Claude Sonnet 5.5 Explained: Performance, Cost, API and Migration

Research checked: September 29, 2026

Anthropic released Claude Sonnet 5.5 on September 28, 2026. It keeps Sonnet 5’s API token price while Anthropic claims substantial gains in output speed, coding, agentic work and token efficiency.

The core specification is a 1-million-token context window, up to 128,000 standard output tokens, and pricing of $2 per million input tokens and $10 per million output tokens.

Claude Sonnet 5.5 keeps Sonnet 5’s token price but changes important API behavior. It adds stronger coding and agent performance, new thinking behavior and provider-specific model IDs that can matter during migration.

Claude Sonnet 5.5 specifications

Specification	Claude Sonnet 5.5
Release	September 28, 2026
Native Claude API ID	claude-sonnet-5-5
Context window	1,000,000 tokens
Standard maximum output	128,000 tokens
Batch output beta	up to 300,000 tokens
Input	$2 / 1M tokens
Output	$10 / 1M tokens
5-minute cache write	$2.50 / 1M tokens
1-hour cache write	$4 / 1M tokens
Cache read	$0.20 / 1M tokens
Thinking	Adaptive
Default API effort	High
Knowledge cutoff	June 2026
Training cutoff	June 2026

Source: Anthropic Claude Platform, checked September 29, 2026.

Which Claude Sonnet 5.5 model ID should you use?

The correct identifier depends on the provider:

• Claude API: claude-sonnet-5-5
• Vercel AI Gateway: anthropic/claude-sonnet-5.5
• OpenRouter: anthropic/claude-sonnet-5.5
• Amazon Bedrock: anthropic.claude-sonnet-5-5
• AWS Global Inference: global.anthropic.claude-sonnet-5-5
• Google Cloud: claude-sonnet-5-5
• Microsoft Foundry: claude-sonnet-5-5

This is operationally important. A model slug copied from a gateway is not necessarily valid in Anthropic’s native SDK.

What changed from Sonnet 5?

Anthropic emphasizes two headline improvements:

1. more than 30% faster output
2. up to 30% lower cost per completed task

The second claim does not mean token pricing was reduced.

Sonnet 5 and Sonnet 5.5 both cost $2 per million input tokens and $10 per million output tokens. Anthropic attributes the task-level saving to lower token consumption and fewer tool calls.

That makes the result workload-dependent rather than a universal 30% billing reduction.

Benchmark results

Anthropic reports:

Benchmark	Sonnet 5.5	Sonnet 5	Opus 5.5
Terminal-Bench 4.0	70.6%	10.3%	66.4%
FrontierCode 1.1	46.2% Max / 52.1% Xhigh	42.4%	54.4%
CursorBench 4.0	55.5%	34.1%	57.8%
GDPval-AA v2.1	1844	1449	1846
AA-Briefcase v1.1	1811	1359	1822

These are Anthropic-reported results, not a normalized independent leaderboard.

Max effort is not automatically optimal

One of the most useful details in Anthropic’s launch data is the FrontierCode result.

Sonnet 5.5 reaches:

• 52.1% at Xhigh
• 46.2% at Max

Anthropic says Max triggered additional code-review subagents more often. In some runs, that caused timeouts or out-of-scope modifications.

For production use, teams should therefore benchmark effort settings against their own tasks instead of automatically selecting Max.

Independent results

Vals AI reports a Vals Index of 69.22 ± 0.96% for Sonnet 5.5.

Its published results include:

• Code Migration: 69.83%
• Vibe Code Bench v1.1: 92.39%
• Terminal-Bench 4.0: 53.03%

The Vals Terminal-Bench result differs substantially from Anthropic’s 70.6%.

That does not prove either number is wrong. Agentic benchmarks are highly sensitive to harness configuration, effort settings, tool behavior, fallback policies and timeouts.

Manufacturer and independent benchmark results should therefore be displayed separately.

What does Claude Sonnet 5.5 cost in practice?

A request with 10,000 input tokens and 2,000 output tokens costs:

• Input: 10,000 / 1,000,000 × $2 = $0.02
• Output: 2,000 / 1,000,000 × $10 = $0.02
• Total: $0.04

One thousand requests with the same token profile would cost approximately $40 in token charges.

Prompt caching

If 8,000 of 10,000 input tokens are cache reads:

• 2,000 new input tokens: $0.004
• 8,000 cached tokens: $0.0016
• 2,000 output tokens: $0.02
• Total: $0.0256

The initial cache write is billed separately.

Batch API

Batch processing reduces standard input and output rates by 50%, resulting in effective rates of $1 input and $5 output per million tokens.

Major migration changes

thinking: disabled is no longer supported

Sonnet 5 allowed:

{"thinking":{"type":"disabled"}}

Sonnet 5.5 rejects that configuration.

For reduced up-front reasoning, Anthropic provides:

{"thinking":{"type":"between_tools"}}

This works at Low, Medium and High effort.

Forced tool choice was removed

tool_choice: any and forcing one specific tool are no longer supported.

Anthropic recommends automatic tool selection and Strict Tool Use where valid tool arguments are required.

Thinking blocks are more tightly bound to model and conversation state

Sonnet 5.5 conversations should generally be treated as append-only. Mutating earlier messages, tools or system prompts can invalidate preserved thinking blocks.

Computer Use changed

On Claude API and Google Cloud, computer_20251124 is replaced by computer_toolset_20260801.

Provider behavior can differ.

Effort levels should be re-evaluated

Low, Medium, High, Xhigh and Max should not be assumed to behave identically to their Sonnet 5 counterparts.

Context and output limits

Claude Sonnet 5.5 supports:

• 1,000,000 context tokens
• 128,000 standard output tokens
• up to 300,000 Batch API output tokens in beta

The output limit is separate from the total context window.

Sonnet 5.5 vs Opus 5.5

Model	Input / 1M	Output / 1M
Sonnet 5.5	$2	$10
Opus 5.5	$4	$20

Sonnet 5.5 therefore costs half as much per standard input and output token as Opus 5.5.

Anthropic positions Sonnet 5.5 for well-defined coding, agent and professional workflows where speed and economics matter, while Opus 5.5 remains aimed at more difficult open-ended tasks requiring sustained judgment.

Who should use Sonnet 5.5?

Sonnet 5.5 is especially relevant for:

• coding agents
• tool-based automation
• long-context document workflows
• knowledge work
• high-volume API workloads
• applications where Opus-level reasoning is not always necessary

A blind model swap is less appropriate for workflows that are sensitive to tool selection, reasoning settings or agent-loop behavior.

Bottom line: migration comes before features

Claude Sonnet 5.5 combines a 1M-token context window, relatively low $2/$10 API pricing, and benchmark performance that approaches flagship territory on several bounded tasks.

The main caveat is migration behavior. Sonnet 5.5 is not a fully compatible drop-in replacement in every Sonnet 5 configuration.

Thinking, tool choice, conversation state, computer use and effort levels should all be tested before production rollout.