Claude Opus 5.5 Fast Mode Leaves Preview: Setup and Pricing
Anthropic made fast mode generally available for Claude Opus 5.5 on 2026-10-15. Learn how to configure requests, manage rate limits, and calculate costs.
- Fast mode for Claude Opus 5.5 is out of research preview as of 2026-10-15, eliminating waitlist approvals and the required beta header.
- The inference configuration yields up to 2.5x higher output tokens per second at $8 USD / MTok input and $40 USD / MTok output.
- Dedicated fast mode rate limits are automatically assigned to accounts in the Start, Build, and Scale usage tiers.
- Claude Opus 5 and Claude Opus 4.8 remain in research preview and still require the fast-mode-2026-02-01 header.
- A parallel Compliance API beta endpoint now exposes member conversation preferences across Claude Enterprise workspaces.
What Changed on 2026-10-15 for Claude Opus 5.5
On 2026-10-15, Anthropic announced in its platform release notes that fast mode graduated out of research preview for Claude Opus 5.5 (claude-opus-5-5). Developers no longer need to submit access requests or wait for manual organization approvals to run Opus 5.5 at accelerated speeds. Along with general availability, Anthropic removed the requirement to send the fast-mode-2026-02-01 beta header when invoking Claude Opus 5.5. Applications that still transmit this beta header will continue to function normally without generating errors.
The release also provisions default fast mode rate limits directly to organizations in the Start, Build, and Scale usage tiers on Claude Console. In addition, Claude Opus 5.5 fast mode is now enabled on Claude in Microsoft Foundry for deployments hosted on Azure. However, fast mode is not available for Microsoft Foundry deployments hosted on Anthropic. Earlier Opus models—specifically Claude Opus 5 and Claude Opus 4.8—remain in research preview and still require an active waitlist approval and the fast-mode-2026-02-01 header.
Separately in the same 2026-10-15 platform update, Anthropic expanded the Compliance API with a beta endpoint for Claude Enterprise organizations. Security administrators can now query GET /v1/compliance/organizations/{org_uuid}/user_preferences using a Compliance Access Key and the read:compliance_user_data scope to audit user-saved Instructions for Claude from claude.ai and Global instructions from Cowork. Further context on enterprise administrative tooling appears in the review of Anthropic Compliance API unified chats and export.
How Fast Mode Works: Throughput, Weights, and Capabilities
Fast mode is an alternative inference execution configuration designed to accelerate text generation. When enabled, Claude Opus 5.5 delivers up to 2.5x higher output tokens per second (OTPS) compared to standard execution speeds. It is important to note that this boost is specifically concentrated on output generation throughput rather than time to first token (TTFT). Workloads with large system instructions and minimal output will not see latency reductions at the initial token arrival stage, but pipelines that emit lengthy responses will observe substantially shorter completion times.
The operational behavior, intelligence, context understanding, and model weights of Claude Opus 5.5 remain identical between fast mode and standard mode. The model does not switch to a distilled checkpoint, an alternative quantized branch, or a smaller architecture. Because the underlying model output behavior is unchanged, fast mode works naturally alongside streaming, where real-time token delivery displays the 2.5x throughput improvement most noticeably to end users.
The following table details model support status, required headers, and target environments across the platform:
| Model | Status | Beta Header Required | Supported Environments |
|---|---|---|---|
| Claude Opus 5.5 | General Availability | None (optional backward compatibility) | Claude API, Claude Managed Agents, Azure-hosted Microsoft Foundry |
| Claude Opus 5 | Research Preview | fast-mode-2026-02-01 | Claude API, Claude Managed Agents |
| Claude Opus 4.8 | Research Preview | fast-mode-2026-02-01 | Claude API, Claude Managed Agents |
| Claude Opus 4.6 | Unsupported | Header ignored (falls back to standard speed) | Claude API |
Implementation: Setting Up API Requests and Checking Speed
To request fast mode with Claude Opus 5.5, add the speed: "fast" parameter to the request payload sent to /v1/messages. You no longer need to pass anthropic-beta: fast-mode-2026-02-01. Maintain your standard anthropic-version: 2023-06-01 header and supply your API key as usual. The request accepts standard parameters including max_tokens and structured message payloads.
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-opus-5-5",
"max_tokens": 4096,
"speed": "fast",
"messages": [{
"role": "user",
"content": "Refactor this module to use dependency injection"
}]
}'
In Python, initialize the standard Anthropic SDK and supply the speed parameter directly during message creation:
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-opus-5-5",
max_tokens=4096,
speed="fast",
messages=[
{"role": "user", "content": "Refactor this module to use dependency injection"}
],
)
print(response.usage.speed)
Every successful completion provides confirmation in the usage block of the API response. The usage.speed field indicates whether the execution was processed as "fast" or "standard". If an application submits speed: "fast" against an unsupported model without preview access, the API raises an error. Claude Opus 4.6 is a specific exception: if passed speed: "fast" with the preview header, it will not return an error, but it executes at standard speed and sets usage.speed to "standard".
