Claude Fast mode just left research preview — your agents can now run up to 2.5x faster with zero paperwork. On October 15, 2026, the Claude Platform release notes confirmed Fast mode is generally available for Claude Opus 5.5: no access request, no beta header, and a default fast-mode rate limit in the Start, Build, and Scale tiers.
Tucked into the same release notes is a second launch aimed squarely at regulated industries. The Compliance API now returns the text each member saved as Instructions for Claude in claude.ai and Global instructions in Cowork — in beta for Claude Enterprise organizations. Two launches, one theme: Anthropic is making its flagship model faster to run and easier to audit.
What changed on October 15
Three changes. First, no more beta header for Opus 5.5 — set speed: "fast" and you’re in (old requests with the header keep working). Second, Start/Build/Scale tiers get a default fast-mode rate limit; exceed it and the API returns 429 with retry-after while the SDK retries twice.
Third, Fast mode now runs on Claude in Microsoft Foundry for Azure-hosted deployments. The carve-outs: Opus 5 and Opus 4.8 stay in research preview (keep the header there), and Fast mode still doesn’t work with the Batch API, Priority Tier, or Claude Platform on AWS.
The numbers: Fast mode pricing vs standard
Fast mode costs exactly 2x the standard rate — a premium multiplier, not a separate model:
| Rate per 1M tokens | Standard | Fast mode |
|---|---|---|
| Opus 5.5 input | $4 | $8 |
| Opus 5.5 output | $20 | $40 |
| Opus 5 / 4.8 input | $5 | $10 (preview) |
| Opus 5 / 4.8 output | $25 | $50 (preview) |
| Max output speed | 1x | Up to 2.5x tokens/sec |
Source: official Fast mode docs; caching and residency multipliers stack on top.
Stat highlights from Anthropic’s official pages:
- 2.5x — maximum output-tokens-per-second gain in Fast mode (Fast mode docs)
- 66.4% — Opus 5.5 score on Terminal-Bench 4.0 agentic coding, topping GPT-6 Astra’s 57.9% (Opus 5.5 announcement)
- Under 3 hours — to audit and fix a 200,000-line codebase; Opus 5 took 20+ hours and 2.5x the tokens (Opus 5.5 announcement)
- 72% vs 56% — known bugs caught in code review at low effort vs Opus 5 at high effort, in Deloitte testing (Opus 5.5 announcement)
How Fast mode actually works
Fast mode is the same model with a faster inference configuration — same weights, same intelligence, more output tokens per second. Gains concentrate on OTPS, not time to first token: it shines on long generations and streaming.
Two gotchas: switching speeds invalidates the prompt cache (fast and standard don’t share prefixes), and every response reports its speed in usage.speed — log it if you bill per token.
Fast mode vs Standard: which should you pick?
Fast mode
- Up to 2.5x output tokens/sec
- 2x token price ($8/$40 on Opus 5.5)
- Best for: chatbots, streaming UIs, interactive agents
- Needs: Start/Build/Scale tier (default limit)
Standard
- Baseline speed, lowest token price ($4/$20)
- Best for: batch jobs, background agents, evals
- Works with: Batch API (50% off), Priority Tier
- Keep here when latency doesn’t matter
The Compliance API’s new auditing power
The Compliance API is Anthropic’s audit interface for Enterprise admins. The October 15 addition — GET /v1/compliance/organizations/{org_uuid}/user_preferences — returns the custom instructions each member saved.
Two fields per member: conversation_preferences (claude.ai) and cowork_global_instructions (Cowork) — both account-level, so the same text appears under every org the person belongs to. Anthropic warns a key with read:compliance_user_data can read every chat, file, project, and transcript in every linked org: treat it like a production credential.
How to profit from both launches
Speed you don’t have to ask for, and audit trails you can sell — here is the playbook:
- Flip latency-sensitive workloads to
speed: "fast"this week — chatbots and code assistants first. Measure tokens/sec before and after viausage.speed. - Re-price on cost per finished task. Fast mode doubles token price but finishes up to 2.5x faster — for interactive products the experience win usually beats the premium.
- Keep fast and standard traffic on separate cache prefixes. Mixing speeds kills prompt caching — and cache reads are your cheapest tokens at $0.20 per million.
- Productize compliance for regulated clients. Banks, insurers, and healthcare orgs already pay for audit trails — the
user_preferencesendpoint is a new artifact for quarterly AI-governance reviews on retainer. - Lock down Compliance Access Keys from day one — created by primary or org owners in claude.ai. Scope narrowly, rotate regularly, never ship one client-side.
FAQ
Do I still need to request access to Fast mode?
Not for Opus 5.5 — no request, optional header, default rate limit in Start/Build/Scale tiers. Opus 5 and 4.8 still need the header and remain in preview.
Does Fast mode make Claude smarter?
No — same weights, faster inference config. The gain is output tokens per second (up to 2.5x), most visible when streaming. Time to first token barely changes.
What does Fast mode cost?
$8/$40 per million tokens on Opus 5.5 — exactly 2x standard. Caching and residency multipliers stack on top.
Can I combine Fast mode with the Batch API?
No — not with the Batch API, Priority Tier, Claude Platform on AWS, or Anthropic-hosted Foundry. Pick speed or the discount, not both.
What does the new Compliance API endpoint return?
Each member’s saved Instructions for Claude and Cowork Global instructions — the custom instructions people type once and forget, now auditable via the new endpoint.
Is the Compliance API feature fully launched?
Beta for Claude Enterprise orgs. You need a Compliance Access Key with the read:compliance_user_data scope — treat it like a production database credential.
Sources
- Claude Platform release notes — October 15, 2026 entry (official)
- Fast mode documentation — pricing, rate limits, considerations (official)
- Introducing Claude Opus 5.5 — benchmarks and pricing (official announcement)