October 3, 2026
Claude Opus 5.5 costs less to run: Anthropic claims 40% savings on typical tasks
Anthropic promises to cut costs for typical tasks with Opus 5.5 by 40% compared with Opus 5.
At Quantium, a coding task took 11 prompts and 3 hours of work with Opus 5.5. With Opus 5, it took 38 prompts and 4 days, according to Harley Barnes.
There is another example for long-running assignments. Sean Heinz of Clio reported that Opus 5.5 worked independently for over 18 hours on an engineering task across 6 repositories. In his comparison with Opus 5, the new model moved through stages faster and needed less rework.
What changed in costs
Launch pricing. On Oct 3, Anthropic called Opus 5.5 a leader in agentic coding and knowledge work. The claimed 40% savings apply to typical tasks, while prices for individual token types changed as follows:
| Per 1 million tokens | Opus 5 | Opus 5.5 | | --- | --- | --- | | Input | $5 | $4 | | Output | $25 | $20 | | Cache reads | $0.50 | $0.20 |
Prices are based on Anthropic's Sep 22 announcement. In Fast mode, input costs $8 and output $40 per 1 million tokens. Anthropic claims a speedup of up to 2.5 times for this mode.
Where to enable the model
Opus 5.5 is available to Pro, Max, Team and Enterprise subscribers. In Claude, select the model in the model picker; for the API, use the name `claude-opus-5-5`.
Claude Code works with a project as an AI coding agent. It is a separate tool, while Opus 5.5 is a model. Anthropic's documentation dated Sep 22 gives this installation command for macOS, Linux and WSL:
```sh curl -fsSL https://claude.ai/install.sh | bash ```
After installation, run `claude` in the project directory and sign in. The Free plan does not include access to Claude Code.
| Claude Code plan | Price as of Sep 26 | | --- | --- | | Pro | From $17 per month | | Max | $100–200 per month | | Team | From $20 per seat per month |
How to assign a long-running task
The task in one message. Addy Osmani recommends describing the full outcome, completion criteria and stopping condition upfront. His example: migrate all endpoints, remove the old client and get the tests passing. These recommendations were published on the claude.dev blog on Sep 22.
For long runs, Osmani suggests adding a rule to CLAUDE.md to keep working without unnecessary pauses and ask when blocked or before destructive actions. He recommends keeping a checklist in TASKS.md so the task list survives context compaction.
You can send a clarification to a working agent: type a message and press Enter. For a large audit, Osmani suggests assigning individual services to subagents and checking the evidence from each.
Osmani recommends removing the phrase `think carefully` from prompts and saved instructions: Opus 5.5 thinks before every response. In his test, removing the phrase made responses start faster without a noticeable drop in quality.
Fast mode is enabled with the `/fast` command in Claude Code and billed through extra usage.