Fast Mode Pricing and Token Economics
Fast mode charges a premium rate across all processed tokens, applied evenly across the full context window including prompts exceeding 200k input tokens. For Claude Opus 5.5, fast mode is billed at $8 USD / MTok for input tokens and $40 USD / MTok for output tokens. By comparison, standard Opus 5.5 pricing stands at $4 per 1M input tokens and $20 per 1M output tokens, effectively doubling the baseline unit cost to achieve the 2.5x output generation speed.
For comparison, research preview models Claude Opus 5 and Claude Opus 4.8 cost $10 USD / MTok for input and $50 USD / MTok for output when operated in fast mode. You can evaluate model rates and operational budget boundaries by consulting the /tools/llm-api-cost-calculator/ and the comprehensive pricing table at /tools/llm-api-pricing/.
| Model | Standard Input (per 1M) | Standard Output (per 1M) | Fast Mode Input (per 1M) | Fast Mode Output (per 1M) |
|---|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | $8 USD / MTok | $40 USD / MTok |
| Claude Opus 5 | $4 | $20 | $10 USD / MTok | $50 USD / MTok |
| Claude Opus 4.8 | $4 | $20 | $10 USD / MTok | $50 USD / MTok |
Fast mode pricing stacks multiplicatively with secondary pricing modifiers. If your architecture relies on prompt caching, the platform prompt caching multipliers apply directly on top of the fast mode rates. Likewise, data residency compliance multipliers stack on top of fast mode pricing. Account administrators can track token expenditure and breakdown totals via the Usage and Cost API.
Rate Limits, Headers, and Fallback Handling
Fast mode does not consume tokens from your standard Claude Opus rate limits. Instead, Anthropic allocates a dedicated fast mode token budget that operates independently. When fast mode capacity is exhausted or rate limits are reached, the API returns an HTTP 429 status code containing a retry-after header. The status of your fast mode allocation is tracked across dedicated response headers on every call:
anthropic-fast-input-tokens-limit: Maximum allowed fast mode input tokens per minute.anthropic-fast-input-tokens-remaining: The current pool of available fast mode input tokens.anthropic-fast-input-tokens-reset: Timestamp indicating when the fast input token limit replenishes.anthropic-fast-output-tokens-limit: Maximum allowed fast mode output tokens per minute.anthropic-fast-output-tokens-remaining: The remaining count of fast mode output tokens.anthropic-fast-output-tokens-reset: Timestamp indicating when the fast output token limit replenishes.
Official Anthropic SDKs automatically retry rate-limited calls up to 2 times by default, waiting for the duration specified by retry-after. Because fast mode operates with continuous token replenishment, delay intervals are generally brief. For more details on usage limits across providers, review rate limits and usage tiers at Anthropic, OpenAI, and Google.
If an interactive application cannot pause for retries, you can fall back to standard speed. To execute immediate fallback, configure client retries to 0 for the fast attempt, intercept the 429 exception, and re-issue the request omitting speed: "fast". Be aware that fallback requests will experience a prompt cache miss: standard speed and fast mode requests run on distinct execution infrastructure and do not share cached prefix memory.
def create_message_with_fast_fallback(max_retries=0, max_attempts=3, **params):
try:
return client.with_options(max_retries=max_retries).messages.create(**params)
except anthropic.RateLimitError:
if params.get("speed") == "fast":
del params["speed"]
return create_message_with_fast_fallback(max_retries=max_retries, **params)
raise
except (anthropic.APIStatusError, anthropic.APIConnectionError) as error:
if isinstance(error, anthropic.APIStatusError) and error.status_code < 500:
raise
if max_attempts > 1:
return create_message_with_fast_fallback(
max_retries=max_retries, max_attempts=max_attempts - 1, **params
)
raise
Migrating to Opus 5.5 Fast Mode: Verification Checklist
Before migrating high-volume production traffic to Claude Opus 5.5 fast mode, verify each operational and architectural dependency:
- Confirm Claude Console tier status: Ensure your organization is active within the Start, Build, or Scale usage tier. Accounts in the Evaluation tier must upgrade before fast mode limits activate.
- Remove header dependencies: You can safely remove the
fast-mode-2026-02-01beta header from your Claude Opus 5.5 code paths, or leave it in place during rolling deploys. - Update payload speed parameters: Explicitly pass
speed: "fast"in request bodies or SDK invocations where low output generation latency is required. - Inspect response usage metadata: Validate that
usage.speedreturns"fast"in integration tests rather than"standard". - Monitor dedicated rate limits: Track
anthropic-fast-input-tokens-limitandanthropic-fast-output-tokens-limitin logging systems to anticipate capacity thresholds. - Configure fallback and retry handling: Implement fallback logic to standard speed or tune client SDK retry parameters to absorb transient 429 and 529 error codes.
- Audit cloud hosting endpoints: If using Claude in Microsoft Foundry, verify that deployments are hosted on Azure infrastructure, as Anthropic-hosted Foundry instances do not support fast mode.
- Account for budget changes: Fast mode increases Opus 5.5 token billing to $8 USD / MTok input and $40 USD / MTok output across all prompt lengths.