Efficient mode limits Enterprise members to Sonnet and Haiku and starts sessions on Sonnet 5.5 at low effort. On Anthropic's chart Sonnet costs 41% less than Opus 5.5 there and scores 20.0% against 38.5%.
Anthropic merged chat and agent work into one Claude Enterprise product on 23 October. Billing still follows usage, leaving budget owners to estimate how many tokens each task will use.
Efficient mode, the release's main cost control, cuts the bill at every effort setting but maximum. It also removes Opus 5.5 from members' reach. At low effort, the Sonnet 5.5 it starts on scores about half as much.
Anthropic's admin guide says that with Efficient mode on, org members "can use only Sonnet and Haiku models". Those are Anthropic's lower-cost models. New sessions "start on Claude Sonnet 5.5 at low effort". The guide does not say whether members can raise the effort.
Opus is therefore not a choice inside Efficient mode. The comparison below prices what an owner gives up by switching the mode on. The guide adds that Anthropic "may update the default model and effort level" as cheaper models arrive.
What Enterprise changes
In Anthropic's announcement, chat and Cowork "first came together for Pro and Max plans in September". Cowork is Anthropic's agent mode, which runs tasks in a cloud session. The combined product is now generally available on Enterprise plans.
Pricing did not change, because Anthropic says it "stays usage-based". It adds that "agentic work can use more tokens than a quick chat".
The page also reports generational gains. Opus 5.5 costs 40% less to run than Opus 5 on typical workloads. Sonnet 5.5 costs up to 30% less than Sonnet 5 for most work.
Efficient mode is one of four cost controls on that page. Model entitlements set which models each role can use. Spend limits apply to the organisation, a group or a member. Pooled group budgets give a group one shared monthly amount.
Cost and score by effort level
Effort is the setting that controls how long a model works. Anthropic's Sonnet 5.5 page says "at lower settings, Claude answers faster and uses fewer tokens". The default is Medium in Claude Code and its apps, and High on the Claude Platform.
Sonnet 5.5 lists at half Opus 5.5's price on every rate. Sonnet costs $2 per million input tokens and $10 per million output tokens, against $4 and $20 for Opus. Cache reads, which are repeat input, are halved too. Any other cost gap comes from the tokens used.
Anthropic's Terminal-Bench 4.0 chart prices each effort level per attempt. The test sets complex tasks in a command line interface. It lets a buyer compare both models at every setting.
Sonnet is cheaper than Opus at four of the five settings. It costs 41% less at low effort, 72% less at medium, 50% less at high and 28% less at xhigh. Only at maximum does the order flip.
At low effort Sonnet's $0.76 an attempt is the cheapest on the chart. Opus costs $1.29. Sonnet scores 20.0% against Opus's 38.5%, so each point of score costs about 3.8 cents against 3.4, by simple division.
Medium is the default setting in Anthropic's apps. There Sonnet 5.5 scores 28.8% for $0.83, and Opus 5.5 scores 57.6% for $2.94. Sonnet costs 28% of the price for half the score.
Model choice moves the bill more than effort does. Dropping Sonnet from Medium to Low saves only 8%, from $0.83 to $0.76. Dropping Opus from Medium to Low saves 56%, from $2.94 to $1.29.
So the model gap narrows at low effort. Sonnet's $0.76 is 41% below Opus there, against 72% at Medium.
At maximum effort Sonnet 5.5 costs $12.54 an attempt against $11.24 for Opus 5.5, which is 12% dearer. It also scores more there, 70.6% against 64.8%. The chart shows every setting.
The independent test
The independent Terminal-Bench 4.0 leaderboard shows a larger cost disadvantage. It runs both models on the Claude Code harness, the software that drives the model, at maximum effort. Its costs are for the whole run, not one attempt.
Opus 5.5 resolved 64.8% of tasks for $4,700. Sonnet 5.5 resolved 61.8% for $7,300, which is 55% more for a lower score.
Sonnet used about 2.4 times Opus's tokens, 19.4 billion against 8.0 billion. At half the price per token, that alone would cost about 21% more, by our arithmetic. The leaderboard shows 55%, and it does not break out token types.
Anthropic's chart and the leaderboard agree that Sonnet is dearer at maximum effort. They differ on the score, 70.6% against 61.8%. Anthropic says Sonnet 5.5 suits lower settings, "where it costs less per task". At higher settings it "can perform comparably at a similar cost".
At maximum effort Sonnet 5.5 is 39% dearer on that test, at $29.19 against $21.05.
Efficient mode’s savings come with lower scores. On Terminal-Bench the low effort Sonnet attempt scores 20.0%. Opus 5.5 scores 57.6% at the Medium default. Whether the saving holds depends on how many failed tasks get rerun.
Rollout and access
The release keeps single sign-on (SSO) and the Compliance API, which Enterprise customers already use. Each agent session runs in its own isolated virtual machine, with network egress (outbound traffic) off by default. That isolation is new.
Three groups of customers stay on their current setup. They use zero data retention for Claude Code, a HIPAA ready configuration, or customer managed encryption keys. The admin guide says the new experience is not yet available to them.
The unified experience switches on by default for existing Enterprise organisations on 16 November 2026 unless an owner changes it. Owners turn on Efficient mode separately, on the Models page or for specific custom roles.
Owners can already chart spend by model on Anthropic's usage page. The guide advises two to four weeks of monitoring before expanding. A pilot role started on 16 November would have its answer by 14 December: whether a 41% saving is worth a 20.0% score.