Opus or Fable for my AI assistant? I analysed my own sessions.
I found much the same pushback under both models when I checked my own saved conversations.
I run one AI assistant through a single conversation that coordinates the work. The assistant hands jobs to roles such as writers and builders on more than one machine, with a queue for longer jobs.
I moved to Opus mostly to slow down how fast my usage allowance was disappearing. Opus got the long and overnight runs. I did not assign the models at random, so this was never a fair test.
I wanted to understand my own choices. The saved conversations could tell me how often I pushed back, but they could not support a ruling on the models.
Much the same pushback
I counted pushback whenever I made a correction, gave a reminder, called out a broken rule or repeated a request. Fixes that still failed and annoyance counted too. Problems with my own tools were left out.
AI workers marked every message without knowing which model was running. A second set checked the flagged messages and a random sample of the rest. That sample let me estimate what the first pass missed.
I compared the three weeks when both models ran. Later weeks would have added work only Fable had received.
My estimated pushback per 100 of my messages was 17.2 under Fable 5.1 against 17.3 under Opus 5.5. Those estimates include what the first pass missed, and the ranges are wide.
About five in six of my messages carried no pushback at all. I could not measure a difference in failed retries, rule breaches or forgetting either.
The one lean towards Fable was on more serious pushback, meaning a clear error that cost another attempt or worse. That difference could still have been down to chance. I would not pick a model on that difference alone.
The count only captures what I bothered to say. I said less during overnight runs.
What the lower price meant for me
I compared costs by pricing the usage at the published per-token prices. I pay a subscription, so my concern is how fast the allowance goes. I cannot see whether the allowance counts the models in the same ratio as those prices.
On the price list, read 17 November 2026, Opus costs 40% of Fable 5.1 for input, output and one-hour cache writes. For reading cached text, the figure is 80%. Those reads made up the largest share of Opus's cost at those prices in my data.
Repricing the same Fable work at Opus's rates made it about 40% cheaper. Across all my data, though, Opus came out about a fifth cheaper per message I typed. Opus took more steps for each ask, which reduced the saving.
Fable ran at high effort and Opus at extra-high. Opus also got the longer runs, so I cannot separate model, effort and work when explaining the extra steps. I never varied either model's effort setting, so I learnt nothing about which effort setting works best.
The bigger change came later
After I reorganised how the assistant works in early November, Fable 5.1 drew measurably more pushback. My estimate rose from about 17 in every 100 of my messages to about 24 in every 100. The change between periods was bigger than the difference between models.
My milder pushback barely changed as the more serious pushback rose. In the same reorganisation I switched off Claude Code's built-in auto memory and used my own memory system instead. My assistant forgetting prior agreements became the largest single cause of my more serious pushback.
Switching auto memory off is my leading suspect, not a proven cause. Other causes explained most of my more serious pushback. The work changed between those weeks too. I am deciding whether to improve my own memory system further or switch auto memory back on.
The newest Fable arrived after the reorganisation, with too few messages to judge.
The comparison needed checking too
The first unchecked read favoured Fable on overall pushback. Checking the marks and estimating the misses closed that gap. An early summary from my assistant also told me Fable drew more serious pushback. The checked comparison said the opposite within the weeks both models ran. The early summary had mixed in the later weeks.
My model comparison needed a second opinion of its own.
My assistant ran the study on Fable, with AI workers on Opus marking and checking messages without seeing the model names. Opus was also a contestant, so hiding the names mattered. Both the labellers and the checkers were AI, not people.
A few messages gave away which model was running despite the hidden names. Removing those messages did not change the result.
How I counted
The detail is here for anyone who wants it. I read 1,520 messages I had typed across 59 saved conversations, from mid-October to mid-November. The marking involved 11 AI workers. Three checkers reviewed 256 flagged messages and 180 messages sampled from those without a flag.
The comparison covered 462 messages under Fable and 536 under Opus during the weeks both ran. The miss estimate rests on six misses in a sample of 70 Fable messages and three in 70 Opus messages.
What I decided
I am staying on Fable at the high setting, though the newest Fable's sample is too small to judge. Opus remains the fallback when usage gets tight. My confidence in that choice is low. A fair test would take about 18 working days on each model. I am not running that test unless usage limits start to bite.
During the run, Claude Code's default cleanup deleted the local transcripts of seven sessions before they were read. Claude Code keeps its own local transcripts for 30 days by default, adjustable with cleanupPeriodDays.
If you use Claude Code, its local transcripts are on your disk in ~/.claude/projects/ (documentation read 17 November 2026). My local transcripts gave me enough to reconsider my own model choices. Next time I would change one setting at a time and keep the transcripts longer, so I still have them to compare.
Barry