White House Tells Anthropic It Opposes Plan to Expand Mythos Access to ~70 Companies — WSJ Scoop, Bloomberg Confirms
Wall Street Journal broke the story Wednesday night and Bloomberg confirmed by morning: members of the Trump administration have told Anthropic they oppose the company’s plan to grant Mythos access to roughly 70 additional companies and organizations — a move that would have taken the total Mythos surface from about 50 entities today to around 120. Two named concerns from administration officials: first, that Anthropic doesn’t have enough compute to serve the expansion without degrading the government’s ability to use Mythos for its own purposes; second, that on the same day Anthropic announced the limited release plan, a small group of unauthorized users in a private online forum gained access to Mythos. The breach line is the part of the story that lands hardest with national-security readers: the gap between Anthropic’s controlled-access design and the practical difficulty of containing a model with autonomous-exploit capability is now a documented operating-environment fact, not a theoretical risk.
The structural read is the contradiction inside the same administration. The White House is opposing Mythos civilian expansion this week while simultaneously drafting an executive action that would restore Anthropic across federal agencies in a workaround of the Pentagon’s national-security supply-chain risk designation from earlier this year. The same building is blocking commercial expansion and rebuilding government access at the same time. For Anthropic this means Mythos commercial revenue is now policy-gated — not technically gated — and the IPO narrative for the October window has to hold a position where the company’s most cited frontier capability is sold one customer per White House meeting. The UK negotiation continues in parallel; the asymmetry — block at home, ship abroad — is now the visible shape of Mythos distribution policy.
1M Context Beta Retires Today on Sonnet 4 and Sonnet 4.5 — context-1m-2025-08-07 Header Stops Working
Today is the cliff. As of April 30, 2026, the context-1m-2025-08-07 beta header has no effect on claude-sonnet-4-5-20250929 or claude-sonnet-4-20250514. Requests over 200,000 tokens to those models return a 400 error from this morning forward. The header is still accepted syntactically — that’s the silent-fallback trap to watch for — but the long context surface is gone. Anyone with batch jobs, multi-document ingestion pipelines, or long-trace agent runs still pinned to the old Sonnet 4 / 4.5 model IDs needs to ship the migration before the next prompt lands or take the production failure.
The migration itself is a one-line change. Update the model ID to claude-sonnet-4-6 or claude-opus-4-6; remove the anthropic-beta: context-1m-2025-08-07 header. Both models have 1M context generally available at standard pricing — no beta gate, no surcharge — since the March 13 GA flip. For Bedrock and Vertex paths the same model-ID swap applies. Riptide’s migration writeup is the cleanest production checklist circulating this week; the Anthropic context-windows doc is the canonical reference. Note for Code users: Sonnet 4.6 1M context still has surface bugs on Max 20x plans where only Opus 1M shows in /model — the Code issue tracker has the live thread.
Mythos UK Track Continues as US Track Tightens — AISI Eval Public, CETaS Frames Regulator-Vendor Layer
While the White House blocks domestic expansion, the UK side moved another step in the opposite direction. AISI — the UK AI Security Institute — published its evaluation of Claude Mythos Preview’s cyber capabilities, noting that under explicit direction with network access, Mythos was able to execute multi-stage attacks on vulnerable networks and discover-and-exploit unknown vulnerabilities autonomously, the kind of work that takes human professional teams days. The CETaS piece at the Alan Turing Institute landed yesterday and the framing that’s emerging is the one to watch: Mythos forces a new shape of regulator-vendor relationship, not adversarial but interleaved at the eval layer. That’s the institutional language the UK working group is building around.
The active negotiating surface this morning: UK Treasury, BoE, FCA, NCSC, and AISI on the working group; the Reuters “within days” report from April 16 is now fourteen days past, and Anthropic has declined to commit to a public timeline through this week’s coverage. Project Glasswing remains the named distribution channel. The Bloomsbury Intelligence & Security Institute and ICAEW both ran public-policy framings this week treating Mythos as the shift from “model regulation” to “capability regulation” — the line that ends up in front of the EU AI Office next. Brussels and Tokyo have both still not shown their hand publicly; both are still positioned to do so before mid-May.
