Loading...

Anthropic introduced Claude Sonnet 5.5 on September 28. It says the model generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task in its testing, although the listed input and output token prices remain unchanged. For teams operating AI agents, the interesting question is not whether a new model wins a leaderboard. It is whether the same workflow completes more reliably, with fewer steps, within your latency and cost budget. Source

That distinction matters because an agent's bill includes much more than one model response. Every tool call, retry, failed edit and review cycle adds time. A faster model can be a poor upgrade if it needs more corrections; a model with the same token price can be cheaper if it uses fewer tokens and finishes in fewer attempts. Anthropic's launch makes both claims, but your own production-like evaluation should decide whether they hold for your tasks. Source

What Anthropic actually announced

Sonnet 5.5 is the second model in Anthropic's Claude 5.5 family, positioned as a faster, lower-cost complement to Opus 5.5. Anthropic describes Sonnet as the fit for well-scoped everyday tasks, coding fixes, and document or presentation work. It still describes Opus as stronger on complex, open-ended work requiring sustained judgment. That is a product positioning, not a guarantee that one model will win on every enterprise workload. Source

The headline API prices are $2 per million input tokens, $10 per million output tokens, and $0.20 per million cache-read tokens, the same as Sonnet 5. Anthropic attributes the potential task-cost reduction to using fewer tokens, rather than to a lower token rate. Its launch table also lists $2.50 per million cache-write tokens. These are published list prices; your bill depends on your usage pattern and commercial terms. Source

Anthropic reports a 70.6% score for Sonnet 5.5 on Terminal-Bench 4.0 versus 10.3% for Sonnet 5. It also reports a 55.5% CursorBench 4.0 score and a 46.2% FrontierCode 1.1 main-set score at Max effort, with a 52.1% score at Xhigh. These are benchmark results under specified settings, not predictions for your codebase. Some comparison rows use different effort levels, and Anthropic's footnotes discuss methodology and known pre-release issues. Preserve that context before putting a percentage in a procurement slide. Source

A concrete workflow example

Imagine an agent asked to fix a flaky integration test. It reads the failing test, locates the implementation, proposes a patch, runs the focused suite, reviews the diff and submits a pull request for a human to approve. The model may be faster in its individual responses, but the outcome depends on the whole loop: whether it edits the right file, respects the requested scope, uses the test runner correctly and stops instead of adding unrelated changes. This is an illustrative workflow, not a Sonnet 5.5 benchmark result.

A useful evaluation records wall-clock time, total tokens, tool calls, test passes, reviewer edits and whether the pull request would be accepted. Compare the same tasks against your current model with the same tools and instructions. Run enough varied tasks to expose hard cases, and inspect failures, not only averages. A model that saves thirty seconds on routine cases but breaks a production configuration on one outlier is not automatically a win.

Why task cost matters more than token price

Anthropic says early testers observed Sonnet 5.5 batching tool calls and taking fewer steps than Sonnet 5; its own claim of up to 30% lower task cost is based on its testing. The phrase "up to" matters. It identifies a best-case bound within the provider's reported scenarios, not a universal discount. A fair internal estimate should calculate the complete cost per successful outcome, including retries and human intervention. Source

For a support agent, measure cost per correctly closed case, not cost per answer. For a coding agent, measure cost per accepted change, not cost per generated patch. For a research agent, measure cost per sourced and reviewed brief. Those denominators penalize the failure modes that simplistic token benchmarks hide. If a model answers twice as quickly but requires twice as much manual correction, the organization has not saved time.

The launch also changes the latency equation. Anthropic reports output generation more than 30% faster than Sonnet 5. In interactive workflows, this could make an agent feel more responsive; in batch work, it could improve throughput. But queueing, external APIs, tests and human approvals may dominate end-to-end time. Measure the delay at each stage before attributing a slow workflow entirely to the model. Source

Safety controls still belong outside the model

Anthropic says Sonnet 5.5 improves on or matches Sonnet 5 on most measures in its automated behavioral audit. It also says its cybersecurity capabilities are high enough to warrant safeguards and fallback behavior similar to those used for more capable models. For higher-risk cybersecurity requests, the product may visibly fall back to Sonnet 5; routine software development and bug fixing are described as unaffected. These are provider statements about its own evaluations and deployment, not an independent security certification. Source

