Anthropic's flagship model can't find a market
Issue #021 · 4 min read
Anthropic's flagship model can't find a market
Claude's best model struggles for users, GitHub Copilot moves into Slack and Teams, and GitLab lets an agent write its own automations.
The big story
Anthropic built the model everyone said was the smartest one out there. It's the cheaper, worse models that people are actually paying for, according to a report covered by Simon Willison.
That's not a small thing.
For two years the industry story has been simple: build the best model, and demand follows. Anthropic did the first part.
The second part isn't happening.
Buyers are routing around the flagship and picking whatever is fast and cheap enough to bolt into a workflow. If that holds, it says something uncomfortable about the whole pricing model the frontier labs have been selling to enterprises — that "best" is a feature people will pay a premium for.
For a PM, this is the sharper version of a question you've probably already had internally: are you buying the smartest model, or the one your users will actually keep open in a second tab? Those are turning out to be different products, and only one of them is winning.
What shipped
Linear rebuilt its coding agent pricing and gave it a body. The changelog says Linear Agent can now set up, run, and test code before handing it back, "fewer handoffs and changes that are further along when they come back to you." Pricing is also getting simpler — the post frames it as making AI credit usage "easier to understand and control." If your team has been fighting with opaque credit meters on agent tools, this is worth a look before your next renewal conversation.
GitHub Copilot is showing up wherever your team already argues about work. Two separate releases put it in Slack and in Microsoft Teams, both in public preview. The Teams version turns a channel discussion into an agent session "everyone can see and help direct" — mention @GitHub and the whole thread becomes a shared workspace instead of one engineer's side quest. Worth testing with a real cross-functional thread, not a demo channel, to see if visibility actually changes who weighs in.
GitLab is letting non-engineers write their own automations. The Flow Creator agent, shipping in GitLab 19.3, removes the requirement to learn the Flow Registry's YAML schema before you can automate a multi-step process — you describe the flow in plain language instead. GitLab's own framing is blunt about why this matters: "the people who understand a workflow best aren't always the same ones who know a YAML schema." That's a real gap in most tooling, not just GitLab's, and it's the kind of feature that quietly changes who's allowed to build process automation at your company.
What I'd actually do this week
- If you're evaluating or renewing a frontier model contract, ask your vendor rep directly what usage looks like relative to their cheaper tiers. The Anthropic story suggests "most capable" is not the same sales pitch it was a year ago.
- Pull up Linear's new pricing page and actually model your team's current agent usage against it before your next billing cycle closes.
- If a non-technical teammate has ever asked you to build them a workflow automation, point them at GitLab's Flow Creator and see how far they get without you.
Hit reply if you've tried any of these — I read every one.
Tools mentioned
- Linear Agent coding sessions
- GitHub Copilot in Slack
- GitHub Copilot in Microsoft Teams
- GitLab Flow Creator agent