Mythos and Fable go dark on Évian-eve
— Carney warns the G7 it's 2008.

The first time the US has used export controls on a deployed commercial AI model itself, not the chips it runs on, and the cleanest policy precedent of the cycle. The directive landed at 5:21 PM ET on Friday Jun 12: a letter signed by Commerce Secretary Howard Lutnick to Anthropic CEO Dario Amodei ordering that Claude Fable 5 and Claude Mythos 5 — the Jun 9-shipped Mythos-class frontier models — be cut off from any foreign national, inside or outside the United States, including Anthropic's own foreign-national employees, citing national-security authorities. Because Anthropic cannot reliably gate users by citizenship at API edges in real time, the practical answer was a hard worldwide shutoff: on Saturday Jun 13 the company posted that it had "disabled access to Fable 5 and Mythos 5 for all customers", with reinstatement contingent on negotiating modified safeguards. The government's framing came from David Sacks, chair of PCAST and the former White House AI & crypto czar, who posted on X that a "highly credible trusted partner of both Anthropic and the USG came forward with a jailbreak", that officials "requested Anthropic either fix the vulnerability or remove the model", and that "Dario refused". Anthropic's public counter: the demonstrated jailbreak surfaced only "a small number of previously known, minor vulnerabilities", the cited Mythos cyber capability — getting a model to read a codebase and find flaws — is a routine workflow, and the company blocks China at the product level. Two reads. (1) The export-control-on-the-model shape is the structural shock. BIS has spent two years drawing perimeter around Nvidia silicon and frontier compute clusters; pulling that same authority over the weights of a shipped model — and forcing the lab to take the product offline globally rather than implement a citizenship filter — re-prices every "compute-export-control"-style regulatory hedge that frontier labs have been signalling. The next time a lab ships above an undefined cyber/bio capability threshold, the Commerce letter is now a precedented instrument. (2) The same-week contradiction inside Anthropic's public stack is the under-the-radar political tell. Jun 10: Amodei's Bloomberg essay arguing the government should be able to block unsafe model deployments. Jun 13: Anthropic publicly objects when the government does that. The Jun 17 AI working lunch at Évian walks into both of those reads at once.

The pattern around why Commerce moved so fast, and the structural read on what an Anthropic backer-and-host relationship looks like when the capability question gets political. Fortune (Jun 14), TechCrunch (Jun 13) and Axios (Jun 12) converge on the same chain: Amazon researchers found a Fable 5 jailbreak that produced cyber-relevant content the model was supposed to refuse; CEO Andy Jassy raised it directly with Treasury Secretary Scott Bessent during a government feedback request; Bessent escalated to Lutnick; Lutnick's 5:21 PM ET Friday letter followed. Amazon's position is the structurally weird one — it has committed ~$13B to Anthropic, hosts Claude as its primary frontier inference workload on AWS Bedrock, and is one of two cloud venues where Anthropic enterprise customers reach the models. The same week Bedrock added Fable 5 to its catalog (carried by Claude Code v2.1.170 on Jun 9), Amazon's CEO was working the phones to take it off the menu worldwide. Two reads. (1) The investor-and-host-as-regulator-input shape is unprecedented at this scale. The public defence is straightforward — a jailbreak with cyber implications goes to Treasury because Treasury runs CFIUS-shaped national-security review on inbound AI risk — but the practical effect is that capacity-conflict incentives (Amazon ships Nova; Microsoft ships MAI; Google ships Gemini) now sit one phone call away from a Commerce action against a portfolio model. The next time a non-Amazon-built frontier lab ships an above-threshold capability into Bedrock, the precedent is in the room. (2) The Sacks-Anthropic public spat is the cleanest accelerant. Sacks: Anthropic "refused" to fix. Anthropic: the jailbreak shows "a small number of previously known, minor vulnerabilities" and the cyber capability Mythos demonstrated is a routine workflow. Both sides are now in the public record before the G7 has even opened — the AI working lunch on Wed Jun 17 walks into a live US-government-vs-frontier-lab dispute with no off-ramp staged.