Claude Code 2.1.123 Lands Clean — OAuth 401 Retry Loop Fixed, Bedrock Tier Selector Added Yesterday
Claude Code shipped 2.1.123 on April 29 and the canary cycle held overnight. The headline fix is narrow but high-value: OAuth authentication was failing with a 401 retry loop when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 was set in the environment — the variable a lot of compliance-controlled shops use as a default. Teams that hit the loop after upgrading to 2.1.121 or 2.1.122 can now move clean. The previous day’s 2.1.122 dropped two ergonomics wins worth pinning: an ANTHROPIC_BEDROCK_SERVICE_TIER environment variable for selecting default, flex, or priority on the Bedrock path (the lever finops teams have been asking for since Bedrock pricing tiers landed), and a /resume search box that now finds sessions by pasted PR URL across GitHub, GitHub Enterprise, GitLab, and Bitbucket.
Production recommendation: pin moves to 2.1.123 for any team running with DISABLE_EXPERIMENTAL_BETAS=1; teams without that var can hold on 2.1.121 through end of week and roll together. The Bedrock tier selector is the upgrade trigger for teams running a multi-tier finops stance — pair it with the rate-limits API that shipped a week ago Saturday and the routing logic gets simple. Bedrock and Vertex remain the most resilient deployment surface for production-critical paths after this week’s outages.
Brief Overnight Incident on Claude.ai and API — 01:20 UTC, Resolved Inside the Window
Claude.ai and the Anthropic API hit an availability incident at 01:20:26 UTC this morning per the Claude status page. The service was operational by 09:08 UTC. No public root cause yet; a postmortem typically lands inside ten business days. This is the third multi-surface event in roughly seven days after Tuesday’s 78-minute API outage and Wednesday’s elevated errors on Haiku 4.5 between 12:37 and 13:37 UTC. None of the three has been long enough on its own to be a board-level event — the cumulative pattern is the part platform engineering teams are flagging.
The operational read for finops and SRE: the third incident inside a single rolling week is the threshold where multi-region failover and Bedrock/Vertex secondary deployment paths become a budget conversation, not a posture conversation. The rate-limits API that shipped Saturday gives you the telemetry to instrument it. For shops still on direct API for production-critical paths, this week is the one to use as the case for a secondary deployment surface.
Sonnet 4 / 4.5 1M Migration: One-Liner if You Read the Header, Two-Liners If You Wrap a Client
Practical follow-on to today’s deprecation. For shops calling the API directly: drop the anthropic-beta: context-1m-2025-08-07 header and update the model identifier from claude-sonnet-4-5-20250929 or claude-sonnet-4-20250514 to claude-sonnet-4-6. For shops using a wrapped Anthropic client (LangChain, LlamaIndex, the OpenClaw shim) the two changes are: bump the model in your config and remove the beta header from any custom transport. Bedrock and Vertex paths take the same model-ID swap. Pricing is identical to standard Sonnet 4.6 / Opus 4.6 — 1M context is no longer a premium tier; the long-context surcharge ended at March 13 GA.
Three pre-flight checks before you ship the change: first, search your codebase for the literal beta header string and confirm it’s not silently still in any tool wrapper. Second, run a 250K-token request through the new model ID before the cutover — the silent-fallback behavior on the old IDs after today is the failure mode that bites production. Third, confirm your output-token limit configuration didn’t inherit a “1M = bigger budget” assumption that shouldn’t carry forward. The MindStudio migration guide and the Tygart writeup both have copy-paste diff snippets for the common SDK paths.
iCapital Picks Anthropic for Its Next AI Phase — Compliance-First Posture Cited, Claude Code Already Inside Dev Platform
iCapital, the alternative-investment platform serving wealth managers and registered investment advisors, named Anthropic as its AI partner for the next phase of the company’s AI strategy. The cited reasons in the announcement track exactly with what Anthropic has been positioning into the regulated-finance segment: reasoning capability, interpretability tooling, and the ability to operate inside a compliance-first deployment posture. Claude Code is already deployed inside iCapital’s software development platform; the next phases bring the model into client-facing applications and the operational stack. iCapital is keeping an open AI architecture for the platform overall — this is a partnership, not a single-vendor lock — but Anthropic is the named primary for the next investment cycle.
Strategic frame: iCapital is the first named “wealth platform” anchor under the Anthropic Partner Network umbrella, and it lands inside a vertical — alternative investments — where the deployment surface (advisor-facing, KYC/AML-adjacent, recordkeeping-bound) is one of the strictest compliance contexts in financial services. The win matters for the IPO narrative: Anthropic has the tier-1 banks (Goldman, JPM, others) on direct contracts; the wealth-platform tier was the next named gap, and iCapital fills it. Watch for advisor-network analogs — Envestnet, Orion, Addepar — to be the next named cohort inside the next ninety days.
