Anthropic Launches Claude Sonnet 5.5, but Effort Settings Decide the Bill

Anthropic released Claude Sonnet 5.5 on Monday, September 28, a mid-tier model that the company says comes close to its flagship Opus 5.5 on coding and office work at a fraction of the cost. For enterprise teams, Sonnet is usually the model that runs high-volume work, so a cheaper and faster version could change the math on agents and coding tools. The first independent test, though, shows the bill depends heavily on how hard the model is told to think.

What did Anthropic release?

According to Anthropic's launch page, Sonnet 5.5 keeps the same list price as Sonnet 5, $2 per million input tokens and $10 per million output tokens. The company says the model generates output more than 30% faster and, in its own testing, usually finishes the same work with far fewer tokens, which cuts the cost per task by up to 30% compared with Sonnet 5.

It is available through the Claude Platform and on Amazon Web Services, Google Cloud and Microsoft Azure, with zero data retention available, as with earlier models. Anthropic says Haiku 5.5, the smallest model in the family, will follow in the coming weeks.

The benchmark claims are strong. Anthropic reports 70.6% on Terminal-Bench 4.0, a test of command-line coding work, up from 10.3% for Sonnet 5 and ahead of the 66.4% it lists for Opus 5.5. On OSWorld 2.1, which measures how well a model operates a computer, it reports 80.1% against 57.0% for Sonnet 5. Anthropic still says Opus 5.5 is clearly stronger on complex, open-ended work that requires sustained judgment.

Customers quoted on the page include Box, Zendesk, Atlassian, Slack and Balyasny Asset Management. Zendesk reported tickets processed 20% faster, Box reported 12% fewer total tokens, and Balyasny said Sonnet 5.5 gave the best tradeoff between quality and cost of seven models it tested for high-volume workflows.

Does Sonnet 5.5 actually cost less?

That depends on the effort setting. Sonnet 5.5 offers five levels, from Low to Max. Anthropic sets Medium as the default in Claude Code and its apps and High as the default on its developer platform, and it frames many of its savings at those settings. At Medium effort, for example, it says the model beats Sonnet 5's best Terminal-Bench score for less than a tenth of the cost per task.

At maximum effort, the picture changes. Artificial Analysis, an independent benchmarking firm, ranked Sonnet 5.5 second on its Intelligence Index, two points behind Opus 5.5 at max effort. But it measured about 193,000 output tokens per task at max effort, which it called the "Heaviest token use we have measured." That pushed the cost to run its index to $7.60 per task, about 50% more than Sonnet 5. The firm also found that OpenAI's GPT-6 Sol delivered similar or better results with fewer output tokens, and it noted that its tests used a pre-release build with a bug that Anthropic says is fixed.

The two findings are not contradictory. Anthropic measured savings at moderate settings, and Artificial Analysis measured cost at the highest one. For buyers, the lesson is that the effort level is now a budget decision as much as a quality decision.

What else changed for enterprise teams?

Sonnet 5.5 is the first Sonnet model to ship with the cyber safeguards Anthropic uses on Opus 5.5, according to the launch page. Higher-risk cybersecurity requests visibly fall back to Sonnet 5, while routine bug fixing is not affected. Security teams with legitimate needs can apply through Anthropic's Cyber Verification Program.

Teams running Sonnet with extended thinking turned off will need to switch to a new setting before migrating, Anthropic says.

How should enterprises evaluate it?

Leaderboard results will not tell a company how the model handles its own documents, tickets or code. As Micaela Kaplan of HumanSignal put it at COLLIDE 2025, "Most of the time your data isn't public," which is why she argued for custom benchmarks built on a company's own tasks.

Swapping in a cheaper model is also rarely what decides whether a project succeeds. "Model selection is not the challenge," Shan Zhou, Director of Sales Engineering at Kindo, said at COLLIDE 2026 this week, arguing that systems, governance and cost matter more. His session covered why AI governance has to run at runtime.

A practical plan for Sonnet 5.5:

• Test at the effort level you will run. Compare Medium and High against your current model on 20 to 50 real tasks, and record cost per completed task, not just price per token.
• Cap the maximum effort. Max effort can multiply token use, so decide which workflows, if any, justify it.
• Watch the migration notes. Thinking settings and cyber fallbacks may change how existing agents behave.

Cheaper per-task costs only help if someone tracks them. As Verisk's experience with making AI pay for AI shows, savings that are not measured tend to disappear into new usage. The next test is Haiku 5.5, which will show whether Anthropic can push the same efficiency into its lowest-cost tier.

Frequently asked questions

When was Claude Sonnet 5.5 released?

Anthropic released Claude Sonnet 5.5 on September 28, 2026. It is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure.

How much does Claude Sonnet 5.5 cost?

List pricing is $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. Actual cost per task depends heavily on the effort setting used.

Is Sonnet 5.5 better than Opus 5.5?

On some benchmarks, such as Anthropic's reported Terminal-Bench 4.0 result, Sonnet 5.5 scores higher. Anthropic says Opus 5.5 remains clearly stronger on complex, open-ended work requiring sustained judgment.

Disclosure: AI Enterprise Magazine uses AI tools in its editorial process, including Anthropic's Claude models. Anthropic had no role in this article.

Filed under: AI Strategy & ROI, Agentic AI, Observability & Data Quality