MCP Publishes a New Roadmap, and It Reads Like an Enterprise Punch List
The Model Context Protocol core maintainers published an updated roadmap this morning, covering the next specification release and beyond. It is organized into five priority areas: agentic messaging primitives, HTTP-native transport unification and hardening, agent identity and enterprise-ready security, improved primitives, and SDK developer experience. Several items graduated from “on the horizon” in the last roadmap — server-initiated events, result type improvements, and agent identity — into full priorities.
Two of the five are worth reading closely. Agent identity takes direct aim at the fact that MCP authorization today assumes a human clicking approve in a browser, which is not what is happening when a cloud agent calls a server on behalf of an absent user or delegates narrower authority to a sub-agent. The fix being pursued is DPoP, Workload Identity Federation, the ID-JAG grant behind Enterprise-Managed Authorization, and standard token exchange — existing standards, explicitly “rather than pasted API keys and long-lived tokens.” The other is progressive discovery: connect to a server with a hundred tools today and the model pays for that entire surface before the user asks anything, and selection quality degrades as the list grows. The proposal is a small entry point that reveals more of the catalog as the conversation narrows. Also flagged: a tools/call response can carry the same output in more than one form, and server authors have no way to know which form a given client shows the model. That gets standardized to one contract.
The governance note matters too. SEPs inside these five areas get expedited review; proposals outside them are not auto-rejected but go to the back of a scarce queue. If you are planning a protocol contribution, the roadmap just told you where to aim it.
Ode with Anthropic Buys Casper Studios, One Month In
Ode with Anthropic — the enterprise AI services venture Anthropic launched in July with Blackstone, Hellman & Friedman and Goldman Sachs — announced its first acquisition yesterday: Casper Studios, a US AI product agency founded four years ago. Casper is small, roughly a dozen technical consultants by some counts and around 41 employees by others, and it was a Select tier partner in the Claude Partner Network. Its stated client list includes Netflix, Pepsi, private equity firms and hedge funds.
What Casper actually does is the tell: it builds the skills, connectors and context that put Claude to work inside the systems enterprise teams already use. That is not model work. That is integration labor — the unglamorous middle layer between a capable model and a company that cannot use it. Buying a services shop one month after standing up a services venture says Anthropic thinks the bottleneck on enterprise revenue is deployment capacity, not capability. Forbes framed it as the AI services race heating up, and that is roughly right: the competition is moving from whose model is better to who can actually land it in a Fortune 500 workflow.
Citigroup Joins the Top Bank Lineup as a Filing Nears
Following yesterday’s reporting that Anthropic expects its IPO to match or top SpaceX’s record, the bank lineup firmed up: Citigroup is set to join Morgan Stanley, Goldman Sachs and JPMorgan Chase in the top ranks of advisers. Anthropic is weighing a public filing as soon as the end of this month, with a market debut potentially in October and investors close to the company anticipating a valuation north of $2 trillion.
The detail buried under the underwriter news is the more interesting one. Citigroup was part of the lending group behind Anthropic’s $2.5 billion credit facility, which is reportedly being expanded to more than $10 billion. A four-fold expansion of a revolver in the run-up to a listing is not decoration — it is the balance-sheet cushion a company builds when it plans to keep spending hard through the transition to public markets. For context on the trajectory: the last private mark was $965 billion in May after a $65 billion Series H. As always, this is sourced to people familiar with the matter, not to Anthropic.
Python SDK v1.0: httpx2, Python 3.10+, and a Pile of Removals
The Anthropic Python SDK shipped v1.0, and this one needs a real migration pass rather than a version bump. The HTTP layer moves from httpx to httpx2, a maintained API-compatible fork. Practically: build your custom http_client, Timeout and transport objects from httpx2 (the DefaultHttpxClient helpers are unchanged), and if you rely on tracing or mocking libraries that patch httpx, call httpx2.alias_httpx() at startup.
The removals are where builds break. Python 3.10 is now the floor. The legacy Text Completions API is gone. So are temperature, top_p and top_k on Messages methods — if those are hardcoded in your inference wrapper, that is a hard failure, not a warning. The tool runner’s client-side compaction_control is out. On the async client, .with_raw_response results now require await response.parse(). And AnthropicBedrock raises instead of silently defaulting to us-east-1 when no AWS region is configured — which will surface a class of latent misconfiguration that has been quietly working by accident. There is a v1 migration guide with before-and-after snippets; use it.
2.1.239 Fixes a Bedrock Bug That Was Doubling Your Bill
Claude Code v2.1.239 landed yesterday with 59 changes, and one of them deserves to be read twice: Bedrock streaming behind proxies that strip the response Content-Type header was silently doubling billed API calls, because every turn was being re-run non-streaming. If you run Claude Code on Bedrock behind a corporate egress proxy, that is a fix worth diffing against your last invoice. A related one: startup hangs behind an HTTPS proxy on Bedrock with an SSO profile and awsAuthRefresh — the credential pre-check now honors HTTPS_PROXY.
