Back to Claude
Ben's Bites ● Covered by 4 sources
Anthropic’s Claude Opus 5.5 is out, and it’s now the go-to agent for some users. OpenAI also cut prices on GPT-6 Luna and Sol, but they still trail Opus.
Based on reporting by Ben's Bites — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic has a new flagship out in the wild: Claude Opus 5.5. The pitch is simple enough. It’s beating Fable 5.1 on benchmarks, costs less than Opus 5, and, in practice, it’s handling tasks well enough that some users are switching their default agent over to Claude Code.
There’s a practical catch, and Anthropic seems to have softened it a bit. The five-hour limits in Claude Code are up by 20%, and there’s a banked rate-limit reset now. Inside the app, Opus 5.5 defaults to medium thinking, though higher settings are still available. One user said they ran it at high and xhigh for two days and only hit the five-hour limit once.
The same roundup says the model was put to work on real side projects: updating filtering on ossean.com, rewriting essays on shortandlongtail.com, and adding the new models to detroitbench. The verdict from that testing is blunt: it’s pulling people back from Codex and making Claude Code the preferred agent again.
OpenAI, meanwhile, has GPT-6 Luna and GPT-6 Sol, but they land more as price cuts than breakthroughs. Both are only in Codex and ChatGPT Work for now, not regular chat, and both are 50% cheaper than before. The listed prices are $0.10/$0.50 for Luna and $2/$10 for Sol, per million input/output tokens. The catch is that the intelligence jump is described as minor, and Sol is said to sit clearly behind Opus 5.5.
One test of Sol pushed it into a long task: building a Duolingo-like course for marketing. It stopped after 1.5 hours with a decent version, then kept going, and after about 4 hours it had used 22% of a weekly limit on a $100 Codex plan. That’s not exactly a model built for thrift. GPT-6-Astra burns through limits even faster, according to the same report.
My take — AI-written commentary, not fact-checked reporting
The pattern here is getting obvious: the market keeps rewarding the model that feels reliable, not the one with the loudest naming scheme. Opus 5.5 sounds like the boring choice, which is usually the dangerous one for everyone else. Fancy AI assistants can keep wearing little pendant-shaped costumes; people still want the thing that finishes the job.
Read more about this at: Ben's Bites