$30B ARR Reset Holds Going Into May — OpenAI CRO Pushback Aside, Investor Narrative Locks In
Three weeks past the April 7 milestone, the $30B annualized revenue run-rate datapoint is now the headline number every IPO model is being built around. Anthropic crossed it in March, up from $9B at end-2025 — roughly 1,400% year-over-year and a four-month $9B-to-$30B sprint that’s the steepest segment in private-software history. OpenAI’s CRO disputed the figure as overstated by ~$8B based on cloud-partner accounting, but Anthropic’s position — principal in the cloud transactions, gross revenue recognition is correct — has held in coverage. The 1,000+ business customers spending over $1M annually has more than doubled since February.
The strategic read into the October window: today’s Mythos block, the Goldman HK carve-out, the iCapital pickup, and the $30B reset are four data points in the same investor narrative. Anthropic is now a frontier lab whose risk surface is itself the product of its market position — and whose revenue is now category-leading. Bull thesis: regulated-enterprise depth plus creative-work breadth plus sovereign-aligned compute. Bear thesis: policy-gated frontier capabilities, jurisdictional exposure, agentic-failure liability. The S-1 narrative has to hold all six positions; the $30B number is the gravity well that holds the bull side together.
MCP Ecosystem Hits 500+ Servers, 97M Monthly SDK Downloads — OAuth 2.1 Lands in the Spec This Month
The MCP ecosystem rolled past 500 active servers and 97 million monthly SDK downloads heading into May, per the April rollup pieces circulating this week. Official client support is now stable across every major editor (Claude Desktop, Code, VSCode, Cursor, Windsurf, Zed) and the enterprise security story holds up under audit — the OAuth 2.1 authorization addition with incremental scope consent landed in the protocol spec this month, which is the missing piece for regulated-enterprise deploys. The CVE MCP server out of github.com/mukul975/cve-mcp-server is the standout new third-party server — 27 tools across 21 vulnerability APIs, configured for both Claude Desktop and Claude Code on first install.
The pattern for technical-publications and vertical-tooling teams: the MCP surface is now a credible delivery channel for domain-specific AI capabilities, not a science project. Trimble’s SketchUp MCP from earlier this week is the “binary file format end-to-end” reference architecture. The CVE MCP is the “multi-API tool-call orchestration” reference. Both ship with first-install configs that work in the major Claude clients without bespoke setup. For S1000D / CGM / IETM-adjacent shops looking at the agentic-publishing roadmap, the MCP path is the one that has the most pattern density to copy from this week.
One Building, Two Doors: The White House Is Blocking Mythos Civilian Expansion While Drafting Federal Re-Onboarding
Today’s lead story sits inside a contradiction that has to be named directly. The same Trump administration that is opposing Anthropic’s expansion of Mythos to roughly 70 commercial customers this week is also drafting an executive action that would restore Anthropic across federal agencies in a workaround of the Pentagon’s national-security supply-chain risk designation. One building, two doors. The civilian door is closing on Mythos this morning; the government door is being unlocked on every Anthropic surface the Pentagon previously walled off. The reason the contradiction is not actually inconsistent — from the policy planner’s seat — is that the same capability that makes Mythos too dangerous to put behind 70 commercial NDAs is the capability the federal government wants to reserve for its own use first. That is the argument the executive action is going to make, and that is the argument the civilian-expansion block is the first half of.
The investor read for the IPO window: this is the day the “policy-gated frontier capability” bear thesis got its first concrete data point on a US-side action. Anthropic’s strongest commercial differentiator at the high end is now bundled under a distribution channel where the gating function is a White House meeting, not a sales process. Two outcomes follow. First: the regulated-enterprise wedge (iCapital, Goldman ex-HK, Accenture / Deloitte / NEC / Caylent depth, $30B ARR) becomes a structurally larger share of the S-1 narrative than the frontier-capability wedge, because the regulated-enterprise surface is sales-process-gated and the Mythos surface is policy-gated. Second: the UK negotiation now matters more, not less — if the UK becomes the first non-US jurisdiction to take Mythos commercially, the asymmetry “blocked at home, shipped abroad” becomes the structural shape of the frontier business. That’s a different shape of company than the investor narrative was built around six months ago, and the October prospectus will have to read like one.