Cursor goes to SpaceX for $60B, Cowork ships on Opus
— the harness war picks sides.

The largest dedicated coding-agent transaction in history, and the first time a public AI- infrastructure ticker uses its own equity to consolidate a model-agnostic harness. Per CNBC, Yahoo Finance and the AI Business Tuesday writeup, the Jun 16 Agreement and Plan of Merger between SpaceX, wholly owned subsidiary X67 Inc. and Anysphere converts every Anysphere common and preferred share into the right to receive SpaceX Class A stock at an implied equity value of $60.0B — exercising the option Musk took out in April for either a ~$10B partnership tap or full acquisition. Cursor runs roughly $2.6B annualised revenue, with the TechTimes Tuesday tag reading "four days after the record IPO": SPCX priced at $135 Jun 12, closed day one +19% at $160.95, then closed $192.50 on Jun 15 and $211.85 on Jun 16 per the StockTwits ticker — a ~57% move from pricing into the day of the merger filing. Two reads. (1) The harness-as-acquisition-target tape rewrites yesterday's thesis. For ten weeks the debate was "does the model own the runtime or does the runtime own the model?" — SpaceX just answered "the buyer owns both, and pays $60B for the runtime even though the model on the other side of the curb is offline today." Cursor's value-add — the routing layer that picks Composer, Claude, GPT-5.5 or local — gets worth more, not less, the day the Mythos-class flagship goes dark for foreign nationals worldwide. (2) The $4B termination fee / $10B walkaway disclosed in the SPCX proxy is the second-order tell. Either number is bigger than the entire Cursor ARR; the deal is being priced as a strategic capability with a regulatory beta, not as a multiple on revenue. Q3 close pending antitrust review.

The single largest deployment surface for a third-party agent runtime in the enterprise landed today on Opus 4.8, with the Mythos-class seat explicitly left empty — and with a billing primitive that breaks Microsoft's fifteen-year per-seat default. Per the Microsoft 365 Blog Tuesday and Charles Lamanna's thread, Copilot Cowork — built "in close collaboration with Anthropic" — runs every tenant on Anthropic Opus 4.8 and Sonnet 4.6 at launch; Frontier customers additionally get GPT-5.5; Microsoft's own Cowork 1 ("a secure, fine-tuned model designed to handle everyday Copilot tasks at a substantially lower cost") is staged "in the coming weeks". Pricing breaks the per- seat default: a Microsoft 365 Copilot USL remains the prerequisite, then every Cowork task is metered in usage-based Copilot Credits across light, medium and heavy bands based on "model use, context retrieval, tool calls, and runtime". Coverage runs inside the Microsoft 365 trust boundary: audit logs, DSPM, eDiscovery, Insider Risk Management, Data Lifecycle Management (GA Jun 22); DLP "coming soon". Two reads. (1) The Opus-as-default-Cowork- brain is the loudest endorsement Anthropic has received this week — and the cleanest read on what "close collaboration" has been buying. Five days into the Fable 5/Mythos 5 suspension, Microsoft is willing to stake the M365 agentic tier on the Anthropic models that are available — a tier-2 default that holds even if the curb extends weeks. (2) The Copilot Credits meter is the structural shift the SaaS press buried under the model news. Once M365's most-used agent layer is metered per task rather than per seat, every competing agent platform inherits the same pressure to break per-seat — exactly the billing shape Anysphere, Codex and Antigravity already ship on. The pricing primitive of the agent layer just standardised under Microsoft's name.

The first time the underlying instrument behind a deployed US frontier-model suspension has been read in full by the public — and the language is materially harder than the "safeguards" framing the administration used over the weekend. Per Bloomberg and the parallel Bloomberg Law piece dated Jun 16, Commerce Secretary Howard Lutnick's Friday Jun 12 letter to Dario Amodei requires Anthropic to obtain government permission before exporting Fable 5 or Mythos 5 "to any destination worldwide, or to any foreign national regardless of location" — and threatens criminal and civil penalties if the company fails to comply. Lutnick told reporters the rationale was a fear the models "could be deployed by military intelligence users in China, Russia or other countries of concern". Two reads. (1) The "to any destination worldwide" phrasing is the part downstream lawyers will read first. It does not carve out US citizens inside the United States by construction — Anthropic's globe-wide blackout was a compliance read of that literal phrase. Until the letter is amended in writing, the permission gate is the entire export surface. (2) The criminal-penalty language resets the political math. "Refused" in Sacks's Friday post implied a negotiable safeguards dispute; "criminal penalties" imply a Bureau-of-Industry- and-Security enforcement posture that survives a tone-shift, which is the line the Bloomberg reporting flags as the reason the negotiations in Washington have moved at the speed they have. The Évian lunch tomorrow inherits this letter, not the "reluctantly" framing.

