Claude Opus 5.5: API Pricing, Quota Resets, and Migration Checks

When did Opus, Sonnet, and Haiku 5.5 launch?

Haiku 5.5 launched on October 7, 2026, with API ID claude-haiku-5-5. Opus 5.5 launched September 22 and Sonnet 5.5 September 28. All three are now released rather than awaiting a future family launch.

Opus 5.5's official API ID is claude-opus-5-5. Its announcement covers Claude Platform and channels on AWS, Google Cloud, and Microsoft Azure. Verify organization, region, and model permissions on the platform you use.

The September 22 Opus announcement described Sonnet and Haiku 5.5 as future plans. Those are historical plans, not their current status. Separate Sonnet pricing and migration and Haiku pricing and 100K-threshold guides explain their respective costs and conditions.

Calculate API costs separately from mode, cache, and subscriptions

At the September 29, 2026 check, Opus Standard API pricing per million tokens was $4 input, $20 output, and $0.20 cache reads. Fast mode was $8 input and $40 output.

Fast remains a research preview. Its API is supported only on Anthropic's first-party platform, not Claude Platform on AWS, partner cloud platforms, or the Batch API.

Consult the relevant official prices for cache writes, cache duration, batch processing, and cloud surcharges. Cache-read pricing is not cache-write pricing. This revision updates family availability; recheck current prices before an actual call.

Editorial calculation, assuming Standard with no caching, additional tools, or platform fees: 10, 000 input and 2, 000 output tokens cost $0.08 in model charges. Multiply input and output quantities by their respective unit prices, then add them. This is a calculation, not a measured bill, speed test, or quality result.

Anthropic says typical workloads at default settings cost less than Opus 5. That comparison comes from vendor testing; this site has not measured it. Include retries, tools, human review, and rework when evaluating total task cost, rather than comparing token prices alone.

What do higher five-hour limits and resets mean?

Anthropic announced higher five-hour limits for Pro, Max, Team, and seat-based Enterprise, and bankable rate-limit resets for some users. The announcement does not list new numerical limits for every plan. This guide does not establish that an entitlement has arrived in every account.

Suggested check: confirm the account, workspace, and plan, then inspect the usage or limits area for remaining capacity, recovery time, reset entitlement, and expiry. Follow the displayed reset rules and check the same account afterward. If no reset entry appears, verify eligibility rather than repeatedly switching accounts.

Anthropic's rules and OpenAI's reset descriptions are separate. Codex users can consult the checklist for automatic resets, banked resets, and scheduled recovery. API billing, subscription resets, and service incidents are different evidence.

Verify each channel's model and permissions before migration

Claude Platform uses claude-opus-5-5. The official overview lists anthropic.claude-opus-5-5 for Amazon Bedrock, and claude-opus-5-5 for Google Cloud and Microsoft Foundry. Use your platform's current documentation and actual permissions; do not copy another platform's identifier format blindly.

The model overview lists a 1M-token context window, up to 128K output tokens, always-enabled adaptive thinking, and default medium effort. Maximum values are not guarantees for every request. Check SDK support, platform limits, and budgets together, including whether parameters from older examples remain supported.

If the model is missing or access is denied, first check its official ID, channel, region, and organization permissions. Changing IP does not grant model access. Repeatedly guessing names and calling them is not a permission check.

Migration: Compare representative tasks under the same conditions

The following is editorial guidance, not an Anthropic-mandated process. Step 1: save the current model, prompts, tool definitions, permissions, budget, and fallback configuration.

Step 2: select representative tasks whose results you can verify, such as fixing a known bug, answering questions against source material, or extracting a fixed format. Do not upload code or business data without authorization.

Step 3: keep inputs and acceptance criteria constant in an isolated environment, changing only the explicitly supported model configuration. Check factual citations, output fields, error handling, tool parameters, and actual actions. Use test data and confirmation steps for writes, external messages, or paid actions.

Step 4: record task completion, errors, token usage, time, and human rework, then compare the budget. Migrate gradually by task type only when comparable results support it. One successful example does not justify replacing every workflow.

When something fails, separate format, tools, and budget

If output structure is wrong, check model and SDK parameters against application parsing and acceptance criteria. If a tool is denied, check permissions and sandbox scope. Record request identifiers and times separately for 403 responses, exhausted quota, incorrect model IDs, and connection timeouts instead of labeling everything model instability.

Stop expanding usage and return to the verified configuration if tools act incorrectly, data boundaries are unclear, or costs exceed a stable budget. Keep reproducible, redacted failure examples for diagnosis. Mark causes as unknown when not reproduced or established.

Use a connection troubleshooting checklist only when evidence points to DNS, TLS, proxy 407, or connection timeouts. Subscription eligibility and model permissions cannot be resolved through network setup.

Comparing with GPT-6 Sol and Luna

When evaluating OpenAI models too, consult their pricing and limit guide and hold task, data, permissions, tools, and acceptance criteria constant. Record model price, tool calls, completion quality, and human effort separately.

This page provides no measured cross-vendor ranking. Vendor benchmarks, product positioning, and your actual workload address different questions. Base model choices on results you can verify.

What this guide will keep tracking

This page retains cost calculations, entitlement interpretation, and migration methods for the Opus announcement, now incorporating Haiku 5.5's October 7 release. Actual pricing, region and account access, and limit details may change; prioritize the facts that affect your choice.

The family has launched without requiring a single model for every task. Anthropic positions Opus and Sonnet 5.5 for more complex agentic coding and workflows, and Haiku 5.5 for summaries, context compaction, and narrower agent tasks. Compare completion quality and total cost on the same tasks before deciding what to migrate.