Claude Opus 5.5 Launches With Lower Pricing, Faster Output, and Higher Claude Code Limits

Anthropic launched Claude Opus 5.5 on September 22, 2026. It is the first model in the Claude 5.5 family and is available in Claude Code and the Claude Platform under the API identifier claude-opus-5-5 (Anthropic's launch page). Anthropic positions it as an Opus-class model for agentic coding, computer use, knowledge work, and long-running tasks.

The practical story is lower operating cost without giving up the Opus tier. Anthropic says Opus 5.5 is about 40% cheaper to run than Opus 5 on typical workloads and produces output more than 30% faster. The launch page also reports stronger results on Anthropic's agentic coding and knowledge-work evaluations, including Terminal-Bench 4.0, CursorBench 4.0, and GDPval-AA v2.1 (Anthropic's model results).

Key takeaways

• The model ID is claude-opus-5-5. Claude Code users can run /claude-api migrate to switch to it.
• Standard API pricing is $4 per million input tokens and $20 per million output tokens. Cache reads are $0.20 per million tokens, down from $0.50 for Opus 5.
• Fast mode is available in Claude Code and the Claude Platform. It costs $8 per million input tokens and $40 per million output tokens and can run up to 2.5 times faster.
• Thinking is always on. ClaudeDevs says Medium is the default effort setting, and users can adjust effort.
• Claude Code subscription limits are expanding. Pro, Max, and Team users get higher five-hour limits and a reset they can bank for later use.
• Long-running agents get a new compaction path. Anthropic's API can summarize older turns on the server while keeping recent turns in the active context.
• Safety routing remains part of the product. Anthropic says Opus 5.5 has Fable 5.1-class safeguards for cyber, biology, and frontier LLM development requests, with flagged requests falling back to another model.

Pricing and availability

Opus 5.5 is available on the Claude Platform and through Amazon Web Services, Google Cloud, and Microsoft Azure. The standard Claude Platform price is $4 per million input tokens, $20 per million output tokens, $0.20 per million cache-read tokens, and $5 per million cache-write tokens (Anthropic's pricing table).

That is a 20% reduction in input and output token prices versus Opus 5, and a 60% reduction in cache-read pricing. Anthropic says cache reads make up most of the cost of agentic and coding workloads, which is why the per-task savings can be larger than the headline token-price change (Anthropic's launch page).

Fast mode is an optional higher-speed path in Claude Code and the Claude Platform. It costs $8 per million input tokens and $40 per million output tokens, with Anthropic describing it as up to 2.5 times faster (Anthropic's launch page).

What changes in Claude Code

Claude Code uses Opus 5.5 as the default Opus model for paid plans, according to ClaudeDevs. The same announcement says Pro, Max, and Team users get a 20% increase in five-hour session limits, and that the lower model price makes those limits go 25% further. Subscribers also receive a rate-limit reset that can be saved and used at any time until October 22, 2026 (ClaudeDevs on X).

To migrate an API-backed Claude Code setup, use the new model identifier:

claude-opus-5-5

ClaudeDevs also says Claude Code users can run /claude-api migrate to switch to the new identifier. The practical migration is therefore small for users who let Claude Code manage model configuration, but teams with their own API wrappers should update the model name and re-run their cost and quality checks (ClaudeDevs on X).

Effort levels and always-on thinking

Thinking is always enabled for Opus 5.5. ClaudeDevs says the default is Medium and that users can adjust effort; the announcement also describes Fast mode as the option for users who want more speed (ClaudeDevs on X).

This makes effort an operational control rather than a choice between a thinking and non-thinking model. Start at the product default, then measure a representative workload before raising effort. Higher effort can increase quality on difficult work, but it also increases latency and token use. Anthropic's launch page reports its benchmark numbers at the effort level specified for each evaluation, so do not treat one benchmark setting as a universal recommendation (Anthropic's model results).

Context compaction for long-running agents

Opus 5.5 is also announced alongside Anthropic's server-side context compaction. Compaction is in beta and replaces older conversation turns with a summary written by Claude, keeping the active context smaller while preserving recent turns (Anthropic's compaction documentation).

There are two modes:

On-demand compaction can run in the background while work continues. The API also supports a custom summarization prompt and options for preserving recent turns. Threshold compaction runs inside the request that reaches the configured trigger. For Opus 5.5, Anthropic says thinking must be removed or dropped from turns that are re-inserted after threshold compaction (Anthropic's compaction documentation).

The implication for coding agents is straightforward: long sessions no longer have to choose between carrying every old turn forever and starting over. A team can let the server summarize earlier work, keep the active task and recent tool results intact, and use a custom summary prompt when a domain-specific detail must survive.

Safety and availability boundaries

ClaudeDevs describes Opus 5.5 as the first Opus model with safeguards in the class of Claude Fable 5.1 for cyber, biology, and frontier LLM development work. Flagged requests fall back to another model, and Anthropic says it is working to reduce incorrect flags (ClaudeDevs on X).

Anthropic's launch page says Opus 5.5 was evaluated with production safeguards enabled. It also describes a classifier that screens actions before they run, an open-source sandbox that security teams can audit, and stronger resistance to prompt injection than Opus 5 (Anthropic's safety notes).

That means the new model is not an unrestricted replacement for every security or biology workflow. Teams should expect some requests to route to another model and should keep human review around high-impact tool actions.

Who should upgrade

Upgrade first if you run long agentic coding sessions. The lower cache-read price, lower standard token price, faster output, and server-side compaction all target the costs that accumulate over many tool calls and turns.

Upgrade with an evaluation if you run a custom API client. Change the model identifier, check any assumptions about thinking blocks, and compare cost per completed task rather than cost per request. Keep a fallback path for requests that trigger safeguards.

Use Fast mode selectively. It is a useful latency lever for interactive work, but its token price is twice the standard Opus 5.5 rate. Reserve it for workflows where response time matters more than minimum spend.

Do not treat the headline benchmark numbers as a guarantee. Anthropic's results are vendor-reported and use different effort settings and safeguards by evaluation. Measure Opus 5.5 on your own repository, tools, prompts, and approval flow.

FAQ

The structured FAQ in this article's metadata covers the model identifier, pricing, availability, Claude Code limits, and migration command in short form.

Sources

• Anthropic, “Claude Opus 5.5” (launch page, pricing, availability, benchmarks, safety; checked September 22, 2026): https://anthropic.com/claude-opus-5-5
• ClaudeDevs, “Opus 5.5 performs at the level of Fable 5.1...” (Claude Code limits, default, and reset; September 22, 2026): https://x.com/ClaudeDevs/status/2102438800836489554
• ClaudeDevs, “On the Claude Platform, Opus 5.5 is $4/$20...” (platform pricing and migration command; September 22, 2026): https://x.com/ClaudeDevs/status/2102438805253091551
• ClaudeDevs, “Thinking is always on for Opus 5.5...” (effort, compaction, and Fast mode; September 22, 2026): https://x.com/ClaudeDevs/status/2102438806465245246
• ClaudeDevs, “Opus 5.5 is available now...” (availability and safeguards; September 22, 2026): https://x.com/ClaudeDevs/status/2102438808952467507
• Anthropic, “Compaction” (server-side context compaction; checked September 22, 2026): https://platform.claude.com/docs/build-with-claude/compaction