Claude was the main coding agent in my workflow for about a year and a half. For the first time in over a year, an Anthropic model isn't in my daily rotation.
I care about that beyond my own setup because of how the work is orchestrated. We review a plan before anything builds. Then a product harness — for me that's Cursor's cloud agents — carries the plan out, with checks that the code still matches what we approved. Once that loop exists, you stop needing the flagship model on every turn. The expensive judgement sits in the plan. The build model is a worker inside Cursor's harness, running against a decision we already made. That is why I can hand Composer 2.5 a job and walk away. I don't need Opus to carry out a plan we already agreed on in a single shot.
Anthropic still has a strong harness in Claude Code, and a top-tier model. What got harder was depending on them for a full working week: reliability in agent loops, how much reasoning Sonnet burns, and how expensive and unpredictable Fable access has become.
What I liked about Claude for so long
I stuck with Anthropic because Claude Code felt built for agent work, and Opus was the model I reached for when the problem was hard. I was happy there. The switch was about what I could depend on week to week, not a change of taste.
Opus 4.8 and Sonnet 5
Opus 4.8 is still capable. In day-to-day agent use I kept hitting reliability issues: tool calls that wouldn't parse, sessions that stalled, the "retry also failed" loop where nothing recovers. I see the same reports on reddit and X. When most of my day is an agent calling tools, that is enough to knock a model out of rotation even if it still looks fine on a bench.
Sonnet 5 is smart, and in a long coding loop it also thinks for a long time. Higher effort means more reasoning tokens, more waiting, more cost. For a lot of the work I do I don't need that depth on every turn, especially once the hard thinking already happened in the plan. Composer 2.5 feels faster in that same kind of loop while being cheaper. Once models cleared a basic quality bar for me, token output speed became something I actually valued. That wasn't true a year ago, when getting the model smart enough was still the whole fight.
Fable access
I still want Fable for hard problems. Access has been hard to plan around. In the subscription, then capped at part of your weekly limits, then extended at the last minute, then pointed at credits that burn quickly. As of mid-July the free-included window had been pushed again to the 19th, after earlier deadlines. Operating Fable looks too costly to leave wide open for everyone to burn all day, so they limit it. Then, under pressure from GPT-5.6 Sol, they keep extending included access anyway.
That kind of access makes a model hard to treat as something you build a week of work on. Their public communication around this hasn't been great either.
Good enough
For a long time the game was latest and greatest. Whoever had the smartest model that week won the conversation. That has cooled off for me, and for the people I work with who treat frontier coding models as close enough for daily work now.
Close enough only works if something else is holding the work together. Drop a cheaper model into a bare chat window and you just get worse chat. Put it inside a product harness that can run a loop, and put that harness behind a plan you already reviewed, and the same model is suddenly useful: it builds against something you trust, and you catch it when it wanders. Peak IQ still matters for the planning step. It matters less for the build once that plan exists.
My stack change is one example of that, not a survey. Asking which model is best and asking which model I can run all day inside that setup are different questions.
OpenAI's package
OpenAI's harness and top model still trail Anthropic's in my experience. Sol is strong, but I wouldn't say it's clearly ahead of Fable. I no longer need the absolute best model to get great results from a build. And for how I work, OpenAI's overall package is simply better: more generous tiers, banked resets, currently no five-hour cliff of the kind I used to hit elsewhere, open to more harnesses, faster in the loop, and communication that has been easier to plan around than Fable's on/off/included dance.
Just better, for a week of real work.
How I work now
Most days I write a detailed implementation plan, get it reviewed, and hand the build to a Cursor cloud agent on Composer 2.5 Standard (not Fast). I wrote about that week here: I ran Cursor's cloud agents for a week.
Fast is what you want when you're watching the IDE. Standard is the same model without paying the interactive premium, which matters when the agent is running in the background while I'm in meetings, or overnight. Cursor's published averages put Standard around $0.07 per coding task and Fast around $0.44, with Opus-class frontier work more like $4–$5 in those same numbers — a small sample and task-level pricing, so treat it carefully, though it lines up with what I see on my bill.
I can plan in the morning, kick off builds, and come back to PRs later. The judgement sits in the plan and in the checks around the agent. Cursor's harness runs the loop. Composer only has to be good enough to carry the plan out while I'm doing something else. That setup is doing more of the work than any particular model launch.
Turning point
When several vendors are good enough for daily coding, the package you can schedule a week on beats the model you babysit. For me that package is Cursor's harness plus the way we orchestrate work into it: plan first, review the plan, let a good-enough model build, then verify against what we approved.
I would stop organising my defaults around which model won the last launch. That orchestration is why cloud agents on Composer are usable for my real work, and why leaving Anthropic's daily rotation doesn't feel like giving up quality.