The first formal door slammed on the "partner-country exemption" route bilateral diplomats had been mapping all weekend — and the cleanest read on what shape a restoration deal will take when it lands. Per City AM and the parallel Telegraph sourcing tracked through ResultSense, Downing Street had spent the post-curb weekend lobbying Washington for a UK exemption from the Friday directive that locked foreign nationals out of Fable 5 and Mythos 5; one source familiar with the matter told The Telegraph Tuesday there was "zero chance" of Britain receiving special treatment, and administration officials briefed that any eventual rollback "is more likely to apply globally than be granted specifically to the UK". Two reads. (1) The "zero chance" shape is what makes the prediction-market ~16-day mean restore window internally consistent. If London cannot buy out, the only path back is a classifier fix Commerce accepts for every foreign national — which is the safeguards engineering sprint the Brown/Heck/Carlini delegation booked into the room Monday. (2) The Five Eyes-symmetric-treatment question — already a working topic in The Globe and Mail's G7 preview — gets answered before the leaders sit down to lunch. If the closest US security partner cannot get a carve-out, neither can Canada, Australia or New Zealand; that scopes Carney's sovereignty pitch into a consensus problem, not a bilateral one.

The second prediction-market venue plants a flag on the Fable suspension, and the consensus shape lengthens by a week between Monday and Tuesday. Per CNBC's Jun 16 tape, Kalshi's "When will Anthropic restore Fable 5 access for US customers?" contracts — opened in the wake of the Polymarket series — price restoration ~58% by Jul 1 and ~74% by Jul 10; earlier Tuesday BingX, CryptoNews and Bitcoin.com News all ran the ~68% by Jul 1 snapshot, the quote that anchored the day. The Polymarket book stayed near ~82% by Jul 1 with the clean $120k-and-growing depth. Two reads. (1) The Kalshi/Polymarket spread is the new under-the-radar governance signal — venue-by-venue divergence on a single export-controls outcome lets administration negotiators read sentiment without taking a phone call. The ~14-point Polymarket-over-Kalshi gap implies the crypto venue is pricing a faster lift than the US-regulated venue, a spread the BIS staff can watch in lieu of press coverage. (2) The extension to Jul 10 is what changed Tuesday. Where yesterday's book put weight on Jul 1, today's prints a two-week restoration window — exactly the interval the Lutnick letter's criminal- penalty language warrants once it reads in public.

The Codex CLI cadence the harness war produced on defection day is still running — a fresh pre-release branch opens inside four hours of yesterday's stable cut and runs three pre-releases before sunrise. Per the openai/codex GitHub releases feed: rust-v0.141.0-alpha.1 at 01:00 UTC, alpha.2 at 01:48 UTC, alpha.3 at 06:07 UTC on Jun 16 — 151 assets per release, the 0.141 branch opening less than four hours after 0.140.0's 21:06 UTC Monday stable tag that landed /import, codex delete, /usage daily/weekly/cumulative meters, managed Amazon Bedrock API-key auth, and corrupted-SQLite auto- recovery. Two reads. (1) The pre-release-branch-opens-inside-the-shift rhythm is the structural read on what shape the harness war takes during the export- controls window. The Codex team cannot ship a faster Mythos-class model than Anthropic — but they can pay their tax on the runtime in release cadence, and the 0.141-alpha overnight band is the visible bill being run up. (2) The three-alphas-in-six-hours shape under SPCX's Cursor move tightens the "OpenAI vs Cursor-as-SpaceX-asset" competitive frame. Codex is the model-vendor-controlled harness and Cursor is now the satellite-vendor-controlled harness; the cadence under both names compounds on pricing rather than features.

Anthropic's public answer to the question "does the harness still ship while the flagship sits in Commerce?" turns out to be "yes, on a calmer cadence" — the day after the Brown/Heck/Carlini delegation negotiation, Claude Code cuts a housekeeping point release rather than a feature one. Per the code.claude.com changelog Tuesday 20:22 UTC, v2.1.179 ships without a Tool(param:value)-grade headline: mid-stream connection drops now preserve partial responses instead of raw errors; the spinner-stuck-at-running-tool regression is gone; WSL2 mouse-wheel scrolling under Windows Terminal and VS Code is fixed (a regression from v2.1.172); Linux sandbox denyRead/ allowRead globs over large directory trees no longer hang sessions; Ctrl+O now shows subagent transcripts; remote-session background tasks no longer appear stuck as "still running" between turns; feedback survey single-digit replies no longer get captured as session ratings. Two reads. (1) The stability-only release-after-the-Lutnick- meeting shape is the structural tell. Monday's v2.1.178 shipped Tool(param:value) and classifier-vetted auto-mode 29 minutes after Codex went stable; Tuesday's v2.1.179 says the second-order priority while the Commerce negotiation is live is "don't break the installed base" — which is the right call when the prediction-market window is ~14 days. (2) The fact that the cadence held at all — Anthropic shipped three Claude Code releases in three working days while the flagship was offline — is the loudest disagreement with the "refused to fix" Friday framing. The team that wrote the classifier just shipped a third clean point release in 72 hours.