One of Anthropic's more important qualifications is that no set of evaluations catches every failure. It reports that Sonnet 5.5 came close to Opus 5.5 on newer containment evaluations and was the least likely of its models to probe container limits, while Opus 5.5 performed slightly better overall across its audit. That does not make a model-generated instruction safe to execute with unrestricted access. Separate the model's recommendations from the permissions to read data, call tools and write to external systems. Source

Our AgentOps practice treats model upgrades as control changes. Keep a scoped tool allowlist; require explicit approval for expensive or irreversible actions; log the tool calls and their results; and maintain a tested rollback path. A model's safety posture and the application's permission boundary solve different problems. The same principle applies whether the agent is a coding assistant, customer-support worker or internal analyst.

Anthropic's engineering team has described a related architecture for managed agents: separate the agent session, harness and sandbox, and keep credentials outside the environment where generated code runs. That is a useful design pattern for anyone building long-running workflows, though it does not tell you how Sonnet 5.5 will behave in your environment. Source

A practical migration test

1. Freeze the task set

Collect representative tasks from recent work, remove sensitive data where possible, and include both successes and known failures. A useful set has easy, moderate and difficult tasks, plus adversarial cases such as contradictory instructions, malformed tool output and prompts that ask an agent to exceed its scope. Do not change the task set after seeing the model's first results just to flatter the new version.

2. Keep the harness constant

Run Sonnet 5 and Sonnet 5.5 with the same prompt, tool definitions, retrieval sources, sandbox, timeout and review rubric. If you change the model and the harness together, you will not know which change caused the result. Anthropic has noted in its own engineering writing that harness assumptions can become stale when models improve; a second experiment can tune the harness after the controlled comparison. Source

3. Grade the outcome, not the prose

For a coding agent, tests and reviewer acceptance beat a confident explanation. For customer support, use a rubric that checks accuracy, policy compliance and whether the case was actually resolved. Have human reviewers inspect a sample of traces and tool actions. Record both helpfulness and safety: a model that refuses every task has not succeeded, and a model that completes tasks by exposing private data has failed.

4. Calculate cost per success

Track input, output, cache reads, cache writes and retries, then divide the total by verified successful tasks. The published API prices are a starting point, but your effective spend depends on the mix. Include any human review time that changes between variants. For a user-facing application, also track p50 and p95 completion time, because the slowest cases shape trust. Source

5. Roll out with an escape hatch

Start with a small slice of low-risk traffic. Keep the old model route available, watch error rates and user complaints, and define rollback thresholds before deployment. Where your agent can change code, send messages or access customer records, confirm that permission checks and audit logging still fire in the same way. A model swap should not silently widen the system's authority.

Choosing where Sonnet 5.5 belongs

The launch positions Sonnet 5.5 for well-scoped work and Opus 5.5 for difficult, open-ended judgment. That suggests a routing experiment rather than a blanket migration. Try Sonnet on bounded coding repairs, document transformations and routine analyses; reserve a stronger model or a human escalation for ambiguous, high-impact cases. Your evaluation should decide the boundary, because Anthropic itself says benchmark performance is only one facet of capability. Source

For example, a development team might let Sonnet 5.5 prepare a patch and run tests inside a sandbox while requiring a human to approve any merge. An operations team might let it summarize incidents from approved sources but not change production settings. A support team could give it read-only case context while holding refunds or account updates behind separate controls. These are deployment patterns to test, not product features Anthropic promises.

Anthropic says Sonnet 5.5 is available on its platform and through AWS, Google Cloud and Microsoft Azure, with the model identifier claude-sonnet-5-5 on the Claude Platform. It notes a migration detail for users running Sonnet with thinking off: they need the new between_tools setting to keep up-front thinking off. Check the provider's current migration instructions before changing a live integration. Source

A model launch is an occasion to retest the entire system, not merely replace an identifier. If your team needs to map agent permissions, evaluation criteria and rollback paths before a rollout, Codesprint's IT consulting team can help assess the architecture, or contact us to discuss an AgentOps review.