bullet points or capping the response at 500 words cuts output tokens immediately and usually produces more useful answers.

Cache your system prompts

If you are sending the same large system prompt with every request, you are paying for it every time. Keep reusable prompt components separate and use caching layers to avoid resending static context repeatedly.

Route requests to the right model

Not every task needs Claude Opus or Fable. A simple FAQ query does not need the same model as a complex multi-step reasoning task. Building a basic routing layer that sends simple requests to cheaper models can cut costs by 50% or more without affecting output quality for those tasks.

Compress context before sending

Run a lightweight compression step before sending large context windows. Taking 40,000 tokens down to 4,000 through summarization before Claude sees the request is often possible without meaningful information loss.

Use structured outputs

Asking Claude to respond in a defined JSON format produces shorter, cheaper, and more usable outputs than asking for open-ended explanations. Structured responses also make downstream processing significantly easier.

The honest benchmark:
Applied together, these techniques typically reduce token costs by 70 to 95%. The teams with the lowest AI bills are not using smaller models. They are sending less, retrieving smarter, and measuring everything.

Before every API call, ask: does Claude actually need to see all of this? If the answer is no, it probably does not.

Previous AI voice experiences had a consistent problem. They felt assembled rather than natural - separate models stitched together, awkward pauses, turn-based interactions that broke the conversational flow.

GPT-Live rebuilds the architecture from the ground up.

What actually changed:

In OpenAI's own testing, users preferred GPT-Live over the previous Advanced Voice Mode in roughly three out of four head-to-head comparisons. Expert science scores nearly doubled. Watch it in action here.

Sam Altman has talked for years about building voice AI that feels like the fictional assistant from the film Her. Previous versions always had friction that broke the illusion. Live removes most of it.

The bigger signal: 👉 Full-duplex voice changes what AI interaction actually feels like. For users who prefer speaking over typing, this is the upgrade that makes voice a genuine default rather than a novelty.

A few tools quietly worth exploring:

See you next issue.