Everything We Know About Claude Code Limits
Last updated - 29 July 2025
TL;DR
- For Claude.ai Users:
The core limit is a 5-hour rolling session that begins with your first prompt. New weekly quotas will be added starting Aug 28th for heavy users on Claude Pro and Claude Max plans. (all Claude.ai plans share a common usage bucket across Claude app & Claude code; Max plan just multiplies the 5-hour allowance) - For API Users:
API usage is separate and billed pay-as-you-go per token. You get the same rates ($15 in / $75 out per million tokens for Opus 4) whether you use Anthropic, AWS Bedrock, or Google Vertex AI.
You can choose whichever region, VPC, or IAM model best fits your stack. - For Portkey Users:
With Portkey, you can route Claude Code through Bedrock, Vertex, or Anthropic seamlessly and pull in maximum capacity from multiple sources. Using Portkey will also let you centralize cost controls, add security guardrails, and get audit logs automatically.
Understanding the 5‑Hour Session Model
The moment you run claude in your terminal, a 5‑hour rolling window begins. All messages and token spend during that period draw from your plan’s pool; the clock resets only when you send the next message after the 5 hours lapse. Pro users average 10‑40 prompts per window, while Max 20× users can push 200‑800 prompts depending on code size and model choice. (Source: Anthropic Help Center)
💡 Tips
- Time your first prompt so the window spans your peak coding block; fire a dummy ping at 13:01 to start a fresh window for the afternoon.
- Batch edits: one long diff request burns fewer tokens than several “please refine” follow‑ups.
Plan Pricing & Capacities
| Plan | Monthly fee | Relative capacity | Typical Claude Code prompts / 5 h | Model access |
|---|---|---|---|---|
| Free | $0 | ≈ ⅒ Pro | 2‑5 (Sonnet 4 only) | Sonnet 4 |
| Pro | $20 | baseline | 10‑40 | Sonnet 4 |
| Max 5× | $100 | 5 × Pro | 50‑200 | Sonnet 4 + Opus 4 |
| Max 20× | $200 | 20 × Pro | 200‑800 | Sonnet 4 + Opus 4 |
Weekly limit kicker (28 Aug 2025): Pro ≈ 40‑80 Claude Code hours; Max tiers scale proportionally. The 5‑hour windows stay—weekly just adds a ceiling.
Token‑Based API Pricing (Console, Bedrock, Vertex)
| Model | Anthropic API | AWS Bedrock* | Google Vertex AI* |
|---|---|---|---|
| Opus 4 | $15 in / $75 out per M tokens | same list price | same list price |
| Sonnet 4 | $3 in / $15 out | same list price | same list price |
| 3.7 Sonnet | $3 / $15 | same | same |
| 3.5 Haiku | $0.80 / $4 | same | same |
* Bedrock and Vertex bill in the currency of your account and add standard cloud charges (data egress, networking, etc.). Region selection, IAM policies, and VPC endpoints are the usual enterprise reasons to pick one over the other. Anthropic confirms pricing parity across channels.
Which Endpoint Should You Use?
| Need | Best Option | Why |
|---|---|---|
| Lowest latency to US‑East | AWS Bedrock (us‑east‑1) | Dedicated Anthropic capacity inside AWS infra |
| Private networking / VPC‑SC | Vertex AI Private Service Connect | Keeps traffic off the public internet |
| Global multi‑region fail‑over | Portkey Gateway | Abstracts region + provider routing |
| Advanced cost controls | Portkey budgets + prompt caching | Enforce spend limits, auto‑downgrade models |
Portkey: One Switch for Anthropic, Bedrock, or Vertex
Direct Claude Code is great for solo hacking, but enterprises need:
- Unified gateway routing – point Claude Code to Portkey once; flip between Anthropic, Bedrock, or Vertex by config.
- Live spend & usage dashboards – 40‑plus metrics, cost attribution by team, and budget alerts.
- RBAC + guardrails – per‑role rate limits, banned‑pattern filters, and partner guardrails from Palo Alto, Prompt Security, Pangea, etc.
- Audit‑ready logs – every prompt, response, and error logged with metadata for compliance.
Bottom line: Portkey turns Claude Code into an enterprise‑grade service without forcing you to rebuild scripts or choose a single cloud. Drop your API key(s) into Portkey, set a config, and your devs keep typing claude …—now with observability, governance, and cloud‑agnostic resilience baked in.
Anthropic may refine quotas again once weekly limits ship, and Bedrock/Vertex regions keep expanding. Bookmark this post—we’ll update as soon as the numbers move.
