Anthropic cuts Claude Opus 5.5 pricing to $4 per million input tokens

Anthropic has updated its Claude Opus 3.5 pricing, cutting costs to $3 per million input tokens. To maximize value from this change, teams can follow a seven-step 'mise en place' checklist for disciplined AI deployment. This structured framework ensures organizations achieve reliable and secure results by prioritizing governance and careful preparation.

The 'mise en place' methodology, borrowed from professional kitchens, has become a best practice for disciplined AI implementation. This step-by-step prep list keeps teams focused on foundational setup rather than chasing complex prompts. The following seven stations combine official guidance on Claude Opus 3.5 with forward-looking governance principles to help you reproduce dependable results.

Station 1: Print the ticket (context)

Begin by creating a clear system message defining the goal, audience, output format, and key constraints. While Opus 3.5 offers a large context window, clarity remains more critical than volume. The model features adaptive thinking capabilities and can generate substantial output tokens, improving coherence for long tasks. Use the large window to provide complete source documents rather than making the model infer details.

To achieve consistent and usable outputs from Claude Opus 3.5, focus on systematic preparation before generating content. This 'mise en place' approach emphasizes defining clear context, implementing security controls, and running structured tests. This disciplined setup is more effective than simply adjusting model settings for performance.

Station 2: Pick the Team plan and model

Team workspaces automatically use the latest Opus 3.5 engine. The model overview confirms the new pricing structure with reduced rates for both input and output tokens. For initial drafts, keep the per-message effort set to 'medium', as higher effort settings rarely improve quality and can increase latency and cost.

Station 3: Flip privacy and memory switches

By default, Anthropic does not use Team plan data for model training. Workspace owners can enable memory for Claude to recall previous conversations, but be aware that deletion is permanent and immediate. For security, restrict memory to channels with a documented data retention schedule. Following enterprise best practices involves encrypting prompts and logs at rest.

Station 4: Add connectors with approvals on

Before integrating Claude with tools like email, code repositories, or ticketing systems, configure every tool to require human approval for any action that writes, deletes, or posts data. While Anthropic includes a safety classifier, manual sign-off is a crucial defense against prompt injection and unintended actions. Use a dedicated service account for connectors to ensure audit logs clearly trace the origin of each command.

Station 5: Write reusable skills

Skills are standardized, reusable prompt snippets for recurring tasks like 'summarize a pull request' or 'draft a Jira comment.' Store these skills in a shared, version-controlled repository. Naming each skill and its inputs clearly simplifies future prompts and enables controlled, predictable automation.

A compact starter set:
- Summarize meeting transcript into bullet actions
- Rewrite text to plain-language grade 8
- Extract entities into JSON schema
- Generate unit tests from code patch
- Draft stakeholder update in 150 words

Station 6: Run three test orders

Dedicate at least 20 minutes to hands-on testing with three representative tasks: one easy, one average, and one complex edge case. This process helps identify potential hallucinations or performance issues early. Record key metrics like latency, token cost, and the number of revisions needed. This data serves as a baseline governance artifact, proving the model met acceptance criteria before a wider rollout.

Station 7: Open for first service

After completing the previous six preparation steps, invite a small pilot group to begin using the model. During the first week, monitor daily for performance drift, bias, or security events. Implement alerts that trigger if response lengths exceed policy or if quality scores drop below your established baseline. Once the service proves stable for a week, you can scale the deployment to the broader team using the same controls.