India Steps Up Bank Cyber Preparedness on Mythos — NPCI Seeks Early Access, RBI Consulting Fed and BoE
India formally moved Monday. The Centre asked scheduled commercial banks to harden cyber posture against Mythos-class exploit risk, the Reserve Bank of India confirmed consultations with the U.S. Federal Reserve and the Bank of England on a coordinated supervisory read, and the National Payments Corporation of India — the body that runs UPI — is in active negotiation with Anthropic for early Mythos access for a small group of Indian banks. The aim is the same one the UK Treasury is pursuing: get supervised access ahead of any wider rollout so zero-day exposure inside critical national-payment infrastructure can be patched before the model meaningfully changes the attacker calculus. The National Payments Corporation has data-localization complications to work through — India’s 2018 rules require payment data to stay on Indian servers — and that is the live diplomatic surface.
What this puts on the board: India is now the second sovereign formally inside the Mythos negotiation, after the UK Treasury and AISI. Australia’s Home Affairs has confirmed direct engagement; Germany’s Bundesbank is publicly pressing for universal access; the EU Commission has not yet shown its hand. The pattern across G7 and G20 supervisors is becoming the regulatory baseline: nobody commits to a public timeline, every supervisor convenes its own working group, and the institutional posture is that frontier-AI cyber capability has moved from annual-review to quarterly-review on every banking regulator’s docket. Watch for the EU and Japan to make their first formal moves inside the next two weeks — that’s where the negotiation continues to expand.
UK Mythos Rollout: Eleven Days Past “Within Days,” Still No Date
The UK Mythos rollout that Reuters reported as “within days” on April 16 is now eleven days old with no public timeline commit. What changed this weekend: the Bank of England and UK Finance’s cross-market operational resilience group convened to discuss Mythos and broader frontier-AI cyber risk; Technology Minister Liz Kendall and Security Minister Dan Jarvis went on the record warning that frontier-AI cyber capability is accelerating faster than the threat models the public sector was working against last quarter; and the UK AI Security Institute’s evaluation of Mythos — the only one published by a non-Anthropic body so far — is what India’s RBI is now using as the reference document for its own consultation with the Fed and the BoE. Reuters added another report Friday confirming the negotiation is ongoing.
The cross-supervisor frame is becoming the story. The UK was first; India formalized today; Australia is in direct engagement; Bundesbank is publicly pushing for universal access. None of those negotiations is moving at marketing speed, but each one represents a permanent change in how a sovereign supervisor relates to a frontier lab. Mythos is the model that’s forcing those relationships to be written down. Watch for the next move from either Brussels (EU AI Act enforcement track) or Tokyo (Japan FSA) inside the next sprint.
Claude Code 2.1.119 Holds Clean Through Monday Morning — Pin 2.1.117 Still the Production Recommendation
The auto-rollback from 2.1.120 to 2.1.119 that landed Saturday at 02:35 UTC has held clean through Monday open. The eight regressions named on the survival-checklist gist — the --resume / --continue JS crash, the sandbox-required false alarm, the Antigravity spawn failure, the UI-duplication bug, the WSL2 freeze, the silent model swap, the auto-update break, and a permissions edge case — are no longer firing in fresh issues, and the GitHub tracker has gone quiet on the headline list. The status page is green. The auto-update channel is reverted; teams that hadn’t opted out are landing back on 2.1.119 on the next check-in.
The production recommendation has not changed: pin 2.1.117, set DISABLE_UPDATES=1 in the shell profile, and watch the GitHub tracker for two days after each minor before rolling forward. The Bedrock and Vertex paths remain the most resilient surfaces for teams that need uptime over latest features. The next thing to watch this week is whether 2.1.121 ships with a clean changelog or whether the same canary gap shows up again — the signal lives on the issue tracker, not on the marketing channels.
