Claude API July 2026: Rate Limits Up, Keys Expire, Act by July 24

Anthropic shipped five developer platform changes in June and July 2026. Four are upgrades with no downside. One has a hard deadline: if you are running Claude Opus 4.7 with speed: "fast" in production, your calls will start throwing errors on July 24. That is 11 days from now.

Opus 4.7 Fast Mode Is Gone on July 24

Fast mode on Claude Opus 4.7 was deprecated June 25 and is removed on July 24. After that date, any request to claude-opus-4-7 with speed: "fast" returns an error. Unlike Opus 4.6, there is no silent fallback to standard speed — it fails hard.

The fix is a one-line change:

# Before — breaks July 24
client.messages.create(
    model="claude-opus-4-7",
    speed="fast",
    messages=[...]
)

# After — switch to Opus 4.8
client.messages.create(
    model="claude-opus-4-8",
    speed="fast",
    messages=[...]
)

The model itself — Opus 4.7 without fast mode — stays available. If you do not use speed: "fast", nothing changes. The migration to Opus 4.8 fast mode also brings a cost drop: $10/$50 per million input/output tokens versus $30/$150 on Opus 4.7. Anthropic is paying you to upgrade.

Rate Limits Are Up Across the Board

Anthropic consolidated four usage tiers into three — Start, Build, and Scale — and raised limits at every level. The more significant change is model parity: Claude Sonnet and Haiku now carry the same RPM, ITPM, and OTPM limits as Opus at each tier. Previously, a high-volume Sonnet workload hit rate limits that Opus users never saw. That distinction is gone.

Monthly spend caps are $500 for Start, $1,000 for Build, and $200,000 for Scale. Organizations above Scale can contact Anthropic for a custom tier. No action is required — no organization received lower limits, and most moved up a tier. You can confirm your current tier in the Claude Console.

One detail worth noting: cache-read tokens no longer count toward your ITPM on any current model. Cached context is effectively free from a rate-limit standpoint, which matters for long-context agentic workflows.

API Keys Can Now Expire

You can now set an expiration when creating an API key or Admin API key in the Claude Console. Options are preset durations (3 hours, 1 day, 7 days, 30 days), a custom duration, or Never for keys stored in a secrets manager with their own rotation.

• Expiration is set at creation and cannot be changed afterward.
• Anthropic emails the key’s creator 7 days before expiry (for keys with at least a 14-day lifetime) and 1 day before (for keys with at least a 7-day lifetime).
• After a key expires, requests return a 401 authentication_error. Expired keys cannot be reactivated.
• If your organization sets a maximum expiration policy, the Never option is unavailable.

The Admin API now includes an expires_at field on the List API Keys and Get API Key endpoints, making automated key-rotation audits straightforward. Existing keys are unaffected.

Two Small Wins Worth Knowing

Refusals are no longer billed. If a request ends with stop_reason: "refusal" and produces no output, you are not charged — effective June 2, no opt-in required. For agentic workloads with heavy safety guardrails, this quietly eliminates the cost of inputs that produced nothing.

The advisor tool now accepts max_tokens. You can cap the advisor model’s output per call, cutting both latency and token cost. The advisor sub-inference does not stream, so an uncapped advisor run adds hidden time to every response. For workloads that only need brief guidance, setting this to a lower value is now trivial.

REPL State Persistence for Code Execution

All major SDKs — Python, TypeScript, Go, Java, Ruby, PHP, and C# — now support code_execution_20260120, the updated code execution tool that adds REPL state persistence. Variables and state persist between execution calls within the same conversation, making multi-step data workflows practical without re-establishing context on every turn.

This version is also the minimum required for programmatic tool calling, which lets Claude invoke other tools from within sandbox code rather than through individual API round-trips. Set the tool type to code_execution_20260120 — no beta header required. Compatible models include Fable 5, Mythos 5, Opus 4.5+, and Sonnet 4.5+.

What to Do This Week

• Audit now: Search your codebase for claude-opus-4-7 combined with speed or fast. If you find it, change the model ID to claude-opus-4-8 before July 24.
• Check your tier: Open the Claude Console and confirm your new rate-limit tier (Start, Build, or Scale) for capacity planning.
• Add key expiration: For new keys going forward, set an expiration aligned with your rotation policy.
• Update code execution: Bump to code_execution_20260120 to unlock REPL persistence and programmatic calling.

Full Anthropic release notes are at platform.claude.com. The July 24 deadline is the only item that requires urgent action — the rest can go into your next sprint.