What is going wrong with Claude since Sonnet 5?
What is going wrong with Claude since Sonnet 5?
The flagship kept moving. The model most people actually call did not.
On June 30, Anthropic made Claude Sonnet 5 the default for Free and Pro and priced it at $2 / $10 per million input and output tokens. The API id is still claude-sonnet-5. Anthropic's model overview lists it as the current Sonnet, with a knowledge cutoff of January 2026.
In the twelve weeks after that launch, Moonshot shipped Kimi K3, SpaceXAI shipped Grok 4.6 and then Grok 4.7, and OpenAI shipped GPT-6 Astra. Google shipped three more Flash models. Anthropic's September point release was Fable 5.1, at $10 / $50, not a new Sonnet.
That is what is going wrong. Claude did not stop. Sonnet did.
Falling behind is a tier, not a eulogy
Fable 5.1 still leads several independent indexes. Artificial Analysis has scored it 66 on its Intelligence Index and 70 on its Coding Agent Index, ahead of GPT-6 Astra at 61 and 67. SpaceXAI's own Grok 4.7 table still gives Fable 5.1 the high score on CursorBench 4.0 (51.8%), Terminal-Bench 4.0 (57.9%), and GDPval (1,735).
Sonnet 5 was sold as the model that could touch last generation's Opus at a Sonnet price. The class around that price moved. Grok 4.6 and Grok 4.7 are $2 / $6. Gemini 3.8 Flash is $0.75 / $3.75 through December 31. Kimi K3 is an open 2.8-trillion-parameter model. Astra is a flagship, $10 / $50, in Fable's price band, and it still shows up on the same coding charts.
If your review stack calls Sonnet because it was the sensible default in July, the default is now old.
The number that aged
Google's Gemini 3.8 Flash model card puts Claude Sonnet 5 at 53.8% on DeepSWE v1.1. On the same row, Gemini 3.8 Flash is 73.7%, Claude Opus 5 is 74.0%, and GPT-5.6 Sol is 72.7%. SpaceXAI's Grok 4.7 launch quotes 71.0% at high effort, against 65.2% for Grok 4.6.
Those are vendor tables, and harnesses differ. They still point the same way. The long-horizon coding bench that everyone started quoting after June is a place where Sonnet 5's published score sits with the previous generation, and the models shipped in July, August, and September sit together near Opus.
On CursorBench 4.0, SpaceXAI's cost chart includes Sonnet 5 on a lower curve than Grok 4.7. Grok 4.7's published score is 46.3%, against 40.4% for Grok 4.6 and 41.7% for GPT-5.6 Sol. Fable 5.1 is ahead at 51.8%, at a much higher cost per task. Sonnet's job was to be the cheap curve. Grok took that plot.
A freeze is a product decision
Sonnet 5's own launch was a real jump on Sonnet 4.6. Anthropic showed it matching Opus 4.8 on some effort settings, with a 1M context and adaptive thinking on by default. Then the id stopped changing.
Meanwhile the workhorse vendors treated the summer the way Google treated Flash: ship, read the agent failures, ship again. Grok 4.7 is a larger base than 4.6, trained longer on tasks that take many hours, and it checks its own work more carefully. Gemini 3.8 is explicitly based on 3.7, which was based on 3.6. Kimi K3 showed up in July as an open model aimed at the same long engineering sessions. None of them waited for Sonnet to blink.
Anthropic spent the September news cycle on Fable 5.1. That is a coherent lab strategy if the flagship is the product. It is a problem if Sonnet is the model inside Claude Free, Claude Pro, and a large share of API traffic. The people writing pull requests are not all on a $50 output token.
Output price is the quiet half of the slide. Sonnet's permanent rate is $10 per million output tokens. Grok's is $6. A review that thinks longer, which is what these agent benches reward, costs more on Sonnet before it scores higher. The models that passed it got both a newer checkpoint and a cheaper token.
What to do on the next review
Keep Sonnet 5 when you want Anthropic's current mid-tier and you already trust its tone on your codebase. Stop treating the June default as the September ceiling.
- Hard, long diffs — try Grok 4.7 or Kimi K3 in ThinkReview and compare the findings on one real merge request.
- The flagship question — that is Fable 5.1 versus Astra, not Sonnet versus anyone. Different budget.
- The patch people are waiting for — the next piece is what Sonnet 5.1 would have to beat.
The model-by-model pass is in the models that passed Sonnet 5.
Scores cited from Anthropic's Sonnet 5 launch, the Gemini 3.8 model card, and SpaceXAI's Grok 4.7 announcement. Run the comparison on a real PR with ThinkReview.