Managed Agents Beta Header Live — managed-agents-2026-04-01 Is Now the Long-Horizon Surface
The Claude Developer Platform’s Managed Agents service is in public beta with the managed-agents-2026-04-01 header. The pitch is durable state for long-horizon agent work: Anthropic-hosted sessions, harnesses, and sandboxes; stable interfaces for tool access and credential scoping; faster cold starts than rolling your own. For teams that have been running their own agent runtime on top of the API — the production-grade Pydantic / LangGraph / Inngest stack — this is the “don’t reinvent the orchestrator” offer with the lab’s name on the SLA. Sessions, harnesses, and sandboxes are the three primitives.
What it doesn’t yet do: replace the open-source orchestration cohort. Beta means the surface still moves, the pricing isn’t firmly published, and the tool-use semantics are not yet the same shape as on the regular Messages endpoint. The reason to build against it now is to lock in a code path that is going to be the supported long-horizon agent runtime for the next two years — the open-source middleware market just got a clear signal about which problems Anthropic intends to own.
Streamable HTTP Is the Final Transport — SSE Servers Should Migrate Now
Worth restating because the migration window is closing: Claude’s official MCP transport is now Streamable HTTP under protocol version 2025-11-25. The older SSE transport is deprecated. The servers that are still SSE-only are the ones that will start producing connection-warning logs on Cowork and Claude Code clients first, then hard-fail in a future release. The SDKs — Python (FastMCP), TypeScript, and the official Rust crate — have first-class Streamable HTTP support already shipped. The 500K-character result-size ceiling that landed in 2.1.91 is also Streamable-HTTP-aware.
For teams that wrote their own MCP server during the early SSE wave: this is a one-week migration if your server is mostly stateless tools, longer if you depended on long-lived event streams. The good news is that the Streamable HTTP protocol is significantly simpler to operate — standard HTTP/2 keep-alive, no SSE-specific reverse-proxy quirks, native compatibility with managed runtimes that didn’t love SSE in the first place. Get this off the to-do list before the deprecation moves from warning to error.
Monday Cap-Table Read — Forge Pricing $1T, Hiive $955, Google Tranche Settles Into the Tape
Markets opened with Anthropic’s secondary marks at roughly $1T on Forge Global and Hiive marking the underlying around $955/share — both prints higher than the $800B venture-offer floor reported through the weekend, and both ahead of OpenAI’s $880B secondary mark for the first time since the OpenAI for-profit conversion. The cap-stack story is the same one that built across the week: Amazon’s $5B-immediate / $20B-on-milestones tranche on Monday April 20, Google’s $10B-immediate / $30B-on-milestones tranche on Friday, the $100B AWS commitment underneath both, the 3.5GW Broadcom/Alphabet TPU expansion on the silicon side, and the $50B Fluidstack data-center pledge as the build-out vehicle.
Indian and European business press picked up the Google tranche this morning, which is what’s driving the secondary-market intraday move. The IPO arithmetic that the weekend brought into focus — a $30B annualized revenue base, more than a thousand enterprise accounts spending $1M-plus annually (doubled since February), eight of the Fortune 10 named as customers, an October window targeting a $60B raise — is now the working hypothesis on every U.S. equities desk that covers AI-platform names. The diligence overhang remains the same: consumer-trust posture (the Pro-pricing wobble, the Claude Code regression weekend) and the legal exposure surface from Sunday’s “cannot control once deployed” filing. Both are durable. Neither is a deal-breaker. October stays plausible.
GPT-5.5 “Spud” vs Opus 4.7, Four Days In — Two Frontier Models, Two Different Axes
OpenAI shipped GPT-5.5 “Spud” on April 23. It is the first fully retrained base model since GPT-4.5, not an incremental 5.x bump. The four-day read on the leaderboard: Opus 4.7 still leads SWE-Bench Verified at 87.6%, with GPT-5.3-Codex at 85.0% in the #2 slot and Spud-derived agent systems trailing on Verified but pushing on Pro. SWE-Bench Pro has Opus 4.7 at 64.3% (Anthropic-reported); GPT-5.4 (xHigh) reaches 59.1% on Scale’s SEAL mini-swe-agent scaffold, with Spud’s Pro number still being independently verified. Gemini 3.1 Pro sits at 80.6% on Verified, the open-weight MiniMax M2.5 at 80.2%, Sonnet 4.6 at 79.6%.
