Claude Got Cheaper. Your Clients Still Cannot Find You in ChatGPT.

This week Anthropic shipped Claude 5.1 with workloads that are about 25% cheaper.

Google is lining up Gemini 3.8 Flash to undercut it.

OpenAI gated Astra because the model crossed a Critical cyber line.

If you run an agency, none of those headlines is the brief.

The brief is this: intelligence keeps getting cheaper, and your clients now ask a new question. Do we show up when someone asks ChatGPT?

What actually shipped

Three signals in 48 hours.

1. Claude 5.1, Fable for most teams and Mythos behind a gate, cut workload cost and cache-read cost. CursorBench sat at 73.4%. That is a production model, not a demo.

2. Gemini 3.8 Flash is being positioned as the fast counter to Fable. Same story as every cycle. Volume models get cheaper. Agencies get more drafts for the same invoice.

3. OpenAI Astra is gated at launch. Rogue-agent incidents are no longer theory. If you are wiring agents into client work, security is now part of the SOW.

We will swap models. We always do. That is not strategy.

The marketing problem hiding under the model news

Here is the part most agencies will skip.

Clients are no longer buying we use AI. They are buying proof they appear inside the answer.

AgencyAnalytics 2026 benchmarks put helping clients show up in AI search as the top new request for 66% of agencies.

Semrush looked at 50,000 brands in ChatGPT. Only 15% of topics have a clear owner.

Forrester says 94% of B2B buyers already use generative AI in the purchase process.

So the cheap model is useful. It lets you ship 30 citation-ready pages this week instead of three.

It does not put your brand into Gemini, Perplexity, or ChatGPT by itself.

Generative engine optimization is not a new content calendar. It is entity clarity, third-party consistency, and pages an answer engine can quote without inventing your offer.

We have watched thin sites get cited because a Crunchbase line was cleaner than the homepage. That is the game now.

Do this before you change models again

Pick one offer you actually sell.

Write the page so a model can extract four things: who it is for, what it costs in range, what the outcome is, and what proof exists.

Then check the same facts on LinkedIn, Crunchbase, and your case-study titles. If those three disagree, the engine will pick the wrong one.

That is the work this week. Not another model bake-off.

The labs will keep cutting prices. Your client will keep asking if ChatGPT knows their name.

Answer that first.