API Rate Limits Rise; Usage Tiers Collapse to Start, Build, and Scale
Anthropic raised rate limits across the Claude API and simplified how they’re structured. Claude Sonnet and Claude Haiku limits now match Claude Opus at every usage tier, and the old ladder of tiers has been consolidated into just three: Start, Build, and Scale. Most organizations move up a rung, no one gets lower limits than before, and no action is required — you can check your tier and current limits in the Claude Console.
It’s a quiet but meaningful change. Flattening Sonnet and Haiku up to Opus-level throughput removes a real friction point for teams that route cheap, high-volume work through the smaller models, and squashing the tier ladder makes capacity planning something you can explain in one sentence instead of a spreadsheet.
Mark Your Calendar: Opus 4.1 Retires Aug 5, Opus 4.7 Fast Mode Removed July 24
Two deprecation deadlines are now firmly on the clock. Claude Opus 4.1 (claude-opus-4-1) is scheduled to retire from the Claude API on August 5, 2026, after which requests return an error — Anthropic recommends migrating to Opus 4.8. Sooner than that, fast mode for Opus 4.7 is removed on July 24; after that date, calls to claude-opus-4-7 with speed: "fast" will error, and the path forward is fast mode on Opus 4.8.
If you have anything pinned to an older Opus, this is the fortnight to test the swap. The pattern is consistent across Anthropic’s recent housekeeping: retire the trailing edge, point everyone at the current flagship, and keep the deprecation calendar public enough that nobody gets surprised in production.
Claude Code Flips Its Default Permission Mode to “Manual”
Claude Code 2.1.200 changes a foundational default: the “default” permission mode is now “Manual” across the CLI, --help, VS Code, and JetBrains (--permission-mode manual and "defaultMode": "manual" are accepted alongside default). In the same release, AskUserQuestion dialogs no longer auto-continue by default — you opt into an idle timeout via /config if you want the old behavior. The follow-up 2.1.201 also stopped Sonnet 5 sessions from using the mid-conversation system role for harness reminders.
Both changes lean the same direction: less silent autonomy, more explicit consent. As background agents and long-running sessions become normal, a tool that pauses for a human by default — rather than pressing on — is the safer footing for teams handing Claude real repository access. Rounding out 2.1.200 is a long list of reliability fixes across background sessions, plugins, screen-reader output, and terminal rendering.
Claude Managed Agents Add Event Deltas, Per-Session Overrides, and Lifecycle Webhooks
The Managed Agents platform picked up a cluster of upgrades for people running agents in production. Session event streams now support event deltas — opt in with event_deltas[] and you get event_start and event_delta events that preview an agent’s message text as it’s generated, before the full agent.message lands. Listing sessions gains backward pagination via a prev_page cursor, and you can now pass agent with type: "agent_with_overrides" to swap the model, system prompt, tools, MCP servers, or skills for a single session without touching the agent itself.
Two more for the security-minded: vault environment-variable credentials add an injection_location setting that controls whether a secret is substituted into outbound headers, the request body, or both; and webhooks now span the agent, deployment, and deployment-run lifecycle, so you can react to a newly published version, a paused deployment, or a failed scheduled run without polling.
Heads Up: Sonnet 5’s New Tokenizer Bills ~30% More Tokens
Now that Claude Sonnet 5 is the default across Free, Pro, and Claude Code, one detail is worth putting on the radar before it dents a bill: Sonnet 5 uses a new tokenizer that produces roughly 30% more tokens for the same text (the exact figure depends on your content and workload). The headline pricing looks cheaper at $2/$10 per MTok through August 31, but token counts move too — so measure both sides before assuming a net win.
A few other migration gotchas ride along: adaptive thinking is on by default, manual extended-thinking budgets now return a 400 error, and setting temperature, top_p, or top_k to non-default values also errors. If you’re moving prompts over from Sonnet 4.6, run the token-counting API against claude-sonnet-5 and re-baseline your budgets rather than eyeballing it.
Launch Season Is Over. This Is the Bolt-Tightening Phase.
Strip away the model launches of the last few weeks and look at what shipped this cycle: a safer default permission mode in Claude Code, higher and simpler rate limits on the API, a published deprecation calendar for older Opus builds, and finer controls for the agents people are actually deploying — per-session overrides, credential-injection rules, lifecycle webhooks. None of it is flashy. All of it is what you build when your users have stopped kicking the tires and started running Claude in production.
That’s the tell. The flip to Manual permissions in particular reads less like a UX tweak and more like a statement of posture: as autonomy scales, the default should be a pause, not a plunge. Pair that with the token-count reality under Sonnet 5’s new tokenizer, and the message to builders this week is refreshingly grown-up — the models are getting cheaper and more capable, but you still have to read the meter, test your migrations, and decide, deliberately, how much rope to hand your agents.