Claude Opus 5.5 lowers the cost of long-running coding agents

Anthropic released Claude Opus 5.5 on September 22, 2026, introducing the first model in its 5.5 family. Microsoft’s GitHub added it to Copilot for Pro+, Max, Business, and Enterprise customers, while Microsoft Foundry joined the Claude Platform, AWS, and Google Cloud as deployment options. GitHub details the rollout in its Copilot announcement.

The release targets a persistent problem with autonomous coding agents: each planning step, tool call, and retry adds tokens, latency, and cost. Anthropic and early deployment partners report that Opus 5.5 completes difficult tasks with quality comparable to Opus 5 while using fewer steps and tokens.

The price cut compounds

Anthropic says typical Opus 5.5 workloads cost about 40% less than equivalent Opus 5 runs. Standard input and output prices fell 20%, while cached reads fell 60%, according to Anthropic’s model page.

Cached reads reuse previously stored prompt context, while cache writes store that context for later requests. Reads can dominate the bill for agents that repeatedly send repository state, instructions, and conversation history. Actual savings therefore depend on each workload’s mix of fresh input, cached context, and generated output.

An optional Fast mode in Claude Code and the Claude Platform runs at up to 2.5 times the standard speed. It costs $8 per million input tokens and $40 per million output tokens. Anthropic also reports that standard output generation is more than 30% faster than Opus 5.

GitHub’s tests used fewer turns

GitHub tested Opus 5.5 in its Copilot evaluation harnesses before release. The company found that it resolved tasks at a rate comparable to Opus 5 while consuming fewer steps and tokens, and that it recovered faster from errors during multistep work. In VS Code terminal tests, GitHub’s chief product officer said the model solved more tasks than Opus 5 in less than half as many steps.

Deployment partners reported similar results. Lovable measured 33% to 50% fewer steps and lower token use. Optiver said Opus 5.5 matched Opus 5 quality in roughly half the turns, reducing workload costs by 40% to 50%. At Clio, an engineer ran the model unattended for 18 hours while it defined service contracts across six repositories.

Those reports come from selected early users and use different tasks, environments, and scoring methods. They indicate where the efficiency gains may appear, but they do not provide a standardized comparison across deployments.

Vendor benchmarks show a lead

Anthropic’s published results place Opus 5.5 ahead of Opus 5 on agentic coding, computer use, and knowledge-work evaluations. The company also cautions that small benchmark margins have become less reliable predictors of production behavior among high-capability models.

Effort settings control how much reasoning work the model applies to a request. Anthropic says Opus 5.5 at default effort exceeded GPT-6 Astra’s highest FrontierCode score at about one-fifth of the cost per task. It also beat GPT-5.6 Sol on CursorBench by 11 percentage points at roughly one-third of the cost.

Large codebases expose the gains

Anthropic’s early testing emphasizes long-running, multi-file work, where reductions in tokens and turns accumulate across an agent’s tool calls. One tester completed a 680,000-line migration in less than a day. Another audited and repaired a 200,000-line codebase in under three hours; Anthropic says Opus 5 required more than 20 hours and 2.5 times as many tokens for the same work.

In an internal test, Opus 5.5 translated HAProxy from C to Rust in 9.5 hours and cost 51% less than Fable 5.1 on the same task. These large migrations test whether a model can preserve architectural context, coordinate edits across files, run tools, and recover when intermediate changes fail.

Results extended beyond coding. Hebbia reported 86.6% coverage on finance workflows graded against expert rubrics, compared with 60.3% for Opus 5. Deloitte said Opus 5.5 found 72% of known bugs during code review at its lowest effort setting; Opus 5 found 56% at high effort.

Less hedging, shorter answers

Anthropic tuned Opus 5.5 to place important information earlier, reduce jargon, and follow requested writing styles more consistently. The changes address the verbose and heavily qualified prose reported by some Opus 5 users.

Ramp said the model produced a design specification that required minimal editing. Box measured responses that were 40% shorter without a corresponding loss of accuracy. For coding agents, shorter explanations can also reduce the context carried into subsequent turns.

Guardrails can change model behavior

Opus 5.5 is the first Opus release to use the same safeguard class as Anthropic’s Fable 5.1 tier. The controls cover cybersecurity, biology, and anti-distillation measures designed to impede attempts to extract or reproduce model behavior.

• Cybersecurity: Qualifying requests are routed automatically to Opus 4.8, so the model handling a request may differ from the one selected in Copilot.
• Biology: Organizations whose legitimate work is blocked by safeguards can apply to Anthropic’s Life Sciences Verification Program.
• Reasoning context: Preserved thinking prevents API clients from editing earlier reasoning blocks. Applications that modify assistant history before resending it may require changes.
• Thinking mode: Opus 5.5 does not allow developers to disable thinking mode.
• Watermarking: Text outputs include watermarks that Anthropic says support its EU AI Act compliance program.

Calling Opus 5.5

GitHub Copilot users on Pro+, Max, Business, and Enterprise plans can select Opus 5.5 from the chat model picker. Usage is billed at provider list prices through usage-based billing.

API clients use the model identifier claude-opus-5-5. The model is available through the Claude Platform, AWS, Google Cloud, and Microsoft Foundry.

from anthropic import Anthropic

client = Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    messages=[
        {
            "role": "user",
            "content": "Refactor this module..."
        }
    ],
)

print(response.content)

The economics of unattended work

The combined reduction in cache prices, generated tokens, and agent steps changes the budget for tasks that run for hours or days. Long sessions repeatedly process repository context and accumulate output, so modest per-turn savings can produce much larger reductions across an entire job.

Clio’s 18-hour, six-repository run illustrates that pattern. Stripe also reported an Opus 5.5 session coordinating 12 other sessions through a multi-day rebase of 40 stacked pull requests; every request passed continuous integration the following afternoon. Workloads of that scale become easier to justify when each agent loop costs less and requires fewer retries.

Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, extending the same efficiency strategy to lower-priced models.