The clearest read on what an export-controls window does to OSS uptake — and the first time a single self-hostable agent crosses the 180k-stars threshold with a Skills Hub measured in five figures. Per The Agent Report Tuesday and the NousResearch/hermes-agent star feed, the repo sits at 188,781 stars as of Jun 10, up roughly 23,000 in two weeks; the Hermes Skills Hub now lists 90,881 skills across 12 registries, up from a few hundred a quarter ago. Cadence backs the curve: v0.16.0 "the Surface Release" shipped Jun 5 with 874 commits, 542 merged PRs, a native desktop app, browser admin panel, /undo, remote Hermes gateway connectivity and Simplified Chinese; the interim point releases are running weekly. Two reads. (1) The 23k-stars-in-two-weeks bump tracks the calendar of the Fable 5 rollout-and-rollback: builders who lost a Mythos-class default between Jun 9 and Jun 13 showed up at the closest self-hosted answer on the way out. The OSS counter-stack consolidates whenever the closed flagship blinks — and this week the blink lasted 72+ hours. (2) The 90,881-skills shape is the harder second-order tell. Skills are an Anthropic marketing primitive copied verbatim into the Hermes runtime; the catalog crossing five figures means the Hermes Skills Hub can be cited next to antigravity-awesome-skills and Claude Code's built-in catalog as an interchangeable surface, not a long-tail OSS curiosity.

The first G7 closing-day AI lunch with all three US frontier-lab CEOs at the same table, and the cleanest scheduled ratification of the "voluntary-commitments + infrastructure" template the post-Évian governance regime will inherit. Per the Élysée official brief, the Global Banking & Finance Tuesday writeup and the Euronews tracker, the Wed Jun 17 working lunch brings Sam Altman (OpenAI), Demis Hassabis (Google DeepMind), Dario Amodei (Anthropic), Arthur Mensch (Mistral), Aidan Gomez (Cohere), Robin Rombach (Black Forest Labs), Pratyush Kumar (Sarvam AI), Victor Riparbelli (Synthesia), Alexandr Wang (Meta), Marc Benioff (Salesforce) and Ren Ito (Sakana AI) to the same table as the G7 leaders, with OpenAI Chief Global Affairs Officer Chris Lehane publicly briefing a voluntary-commitments package on youth safety as the planned deliverable. Two reads. (1) The Amodei-and-Lutnick-both-confirmed-on-the- ground shape is the under-the-radar lift. The same week Bloomberg published a criminal-penalty Lutnick letter and the administration told The Telegraph "zero chance" on a UK exemption, both principals are in Évian for a working lunch — a face-to-face channel that is strictly higher-bandwidth than the Monday-virtual cadence. (2) The voluntary-commitments-as-deliverable framing is the structural read on what comes after the lunch. Voluntary commitments survive a Macron-Trump frame that a binding declaration cannot — which is the shape OpenAI's pre-summit youth-AI- safety institute push has been pitching for a week, and which the Évian communiqué can graft onto without a unanimous-consent hurdle.

The smaller second beat in the Cowork-GA cycle that recasts "Opus as the Cowork default" as one node in a polycentric model strategy rather than the Anthropic exclusive a casual read assumes. Per the Windows News companion to the Cowork-launch piece Tuesday, Microsoft is "exploring hosted DeepSeek for enterprises" in parallel with the Cowork GA — alongside the in-house Cowork 1 staged in the coming weeks and the Frontier-tier GPT-5.5 option. Two reads. (1) The hosted-DeepSeek pilot is the part the Évian sovereignty argument inherits next. A Microsoft-tenant-grade open-weights Chinese model hosted under M365 governance is the precise shape Carney's "over-reliance" frame wants to see — a tier-2 alternative behind the first-party agent surface. (2) The Opus-Cowork-1-GPT-5.5-DeepSeek quadrant tells you what Microsoft thinks the Mythos-class default for agents gap actually is: priced low enough to need a cheap in-house option, important enough to ship a frontier alternative on the Frontier tier, contested enough to investigate an open-weights non-US option in parallel. Five days into the Fable 5 suspension, that is the loudest "agent layer must be plural" statement any hyperscaler has made on the record.