The cleaner read is that the two frontier models are not competing on the same axis. Spud leads planning-and-execution evaluations (Terminal-Bench 2.0, Expert-SWE, Toolathlon) by meaningful margins; Opus 4.7 leads codebase-resolution evaluations (SWE-Bench Pro/Verified, MCP-Atlas, CursorBench). The implication for routing: a sophisticated agent stack uses both depending on the task type, with Sonnet 4.6 doing the routing-weight work in the middle of the cost-quality curve. Salesforce’s decision Friday to bundle Sonnet 4.5 (not Opus 4.7) into every Developer Edition org as the default coding model is the practical confirmation — the mid-tier reliability is what production wants.
Consumer Connectors Pass the Weekend Adoption Test — Mobile Beta Picks Up the Long Tail
The fifteen consumer connectors that flipped on Friday — Spotify, Uber, Uber Eats, Instacart, TurboTax, Credit Karma, Booking.com, Audible, AllTrails, TripAdvisor, Viator, Resy, StubHub, Taskrabbit, Thumbtack — cleared the weekend without the kind of failure mode that usually accompanies a wide consumer surface launch. The Cowork desktop client is doing most of the work; the mobile beta on iOS and Android is where the long-tail, single-use-case adoption is showing up first. The connectors are explicitly opt-in, not used to train the model, and don’t see other Cowork sessions — the privacy story Anthropic positioned with the launch is the one consumer reporters quoted back through the weekend.
The interesting metric to watch this week is multi-connector composite use. The 200+ connector directory is a count, but where the agent gets useful is the multi-step composite — book a Resy reservation, then call a Uber, then add the calendar invite, then put the playlist on Spotify. That’s the workflow a single-app integration cannot fulfill, and that’s the surface where consumer Claude pulls meaningfully ahead of ChatGPT and Gemini for the next two quarters. The next vertical to watch break through: legal research, healthcare scheduling, supply-chain procurement — the three categories where the “agent that actually does the job” thesis has not yet picked winners.
The Mythos Sovereign Pattern Is Now the Permanent Operating Mode
Look at the supervisor side of the board today. The UK is eleven days into “within days,” with the BoE convening operational-resilience working groups and AISI publishing the only third-party Mythos evaluation. India formalized today, with the Centre asking banks to harden, the RBI consulting the Fed and the BoE, and the NPCI — the body that runs UPI — quietly negotiating early access for a small bank cohort. Australia is in direct engagement on vulnerability tracking. Germany’s Bundesbank is publicly arguing for universal access. Japan and the EU have not yet shown their hands but will inside the next sprint. The pattern is becoming the permanent operating mode: every G7 and G20 financial supervisor will, before the end of Q2, have a working relationship with Anthropic that did not exist 90 days ago.
That’s the structural change worth pricing. Until Mythos, the relationship between a frontier lab and a sovereign supervisor was filtered through general AI-policy bodies (the EU AI Office, the U.S. AISI, the UK AISI). What Mythos has done is force direct, model-specific, supervisor-by-supervisor engagement at the technical level — the kind that until last quarter only happened with critical-infrastructure vendors and arms exporters. That is a permanent step-change in how a frontier lab is treated, and it is happening exactly in the window where Anthropic is preparing for an October IPO. The S-1 conversation about “regulatory exposure” just got a much fuller footnote: it is no longer a category risk, it is an operating cadence. The cleanest version of the bull case for the public listing is that Anthropic now has the regulator-by-regulator process other frontier labs do not. The cleanest version of the bear case is that the same process becomes a tax on the next generation of frontier capability. Both readings are defensible. Both will show up in the S-1.