Also in the drop: /claude-api upgrade, which migrates Python projects from anthropic 0.x to 1.x — shipped in the same week as the SDK 1.0 above, which is not a coincidence. Windows gets cross-session messaging, so SendMessage and ListAgents work across machines the way they already did on macOS and Linux. Cost estimates (/cost, status line, --max-budget-usd) now include the 1.1× US-only-inference premium for data-residency workspaces. And /goal check-ins on long-running background work finally back off — 30 minutes, then an hour, then every two — instead of pinging every 30 minutes forever. Plus a fix for WebFetch retaining expired page content in memory for the whole session rather than the intended 15 minutes.
Sonnet 5 Stays at $2/$10, Permanently
Worth restating as August closes, because the deadline everyone was budgeting around no longer exists: Claude Sonnet 5’s introductory $2 / $10 per million input/output tokens is now the standard price. The scheduled increase to $3 / $15 on September 1 is not happening. If you built a cost model this summer that assumed a 50% jump next week, delete that line.
The read on why is not complicated. With comparably priced competition on the developer side, there was not much headroom to raise a mid-tier model’s price without inviting migration — and Sonnet is the tier most likely to be sitting in high-volume production paths where a 50% increase triggers an actual procurement review rather than a shrug. Locking the price removes a reason to go shopping.
The Watermark Backlash Is About Detection, Not Transparency
Three weeks into global watermarking of Claude’s text and file output — Anthropic’s implementation of the EU AI Act Article 50(2) Code of Practice, applied worldwide rather than just in the EU — the loudest complaints are not from EU compliance teams. They are from users who realized the marks will identify Claude-assisted work at their jobs and in their classes. TechCrunch covered the reaction; the honest summary is that a transparency measure and a detection measure are the same artifact viewed from two sides.
The nuance that keeps getting lost is Anthropic’s own caveat: a detected mark means Claude may have processed the content, not that Claude authored it. Marks can land on human writing that Claude only edited or translated. That distinction is going to matter enormously the first time someone is accused on the basis of one, and it is exactly the sort of caveat that does not survive the trip into a university honor code or an HR policy. Models launched in the EU on or after August 2 support marking at launch; earlier models are in a transition period. Non-compliance exposure under Article 50 runs to €15 million or 3% of global turnover, whichever is higher — which explains the global rollout.
Two Clean Days on the Status Page
After a week that produced three incidents in 24 hours, the status page has recorded no incidents on August 21 and none so far on August 22. All systems operational. The last events were Thursday’s pair — elevated errors across multiple models, resolved 19:42 UTC, and a Google connectors outage resolved 19:01 UTC.
The 90-day numbers still carry the summer: claude.ai 99.37%, API 99.48%, Claude Code 99.4%, Cowork 99.49%, Console 99.86%, Claude for Government 100%. Two quiet days do not move a 90-day average, and this section stays here until the trend does. But it is a genuinely better weekend than last one, and if you were planning to test that failover path today, you have the window.
Nobody Shipped a Model This Week, and That Was the Story
Line up what actually happened over the last 48 hours. A protocol roadmap whose headline items are identity federation and tool catalog pagination. An acquisition of a consultancy that writes connectors. An SDK major version whose defining feature is a list of things that were removed. A Claude Code release whose most valuable line is a billing bug behind a corporate proxy. Not one of these is a capability announcement. All of them are load-bearing.
The MCP roadmap is the clearest artifact. Read what it treats as unsolved: an agent calling a server on behalf of a user who is not present has no standard way to prove who it is, so people paste API keys. A server with a hundred tools poisons its own selection quality and bills you for the privilege before the first question. A tool result can come back in multiple shapes and the server author cannot know which one the client will show the model. These are not research problems. They are the problems you hit in month three of running agents in production, and the fact that they are now the top of the roadmap tells you where the ecosystem actually is.
The Casper acquisition says the same thing from the commercial side. Ode launched in July to sell enterprise AI transformation; one month later its first move is to buy a dozen consultants who know how to wire Claude into systems that already exist. If model quality were the constraint, you would spend that money differently. And the Python SDK 1.0 is the version of this that lands on individual developers: the removals — Text Completions, sampling parameters on Messages, the silent us-east-1 default — are all cases of the library finally refusing to paper over ambiguity. That is what a platform does when it stops optimizing for the demo and starts optimizing for the thing running unattended at 3am.
The uncomfortable corollary: none of this is what gets a company priced above $2 trillion in an October listing. The IPO narrative runs on capability and growth curves. The actual engineering agenda this month runs on identity, transport, billing accuracy and deployment labor. Both can be true, and usually are — but if you are deciding where to spend your own team’s next quarter, the roadmap is the more honest document.