Anthropic released Claude Sonnet 5.5 on September 28, 2026, with model ID claude-sonnet-5-5, per Claude Sonnet 5.5's own overview page. Unlike Opus 5.5's 20% cut the week before, Sonnet 5.5 costs exactly what Sonnet 5 already costs: $2 per million input tokens, $10 per million output tokens. The real story is five breaking API changes that hit any code still calling claude-sonnet-5 the moment the model ID swaps.
What Claude Sonnet 5.5 costs, and what actually changed
That overview page lists Sonnet 5.5's status as "Active (latest)," putting it ahead of Sonnet 5 in Anthropic's current lineup. The model ships with a 1 million-token context window, a 128,000-token maximum output on the Messages API, and a June 2026 knowledge cutoff. Default effort is high. On the Batch API, still in beta, max output extends to 300,000 tokens, but only when the request carries the output-300k-2026-03-24 header.
| Spec | Value |
|---|---|
| Input | $2/MTok |
| Output | $10/MTok |
| Cache write (5m) | $2.50/MTok |
| Cache write (1h) | $4/MTok |
| Cache read | $0.20/MTok |
| Context window | 1M tokens |
| Max output (Messages API) | 128K tokens |
| Max output (Batch API, beta) | 300K tokens |
| Knowledge cutoff | June 2026 |
| Default effort | high |
Anthropic states the pricing story plainly on Sonnet 5.5's what's-new page: the model runs at "the same prices as Claude Sonnet 5, including prompt caching and batch processing rates." There's no re-pricing angle here, in other words, only a model swap.
On Terminal-Bench 4.0, Anthropic's own benchmark table puts Sonnet 5.5 at 70.6%, against Sonnet 5's 10.3%, per Anthropic's Sonnet 5.5 announcement. Both figures come from Anthropic, not an independent lab, so treat the gap as a headline rather than a verified result.
How to migrate from Sonnet 5 to Sonnet 5.5
Sonnet 5.5's own migration notes don't hedge on scope: per the breaking-changes list on Sonnet 5.5's what's-new page, five breaking changes affect code already running on Claude Sonnet 5. Each one returns a 400 rather than failing quietly, so a straight model-ID swap surfaces the problems fast, provided you test before rollout rather than after.
1. Replace disabled thinking with between_tools
Sending thinking: {"type": "disabled"} now returns a 400 invalid_request_error pointing at between_tools instead. That replacement only works at low, medium, or high effort; at xhigh or max, omit the thinking parameter entirely or send {"type": "adaptive"}. Manually setting a budget with thinking: {"type": "enabled", "budget_tokens": N} gets rejected the same way.
2. Replace forced tool_choice with auto
Setting tool_choice: {"type": "any"} or {"type": "tool", "name": ...} now returns: tool_choice: type "tool" and "any" are not supported for this model. Switch to tool_choice: {"type": "auto"} with strict: true, or move the logic to structured outputs.
3. Keep conversations append-only for thinking blocks
Thinking blocks are account-bound. In Anthropic's own words, "thinking blocks that Claude Sonnet 5.5 produces work only in the account that produced them, or in an account linked to it." Outside that scope, they're dropped silently rather than billed. For accounts created on or after August 31, 2026, 00:00 UTC, Anthropic also checks a thinking block's position against any edit to the system prompt, tools, or earlier messages, on the Claude API, Bedrock, and Google Cloud. A mismatch now 400s unless the request sends the thinking-binding-controls-2026-08-01 beta header with block_binding.prefix_mismatch_behavior set to "drop_block".
4. Move off computer_20251124 (Claude API and Google Cloud only)
The error names the model directly: 'claude-sonnet-5-5' does not support tool types: computer_20251124. That rejection applies to the Claude API and Google Cloud. Replace the tool with {"type": "computer_toolset_20260801"}. Amazon Bedrock is the exception: Anthropic's own line is that "on Amazon Bedrock, Claude Sonnet 5.5 accepts the earlier computer_20251124 tool."
