Anthropic Is Hitting a Wall

Anthropic had the developer story every AI lab craved. Claude Code worked. Opus was expensive but worth the splurge. Sonnet was known as the workhorse. For a time, the company felt like the lab that knew builders better than OpenAI did.

That is precisely why the last few weeks are so ugly.

On June 1, Anthropic confidentially filed a draft S-1. It said on May 6 it was increasing Claude usage limits, including doubling five-hour limits for Claude Code for Pro, Max, Team and seat-based Enterprise plans. Then the product experience began to tell a different story.

I know the trouble from my own Claude subscription logs. Visible Opus input and output in the current seven-day rolling window was 1,442,423 tokens. Two other older heavy weeks on the same machine were 8,927,831 and 8,477,395. That is 16.2% and 17.0% of the old weeks.

Even if I put cache creation in and only exclude cache reads, the current non-cached total was 4,563,761 versus old weeks of 22,155,800 and 23,077,312. That is 20.6% and 19.8%.

That is not universal evidence of an underlying rule. That's what one power user has observed. But if a subscription used to have way more visible Opus work, and now burns down fast, while the public message was "higher limits", the trust problem is obvious.

Then the Fable / Mythos mess came along.

On June 12, Anthropic said the US government ordered the company to halt access to Fable 5 and Mythos 5 for foreign nationals, effectively shutting down both models for all customers while Anthropic complied. Many users had complained that Fable's safeguards were overly broad, Anthropic said in the same statement, and admitted that perfect jailbreak resistance is not currently possible. Anthropic said Fable requires 30-day retention of customer data so Anthropic can research and mitigate jailbreaks.

This is a brutal product package: the flagship model is deemed too sensitive for normal access, the safety layer annoys legitimate users, the safety story still doesn't satisfy the government, and customers get a retention tradeoff they didn't sign up for because the model needs monitoring.

Now Sonnet 5 is meant to be the answer.

On June 30, Anthropic released Claude Sonnet 5 and made it the default model for Free and Pro plans, including for Claude Code. Sonnet is cheaper than Opus on sticker price: $2/M input and $10/M output through August 31, then $3/M and $15/M. Opus 4.8 costs $5/M and $25/M.

But the sticker price isn't the whole story. Anthropic says Sonnet 5 can match Opus 4.8 on some tasks, at higher effort. That same post also states that the Sonnet models are significantly worse at cyber than Opus 4.8 and Mythos 5, and the change in the tokenizer means that the same input can map to 1.0-1.35x tokens.

When looking at Anthropic's own cost-performance charts, you'll see Opus looking better or cheaper for hard tasks. Sonnet 5 is only interesting for low-effort tasks, where it competes with cheaper Chinese Open Source models.

That's the daily coding reality. Developers don't care if it is easier for Anthropic to serve the default model. They care about closing the ticket.

Anthropic is now squeezed from both sides. If you want cheap long-horizon work, open-weight models are getting better fast. Z.ai is putting GLM-5.2 right in that lane. OpenAI has GPT-5.5 in Codex and ChatGPT for frontier coding agents, and GPT-5.6 is in preview for API and Codex with wider availability planned.

That puts Anthropic in a rough middle. The best models are gated, costly, or covered in safety/compliance drama. The subscription is tighter. The default model is cheaper, but only if you don't care about effort, tokenization and task completion.

The public company story requires enterprise confidence. The developer story is goodwill. It is easy to spend goodwill.

Perhaps enterprise customers still buy the Anthropic story. Perhaps regulated buyers like the safety stance. Perhaps this doesn't kill the IPO story.

But developers decide which tools actually get pulled from below into companies. If Anthropic keeps hiding the models people want, making subscriptions less useful, and defaulting users to a model that doesn't feel cheap enough or powerful enough, it will lose what made Claude valuable in the first place.

Cheap work is for open models. Frontier coding agents are for Codex. Now Anthropic must explain why ordinary developers should remain in the middle and pay for the squeeze.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com