Sonnet 5.5 versus Opus 5.5: what fifteen coding projects show
Mike Codeur
Can you afford to run every coding task with Opus 5.5? Sonnet 5.5 was priced at half the rate per token. Mike compared delivered projects rather than trusting the price card. He launched fifteen projects in Claude Code with both models, using the same prompts and isolated directories, with no human intervention during the runs. Only thirteen finished on both sides; every comparison below uses those thirteen.
What the test covered
The projects include a 3D black hole, a particle galaxy, fluid simulation, landing page, configurator, dashboard, platform game and tested API. The mix matters: animations, interfaces and backend code have different failure modes. Mike used matching prompts in separate folders inside the same coding tool, then opened the results side by side. That controls part of the setup, not every variable in your own repository.
Results from the thirteen complete matchups
| Measure, thirteen common projects | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Delivered, runnable projects | 13/13 | 13/13 |
| Projects with passing tests | 13/13 | 12/13 |
| Cumulative run time | 34 min 01 sec | 2 h 15 min |
| API-equivalent cost for these runs | $4.52 | $26.34 |
| Tests written | 128 | 222 |
These are observed runtimes, not a promise about delivery time on another task. API-equivalent dollars are calculated for these runs, not a Claude Code subscription invoice or the human cost of review. More tests written does not automatically mean better software.
What happened to projects fourteen and fifteen?
Opus stopped before finishing these two runs. Sonnet has outputs, but there is no completed head-to-head result. Counting them as two Sonnet wins would change the denominator dishonestly. Passing tests and starting an app still cannot tell you whether the interaction, UI and architecture are any good. The video opens the demos and code so you can assess the deliverables, not just the scoreboard.
Token price versus cost of usable work
Anthropic listed Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, roughly half Opus 5.5's per-token price at the time. A realistic decision also includes reasoning tokens, cache, tool calls, retries, elapsed time and human correction. A cheaper token rate may lose its advantage if a run takes many more steps or fails to deliver an artifact.
Vendor and third-party benchmarks sometimes compare different effort settings or execution environments. A Sonnet Terminal-Bench figure at max is not a like-for-like comparison with an Opus figure at xhigh. Mike's fifteen projects are a situated experience, not a universal model ranking.
A test you can repeat on your own work
- Select several representative tasks: a UI, a bug fix and a tested feature.
- Freeze the initial commit, prompt, permissions and acceptance criteria.
- Track completed runs, elapsed time, full cost and post-review corrections.
- Open the resulting app and code: interactions and edge cases matter alongside test output.
- Keep incomplete runs separate instead of forcing them into a win percentage.
On these thirteen complete runs Sonnet was faster and cheaper by API-equivalent cost. Watch the implementations before deciding whether that trade-off holds for your work or whether some open-ended tasks still warrant Opus.
Join The Agentic Dev for more hands-on coding-model tests.