5. Update advisor tool pairings
The advisor tool now rejects Claude Opus 4.8, Claude Opus 4.7, and Claude Sonnet 5 as advisors for a Sonnet 5.5 executor, returning a 400 invalid_request_error. Accepted advisors are Mythos 5.1, Fable 5.1, Mythos 5, Fable 5, Opus 5.5, Opus 5, or Sonnet 5.5 itself. If an advisor-pairing config still names Sonnet 5 or either Opus 4.x model, fix it before the swap goes out, not after.
What changes without any code change
Two behavior shifts land even on integrations that touch nothing. Text that used to stream between tool calls now returns as thinking progress-update blocks, and those blocks are empty by default because display defaults to "omitted". An app built to show that in-between narration to users goes quiet mid-task until display is set explicitly or the flow moves to between_tools.
Effort levels were also recalibrated. An effort value carried straight over from Sonnet 5 doesn't produce the same thinking depth on Sonnet 5.5. Re-run the effort sweep against the new model rather than assuming last quarter's tuning still holds.
Is Claude Sonnet 5.5 actually the same price as GPT-6 Sol?
Yes, on the headline numbers. Sonnet 5.5 and OpenAI's GPT-6 Sol both charge $2 per million input tokens and $10 per million output tokens, per OpenAI's model specs page for GPT-6 Sol, which also lists a 1,050,000-token context window, a 128,000-token max output, and cached input at $0.20 per million tokens. Sonnet 5.5's context window rounds to 1M tokens against Sol's 1,050,000, and the two models shipped within a week of each other, Sonnet 5.5 on September 28 and GPT-6 Sol on September 22, per OpenAI's API changelog entry for Sol and Luna.
That parity matters for a routing or budgeting decision between the two vendors, not for a cost-savings pitch. The rate cards already match, so there's no arbitrage to chase, only a real choice between two tiers priced identically.
Where Claude Sonnet 5.5 is available
Sonnet 5.5 runs on the Claude API for all customers, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, all under the model ID claude-sonnet-5-5 (Bedrock uses anthropic.claude-sonnet-5-5). GitHub added it to Copilot the same day, per GitHub's own changelog for the release, on Pro, Pro+, Max, Business, and Enterprise plans, billed at provider list pricing under usage-based billing.
It's one of several releases this month tracked in AI model releases and pricing: what changes for engineering teams. Claude Opus 5.5 forced four breaking changes the same month, on top of a 20% price cut that Sonnet 5.5 doesn't share. Over at OpenAI, its flagship tier priced and shipped the same month, with its own set of breaking changes to the Responses API and tool calling.
Sonnet 5.5's own docs carry a "Refusals, fallback, and billing" section that points at the same stop_details.category field documented in full in how Anthropic now bills some refusals before any output. That billing change is scoped to Fable 5.1, Fable 5, Opus 5.5, and Opus 5, not Sonnet 5.5, so don't assume the pre-output charge applies here without checking.
FAQ
What breaks if I just swap the model ID from claude-sonnet-5 to claude-sonnet-5-5?
Five things: disabled or manually budgeted thinking, forced tool_choice, thinking-block position checks on accounts created on or after August 31, 2026, computer_20251124 outside Bedrock, and advisor-tool pairings with Opus 4.8, Opus 4.7, or Sonnet 5. Each one returns a 400 rather than failing quietly.
Does Claude Sonnet 5.5 cost more than Claude Sonnet 5?
No. Same $2/MTok input, $10/MTok output, and the same caching and batch rates, per Anthropic's own pricing page.
Is Claude Sonnet 5.5 the same price as GPT-6 Sol?
Yes, on the headline numbers: both charge $2 input and $10 output per million tokens, per each vendor's own model page.
Does the computer_20251124 restriction affect Amazon Bedrock?
No. The Claude API and Google Cloud reject computer_20251124 outright; Bedrock still accepts it.