Opus 4.8 ships, Anthropic prints $965B —
the same afternoon.

Anthropic's first frontier model since Opus 4.7 (41 days ago) lands with the numbers that matter to the agent layer. SWE-Bench Pro moves from 64.3% to 69.2%, USAMO 2026 from 69.3% to 96.7%, and Anthropic claims Opus 4.8 is roughly four times less likely to let a flaw in its own code pass unflagged. Standard pricing holds — $5/M in, $25/M out — but the new fast mode runs at $10/$50 per million, 2.5× the speed for ~three times less money than Opus 4.7's fast mode. An effort dial (low / medium / high / xhigh) appears across Claude.ai, Cowork and Claude Code, with Opus 4.8 defaulting to high. And the research preview people will be talking about all weekend: dynamic workflows lets Claude Code write an orchestration script that spawns parallel subagents — up to 16 concurrent, 1,000 total per run — each planning, executing and verifying a slice of a task and reporting back. Anthropic frames the use cases bluntly: codebase-scale migrations across hundreds of thousands of lines, profiler-guided optimisation audits, security hardening sweeps, "work you need checked twice." The model is available the same day on Claude API, Amazon Bedrock, Vertex AI and Microsoft Foundry (1M-token context on the first three, 200K on Foundry). Read alongside Cognition's $26B mark last week, Opus 4.8 is the model layer's price-anchored answer to the autonomous-coding-agent thesis — same dollar, more agent.

Two weeks ago the round was Bloomberg leaks at "~$900B"; one week ago last edition flagged it as "circling >$30B at ~$900B, ahead of OpenAI"; on Thursday it closed at a different number entirely. Altimeter Capital, Dragoneer Investment Group, Greenoaks Capital and Sequoia Capital led $65B in primary at a $965B post-money — vaulting Anthropic past OpenAI's $852B March mark for the first time, on a reported $47B annualised revenue run and Anthropic's first projected quarterly operating profit. The deal is the second $30B-plus round Anthropic has printed inside one calendar year (after February's Series G at $380B), and it lands the same afternoon as Opus 4.8. The optics aren't subtle: the model ships, the revenue line clears $47B annualised, the valuation clears the rival's, and OpenAI files its confidential S-1 the same week (item 08). Three years ago Anthropic was the foundation lab that valued safety over speed and a few months behind on the product cycle. This week it priced both above OpenAI on the same day it shipped them.

A nine-release train across the Opus 4.8 launch arc. v2.1.146 gave /code-review proper effort levels; v2.1.147 kept pinned background sessions alive on idle and reshuffled /code-review//simplify; v2.1.148 fixed a Bash exit-code-127 regression; v2.1.149 broke /usage out by category, added keyboard scrolling in /diff, taught Markdown rendering GFM task lists and landed PowerShell security fixes; v2.1.150 was internal; v2.1.152 (covered last edition) shipped --fix, MessageDisplay and skill hot-reload; v2.1.153 added a skipLfs option to the git plugin, polished npm auto-update notices and tightened the MCP server approval workflow. Thursday's v2.1.154 is the punch line: Opus 4.8 is the default, high effort is the default for Opus 4.8/4.6/Sonnet 4.6 (xhigh for Opus 4.7), the lean system prompt is now the default, and dynamic workflows is on by default on Max and Team — in the CLI, Desktop and the VS Code extension. The seam between "model release" and "agent harness release" has collapsed into one commit graph.

Codex shipped 0.135.0 the same day as Opus 4.8 — fitting, given how much of the release reads as the harness side of the same problem statement. codex doctor reports richer environment and system information for "why isn't my agent running?" triage. Vim mode picks up text-object editing and a configurable interrupt-turn binding (the equivalent of / reverse-history in Claude Code). /permissions takes named permission profiles — read-only mode by name, finally — and the Python SDK introduces friendly Sandbox presets for thread and turn APIs, which is the foundation for "spin up a hundred parallel agents and have them not eat the host." TUI stability fixes on macOS and Zellij round it out. The codex-vs-claude-code race is now visibly about whose fan-out story scales — and 0.135.0 is OpenAI's response to the dynamic-workflows preview shipped at noon.

Microsoft's Agent Framework lands a release that names what every other harness vendor is now also chasing. HarnessAgent and a new background-agents harness provider formalise the "long-running task in a managed sandbox" pattern Claude Code's dynamic workflows previews from the other side. A2AAgentSession gains task-reference support — i.e. agents can hand off and resume each other's work over the A2A protocol without dropping state — and an experimental Foundry APIs surface plus tool-name alignment across implementations smooth the multi-runtime story. There are breaking changes too (Python-only declarative actions are out, and streaming/citation-metadata bugs are fixed). Read alongside Anthropic's dynamic workflows preview and OpenAI's Codex sandbox presets, all three reference harnesses are now building toward the same primitive — a planning model that orchestrates a fleet of background workers — and naming it almost the same thing.

