Every writeup of Claude Sonnet 5 back in June led with the benchmark table. Almost nobody put the other number on a calendar, and that number lands in 24 days.
The Sonnet 5 Introductory Price Expires, and the Cache Line Goes With It
Here is the part that got buried. Sonnet 5 has been running at introductory pricing since it shipped on June 30, and Anthropic’s own pricing page says that rate holds through August 31 only. Right now you pay $2 per million input tokens and $10 per million output. Starting September 1 it becomes $3 and $15.
Fifty percent is the easy headline. The line that actually decides your invoice is further down the table. Cache hits and refreshes go from $0.20 to $0.30 per million. Five minute cache writes go from $2.50 to $3.75, one hour writes from $4 to $6. Batch goes from $1 and $5 to $1.50 and $7.50. It is a clean 1.5x on every single row, which means there is no clever routing trick waiting for you. If you run a cache heavy agent loop, and anyone running long lived agents does, your cache read line is where most of your spend lives and it is going up by exactly as much as everything else.
Then there is the comparison nobody is making. After September 1, Sonnet 5 lists at $3 and $15, which is precisely what Sonnet 4.6 costs today. Same sticker. Not the same bill. That same pricing page notes that Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text, while Sonnet 4.6 and earlier use the old one. So the per token price matches and the token count does not. My read: anyone planning to “just stay on Sonnet” through the change should model it as more than a 50% increase on identical text, and the exact delta depends on your content shape.
None of this is a scandal. Anthropic published the end date the day it published the model, and the Sonnet 5 announcement is upfront about the introductory framing while claiming 46.8% on Humanity’s Last Exam with tools and 78.5% on OSWorld-Verified. The failure here is downstream. Coverage optimized for launch day and stopped, so the thing that actually hits operators showed up in nobody’s feed. Verdict: matters. Pull your last 30 days of usage this week and multiply the cache read line by 1.5 before you find out in October.
Anthropic Shipped a Hard Spend Cap the Same Month
Timing worth noticing. On August 7, per the Claude Platform release notes, Anthropic added budgets to Claude Managed Agents sessions. You set a hard cap on a session’s spend at public list rates, and when the session hits it, it pauses with a budget_reached stop reason instead of firing another model request. Change or remove the budget and it picks up where it stopped. Deployments pass the same budget down to every session they start.
I have wanted this specific shape for a year. Nearly every cost control in this category is an alert, which is a thing that tells you about the money after the money is gone. A stop reason fails closed. That is the difference between a monitoring feature and a control.
The same day brought three more: an advisor slot in the multiagent roster, so a session’s main thread can consult a model at least as capable as itself mid turn, an inference_geo setting on the agent’s model object for data residency, and automatic skill loading from a mounted GitHub repository’s root .claude/skills directory. Two costs to keep in your head while you play with those. Pinning inference to the US carries a 1.1x multiplier on every token category, and Managed Agents sessions bill $0.08 per session hour on top of tokens, metered only while the session is actually running. If you have been reading my August 7 roundup on the context window token meter, this is the same lesson arriving from the other direction: the config file is where the money is. Verdict: matters.
MCP Went Stateless on July 28, and “Nothing Breaks” Is Wrong
The migration guides that showed up this week keep repeating a comforting line: nothing breaks, deprecations carry a one year grace period. Half right, and the wrong half is the half that will page you.
The 2026-07-28 changelog separates two states, and only one of them gets the grace period. Deprecated means Roots, Sampling and Logging, plus the HTTP and SSE transport and OAuth 2.0 Dynamic Client Registration. Those keep working through a minimum twelve month window under the new feature lifecycle policy. Fine.
Removed is a different word. Protocol level sessions are gone, and so is the Mcp-Session-Id header. The initialize and notifications/initialized handshake is gone; every request now carries its protocol version and client capabilities in _meta. The HTTP GET endpoint and resources/subscribe are replaced by a single subscriptions/listen stream. ping, logging/setLevel and tasks/list are gone outright. SSE stream resumability is gone too, which means a broken response stream loses the in flight request and the client has to reissue it with a fresh request ID rather than replaying from Last-Event-ID.
The honest version of the reassurance is this: version negotiation protects you, not the grace period. A client speaking 2025-11-25 to a server that still answers 2025-11-25 keeps working indefinitely. The morning your server adopts the new revision and drops the old one, every one of those removals lands at once, and if your retry logic assumed stream resumability you will find out during an incident rather than during a migration.
Worth saying what you get for it. A remote server that needed sticky sessions, a shared session store and a gateway smart enough to inspect them can now sit behind plain round robin and run on serverless. That is a real reduction in operational surface, and it is why this revision is worth doing on purpose instead of by accident. Verdict: breaks your stack, and it is the good kind of breakage.
Two Anthropic Dates Already on the Board
Short section, because these are calendar items rather than arguments. Anthropic retired Claude Opus 4.1 on August 5, and requests to claude-opus-4-1-20250805 now return an error on the first party API, though the pricing table shows it still alive on Bedrock and Google Cloud. And the legacy Workbench loses access on August 17, which is nine days from today. Saved prompts, variables and evals do not carry into the replacement playground, so export anything you want to keep before then. Verdict: breaks your stack.
Google Moved Hassabis Sideways and Lost Jeff Dean
Sundar Pichai’s own post confirms it. Demis Hassabis becomes Chair of Google DeepMind and Chief Scientist of Alphabet, reporting to Pichai on strategy and continuing to run Isomorphic Labs. Koray Kavukcuoglu becomes SVP of Google DeepMind with Gemini model development, Frontier AI research and the Gemini app under him, reporting straight to Pichai. Time framed it as a reshuffle aimed at closing the gap with OpenAI and Anthropic.
The reshuffle is the headline and the headline is mostly noise for anyone shipping. No endpoint changed. No model got better or worse on August 6 because a reporting line moved. Org charts are the cheapest possible story to write and the least useful one to act on. Verdict: marketing.
The part I would actually watch is in the same announcement and got a fraction of the coverage. Jeff Dean is leaving Google after 27 years to start an independent public benefit corporation with Sanjay Ghemawat, with Google as a founding investor and Cloud partner. Those two built the infrastructure the modern stack is a descendant of. Two people with that track record starting something together, funded by the company they left, is a more interesting signal than any title change, and it will take a year to read.
The One Thing I Would Do Today
Open your billing dashboard, find your cache read line, and multiply it by 1.5. That is the whole assignment. The MCP work can wait for a planned migration window and the Google story can wait forever, but the Sonnet 5 change is dated, unavoidable, and arrives whether or not anyone told you it was coming. Everything on this site is built with these models, which is exactly why I read the pricing table before I read the benchmark chart.