Anthropic Launches Claude Sonnet 5.5 at $2 Input and $10 Output per Million Tokens

Claude Sonnet 5.5 launched on September 28

Anthropic introduced Claude Sonnet 5.5 on September 28, 2026 as the second model in its Claude 5.5 family. The company positions it below Opus 5.5 for open-ended work requiring sustained judgment and above a future Haiku 5.5 for well-scoped coding, document, slide and spreadsheet tasks. The direct Claude API model ID is claude-sonnet-5-5.

The live model page documents a 1-million-token context window and a 128K-token maximum output. It supports text and image input, text output, multilingual use, vision, tool use and adaptive thinking. Those interface facts establish what developers can request; they do not establish reliable recall across the full context or safe autonomy across a long tool trajectory.

Direct API pricing is $2 in and $10 out per million tokens

Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens, $10 per million output tokens and $0.20 per million cache-read tokens—the same rates it cites for Sonnet 5. The announcement says the new model costs up to 30% less per task in Anthropic’s testing because it typically uses fewer tokens for the same work. That task-economics result is a vendor claim and should not be substituted for a workload forecast.

A production comparison should price the entire request path: uncached and cached input, thinking and visible output, retries, failed tool turns and any surrounding retrieval or execution. A nominally unchanged token rate can still produce a different bill if the model changes output length, tool-call count, cache hit rate or retry behavior. AWS directs Bedrock customers to its live regional pricing table; confirm the rate displayed for the intended account and route before approval rather than assuming the direct Claude API tariff applies.

Amazon Bedrock has a separate global model ID

AWS announced Sonnet 5.5 availability on Amazon Bedrock on September 28. Its example uses global.anthropic.claude-sonnet-5-5 through the Global Cross-Region Inference, or Global CRIS, profile on bedrock-runtime. Developers can use Bedrock Invoke and Converse APIs or Anthropic’s Messages API against Bedrock. The example requires bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream permissions.

Global CRIS is an access route, not a promise that every application can satisfy a fixed in-region processing requirement. AWS tells customers to consult the current supported-region documentation and separately says Claude Platform on AWS is available in North America. Confirm model access, inference-profile routing, data residency, quotas and contract terms in the actual account before moving regulated traffic.

Sonnet 5 migrations have documented breaking changes

Anthropic lists five breaking changes for applications already using Sonnet 5. Up-front thinking can be disabled with between_tools; forced tool use now returns an error; thinking blocks are tied to the model and conversation; an earlier computer-use tool version is not accepted on the Claude API and Google Cloud; and several older Claude models cannot serve as advisors for the advisor tool.

There is also a non-failing response-shape change: text between tool calls can arrive inside thinking blocks. An application that only streams ordinary text may appear silent between calls unless it chooses the documented display behavior or disables up-front between-tools thinking. Treat the migration as a parser, state and tool-contract change—not only a model string update.

Anthropic reports large coding gains and faster generation

Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5. It reports 70.6% versus 10.3% on Terminal-Bench 4.0, 55.5% versus 34.1% on CursorBench 4.0 and 46.2% at Max effort versus 42.4% on FrontierCode 1.1 Main. It also reports a GDPval-AA v2.1 score within two points of Opus 5.5. These are vendor-selected and vendor-reported results under the announcement’s footnotes and settings.

The unusually large Terminal-Bench gap makes methodology review especially important. A buyer should not infer that its repository tasks improve by the same amount. Preserve the exact model, effort, harness, tools, time and token ceilings, scorer and task set when comparing results, then run representative private tasks that were not used to tune prompts or select the model.

Cyber safeguards now apply to a Sonnet model

Anthropic says Sonnet 5.5 is the first Sonnet model launched with cyber safeguards and fallbacks similar to those used for its most capable models because its cybersecurity capabilities are comparable to Opus 5. It says biology safeguards remain the same as Sonnet 5 and that both target a narrow set of high-risk requests. These are Anthropic’s descriptions of its own controls, not an independent safety certification.

Builders should test ordinary authorized security and life-sciences workflows for false refusals, escalation behavior and provider differences while maintaining their own authorization, monitoring and review controls. Provider safeguards do not grant permission to run a task, contain tool side effects or replace application-level access control.

Use a shadow migration before changing production traffic

Start by pinning the current Sonnet 5 baseline and Sonnet 5.5 candidate with the same prompts, tools, fixtures and effort policy. Replay complete workflows in shadow mode. Record accepted-task quality, first-token and completion latency, input, cache-read, thinking and output tokens, tool-call validity, parser failures, retries, refusals and cost per accepted task. Add explicit fixtures for every documented breaking change.

Move a small reversible cohort only after response parsing, tool state and cost gates pass. Keep the old model available during rollback. Wait when Global CRIS conflicts with residency requirements, an AWS account does not expose the expected route or price, or a tool integration depends on old thinking-block or computer-use behavior. Reject unattended authority where model output can trigger consequential writes without deterministic authorization and review.

Disclosures

AccessAllGPT did not receive Anthropic or AWS credentials, credits, early access, a briefing, benchmark data, review or compensation for this article. Anthropic and Amazon Web Services did not sponsor, review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Anthropic, AWS or organizations cited. Publication-wide relationships are listed on the disclosures page.