Short answer for the Anthropic 5.5 models: let Opus 5.5 plan, let Sonnet 5.5 do the hands-on work, keep Medium as your default effort, and skip Max / xHigh unless a task really needs it. Here is why.
Anthropic launched Opus 5.5 one week ago, and just two days ago Sonnet 5.5 was also released.
As with other model releases, some habits might need fine-tuning to make the most of the models, and especially to avoid running out of credits.
✅ DO keep your pipeline split: Opus plans, Sonnet acts
For coding pipelines, Ive been using Opus to orchestrate and plan, with Sonnet doing the hands-on work. With the cache read price decrease of Opus 5.5 this might seem unnecessary, but I am keeping this setup.
For a lot of tasks, Sonnet 5 low was more than enough. I am now substituting Low → Medium and still the tasks seem to cost 2-4x less for around a 2x improvement on the score with Sonnet 5.5. I am also betting on the price and speed difference to prefer it over Opus 5.5 on this pipeline setup (from benchmarks, around half the price and 30% faster)
✅ DO let Opus 5.5 run longer
With Opus 5, I got used to clearing and starting over whenever a session got long. With Opus 5.5, I have loosen up that rule.
Opus 5.5 60% decrease in cache reads (most of the cost I have had in past sessions) and 20% decrease in cache writes compared to Opus 5 allows for this behavior change. [source for Anthropic models prices and comparison]
Reports also show that Opus 5.5 stays coherent much longer, its summaries are easier to pick back up after a break, and it is possible to can change effort mid-session without losing the cache.
Two habits I kept:
– a state file (plan, decisions, open tasks), so nothing important gets lost when context is cleared
– clearing sessions when context > 250k tokens
✅ DO set Opus 5.5 Medium as your default model
My go to for individual ad-hoc tasks was Sonnet. Many times it was not a one-shot effort, but the lower cost compared to Opus would get the work done.
Now? Again due to the 60% price reduction in cache reads Opus 5.5 is my go-to model – smarter, more concise, and I catch myself having sessions with 20 million tokens without worrying about the 5-hour usage limits.
❌ DONT let sub-agents run wild in long sessions
I was caught off guard in the past. Even if prices are reduced and models are smarter, spawning multiple sub-agent still counts against the same usage limits, and over a long session they add up quickly. On top of that, Opus 5.5 is now the Claude Code default, so any sub-agent that inherits the model runs on Opus unless you tell it otherwise.
❌ DONT go for the Max / xHigh effort
For both models, it’s rarely worth it. Opus 5.5 at default effort already beats Opus 5 at max for about a fifth of the cost, and Sonnet 5.5 at medium beats Sonnet 5 best for a tenth. Max can even backfire: Sonnet 5.5 scored lower at Max than at xHigh on one coding benchmark, because it started spinning up extra sub-agents.
What changes have you made to your workflow with the new 5.5 family models?
How are you splitting work between the two?
