Sonnet 5.5 beat Opus on a coding test: which Claude model to use for every task

Anthropic released Claude Sonnet 5.5 on 28 September 2026, and one result surprised me. On a coding test called Terminal-Bench 4.0, the cheaper Sonnet beat Opus. That doesn't mean Opus is done. So here's what the numbers actually say, and my simple cheat sheet for which Claude model to use for which job.

01What Anthropic announced

Terminal-Bench 4.0: Sonnet 5.5 scored 70.6%. Opus 5.5 scored 66.4% (at its xhigh effort setting). The older Sonnet 5 scored 10.3%.
Anthropic says it's the first Sonnet model to beat Pokémon Red using only screenshots.
It writes output over 30% faster than Sonnet 5, and Anthropic calls it the fastest Sonnet so far.
In Anthropic's testing it costs up to 30% less per task than Sonnet 5.
The price per token didn't change: $2 per million input tokens and $10 per million output tokens. That's half of Opus 5.5.

02The catch: Opus still wins most tests

Anthropic's own comparison table has eight benchmarks. Opus 5.5 leads on seven of them. Terminal-Bench 4.0 is the only one Sonnet wins, and some of Opus's leads are small.

So Sonnet is not a better Opus. It's a fast, cheaper model that's now very strong at terminal-style coding. Opus is still the one for complex work that needs careful judgment.

03My cheat sheet: which Claude for every task

Haiku 5.5: quick jobs like sorting, short summaries and simple rewrites. Fastest and cheapest: from $0.10 / $0.50 per million tokens (prompts up to 100,000 tokens).
Sonnet 5.5: everyday coding, bug fixes and docs. $2 / $10 per million tokens. Start here.
Opus 5.5: big projects and complex work that needs careful judgment. $4 / $20 per million tokens.
Fable 5.1: the hardest problems and long agent jobs. $10 / $50 per million tokens.

These are Anthropic's API prices for input / output. On a Pro or Max plan you don't pay per token, but bigger models use up your limits faster.

04My rule: start on Sonnet, move up when it gets stuck

Most of my tasks don't need the biggest model. I start on Sonnet. If the task is big, touches many files, or Sonnet gets the same thing wrong twice, I switch to Opus. I only reach for Fable on the really hard stuff.

Ask Claude which model fits

Here's the task I want to do: [describe it]. Is this a quick job, an everyday coding job, or complex work that needs careful judgment? Tell me if Haiku, Sonnet, Opus or Fable fits best, and why, in two sentences.

05How to switch models in Claude Code

Type /model to open the model picker.
Pick a model. Enter saves it as your default. Press s to use it for this session only.
Or switch straight away with /model sonnet, /model opus or /model fable.
To start a session on a model, run claude --model sonnet.

Note: on Pro, Max, Team and Enterprise, Claude Code starts on Opus 5.5 by default. If you want Sonnet as your everyday model, pick it in /model and press Enter to save it.

Hand a stuck task to a bigger model

Summarise what we've tried on this task so far, what failed and why, and what's left. Keep it short so I can continue it on a different model.

06Fable as an advisor

Claude Code also has an advisor option. Instead of you switching models, your main model can decide mid-task to ask a second model for advice. The docs list it as the --advisor flag and the advisorModel setting. Putting Fable 5.1 in that advisor seat is how I'd use it next to Opus.

Fable is the priciest model. Depending on your plan, Fable usage may bill to usage credits instead of your plan's included limits. The model picker shows "Requires usage credits" when that applies.

07Good to know

The numbers here come from Anthropic's Sonnet 5.5 announcement and pricing page, checked on 20 October 2026.
API names: claude-sonnet-5-5, claude-opus-5-5, claude-haiku-5-5 and claude-fable-5-1.
Haiku 5.5 costs more per token on prompts over 100,000 tokens.
Model lineups change every few months. Check the models page before you build anything that depends on one.