The framework that ate the Python agent middle layer keeps printing. v1.103.0 adds list_prompts and get_prompt to McpServer — the prompt half of MCP that most servers still haven't bothered to implement is now first-class on the client. Anthropic eager_input_streaming lands on OpenRouterModel, message-timestamp round-tripping comes through the VercelAIAdapter, and a clean batch of Bedrock, thinking-mode and toolset-callback fixes shipped with it. If you missed v1.102 (security fix for an IPv6 transition-form SSRF bypass) the upgrade is doubly worth it. Three releases (1.101, 1.102, 1.103) plus two v2 betas inside eight days — Pydantic AI's cadence remains the loudest signal in OSS agent tooling.

The Bloomberg scoop the AI press has been chewing on all week. Twenty-five "power users" inside the Department of Defense have been running side-by-side evaluations of OpenAI, Google and xAI's Grok models against Claude since early March — three days after Defense Secretary Pete Hegseth labelled Anthropic a "supply-chain risk" over Anthropic's refusal to drop usage policies that bar mass surveillance and lethal autonomous weapons. OpenAI signed a classified-network deployment inside hours of Anthropic's blacklisting. Anthropic is fighting the designation in court, saying it could cost billions in revenue. The pragmatic wrinkle Scientific American makes explicit: even if a replacement model wins on benchmarks, "the hard part begins after the model is gone" — every integration, MCP server and skill currently wired to Claude has to be re-wired by hand. The bake-off is real; the rip-and-replace is the project. And every other Fortune-500 risk committee just got an awkward new "what if our regulator labels our model provider a supply-chain risk?" agenda item.

Working with Goldman Sachs and Morgan Stanley, OpenAI filed a confidential S-1 with the SEC targeting a public listing as early as September 2026 at a $852B–$1T valuation. Full financials stay sealed for the SEC back-and-forth; the public S-1 has to drop at least 15 days before the roadshow. Two implications worth holding together: (1) the IPO clock and the Pentagon court fight are now running concurrently for OpenAI's most direct competitor, and (2) the public listing forces a level of revenue disclosure that will, for the first time, let everyone independently price what "an agent business" is actually worth. The same week Anthropic's $965B mark cleared OpenAI's $852B — and now OpenAI is inviting the public market to reset the comparison. A listing in three months would be the most consequential AI-economy moment since GPT-3.

Acrew Capital and a16z crypto led $30M into Catena Labs, with Breyer Capital, General Catalyst and QED alongside, taking total funding to $48M. Catena's pitch: programmable banking rails specifically for AI agents — KYC for non-human identities, programmatic balance and payment APIs an agent can call without touching a human treasurer, the same Compliance / SOC2 / audit surface enterprises already require. To get there, Sean Neville (Circle cofounder) filed with the U.S. Office of the Comptroller of the Currency for a national trust bank charter. That's the part that matters: not a fintech wrapper on top of a correspondent bank, but a directly chartered institution whose primary customer is, by design, an autonomous agent. If the OCC approves it, the U.S. just stood up a regulated banking primitive for the agent economy — and every "agents can hold their own credit pool" pitch (see Anthropic's split billing last month) gets a payment side to match.

Three Mythos beats this week, all pointing the same way. May 22: EU–Anthropic talks over Mythos cybersecurity concerns are stalled, per Spanish officials cited by Bloomberg — the first real friction between Mythos and a major regulator. May 26: Anthropic publishes a new anthropic.com/glasswing hub, confirms the consortium has now found 10,000+ high-/critical-severity vulnerabilities across over 1,000 open-source projects, names a 27-year-old remote DoS flaw in OpenBSD as a representative find, and discloses up to $100M in Mythos credits and $4M in direct donations to open-source security orgs. May 28: the Opus 4.8 launch post and Axios both confirm Mythos-class models are expected publicly "in the coming weeks" — the first time Anthropic has put a near-term date on it. Counter-pressure is real: operational-technology vendors say they're being excluded from Glasswing; an unauthorised forum reportedly accessed Mythos via a contractor's leaked credentials in April. The trajectory is clearer than the timeline: the model that was "too dangerous to ship" in April is now planning its public release on a calendar measured in weeks, not quarters.