The first time a sitting G7 head of state has framed concentrated frontier-AI dependence as a systemic economic risk, and the first multilateral pushback to the new export-controls-on-models regime. In Ireland on Sunday Jun 14, Canadian Prime Minister Mark Carney — Bank of Canada governor through the 2007–08 crisis response and the first non-British governor of the Bank of England — used the Fable 5 / Mythos 5 suspension to argue that overreliance on a small number of US AI providers creates model-risk -shaped systemic exposure analogous to 2008's concentrated bank linkages, telling press "it is never a good idea to have one option" ahead of his arrival at Évian. The same morning, EU Commission spokesperson Thomas Regnier said Brussels is examining the "practical consequences" for European users cut off from Fable 5 and Mythos 5 and warned that the US action "should not be discriminatory" against EU residents and businesses; the EU AI Act's general-purpose-model obligations bind on Aug 2. Two reads. (1) The former-central-banker framing is politically devastating. Carney is the one G7 leader whose job résumé is literally 2008-systemic-risk management; lining up "concentrated AI dependence" alongside "concentrated bank linkages" compresses an entire European-tech-sovereignty argument into a single image and ships it into the Jun 17 AI working lunch with Macron, von der Leyen and the new German chancellor as natural amplifiers. (2) The Regnier review is the second escalation channel. The EU has spent two years writing the AI Act around the premise that frontier-model availability is stable and the questions are transparency and red-team disclosure; the availability question is now back on the table, and the Aug 2 obligations could land in a regime where the US has demonstrably pulled EU access to a flagship model. The sovereignty argument Mistral's parallel ~€20B valuation talks rest on just got a precedent it can point at.

The first G7 leaders' summit where frontier-lab CEOs are present at the room, not consulted on the sidelines — and the first to open with a live US-government-vs-frontier-lab dispute on the table. 52nd G7 Summit runs Jun 15–17, 2026 in Évian-les-Bains, Haute-Savoie, hosted by President Emmanuel Macron (his second G7 after Biarritz 2019). Macron personally invited Sam Altman; Anthropic's Dario Amodei and Google DeepMind's Demis Hassabis round out the three-CEO seat at the Wednesday Jun 17 AI working lunch. The summit framing pre-Anthropic-pull was sovereignty-and-compute; post-pull, the unavoidable question is availability — when an ally pulls a flagship US model from EU and Canadian customers under national-security authorities, what guarantees does the "trusted ally" model leave for the rest of the room? Two reads. (1) Trump and Carney sit at the same table for the first time as PM-and-President with the Anthropic episode as a worked example of Carney's thesis. Indexbox and The Council on Foreign Relations both flag Iran-deal language as the second-largest agenda item; AI is the only topic with a fresh-this-week incident attached. The communiqué language on "international AI cooperation" will be parsed against Lutnick's letter, not the abstract version drafted in May. (2) The CEOs-in-the-room precedent is what survives the summit regardless of the communiqué. Once Altman, Hassabis and Amodei have sat at a G7 working lunch, the next AI safety event involves a phone tree with sitting heads of state on it — the "who do you call" question that has been answered by ad-hoc letters since 2023 now has standing infrastructure. The lunch is the institutional change even if no concrete commitment ships.

The Codex CLI alpha train rolled on through the Anthropic shutdown without a missed beat. openai/codex cut rust-v0.140.0-alpha.19 at 02:03 UTC on Sunday Jun 14 — running count nineteen alpha pre-releases on the v0.140.0 tag across Jun 9 → Jun 14, stable channel still pinned to rust-v0.139.0. Claude Code, meanwhile, has held at v2.1.177 since the Jun 13 01:25 UTC chore-only cut — the first multi-day gap in Anthropic's release feed since the Fable 5 RC train started on Jun 5. Two reads. (1) The asymmetry is the story this cycle. The same forty-eight hours in which Anthropic was negotiating a government-led shutdown saw OpenAI cut a substantive pre-release on Saturday (alpha.18, 17:30 UTC) and another on Sunday morning (alpha.19); the Codex changelog continues queueing Windows sandbox provisioning, app-server token-usage reads, CODEX_API_KEY remote-execution registration and OSC 8 markdown link rendering in the TUI. The harness shipping while the rival is arguing is the competitive read OpenAI gets free into Évian week. (2) The v0.140.0-still-pre-release shape also continues the buffering-as-moat pattern from the prior brief — nineteen alphas deep with no stable cut suggests the Codex team is now staging a substantial stable release for the post-Évian window, after the G7 communiqué clears the news cycle.

The hidden second policy event of the day that quietly re-prices every Claude-Code- driven agent build, and the first time Anthropic has explicitly unbundled programmatic from chat-shaped Claude usage at the billing layer. Announced May 14, effective today, Jun 15: Claude Agent SDK, claude -p, Claude Code GitHub Actions and third-party apps built on the Agent SDK move off the subscription rate-limit pool onto a dedicated, dollar-denominated credit metered at standard API list prices — $20/month for Pro, $100/month for Max 5x, $200/month for Max 20x. Credits are per-user (no team aggregation), don't roll over, and the claim is opt-in via an email Anthropic sent ahead of the cutover. Chatting with Claude on web / desktop / mobile, using Claude Code in the terminal, and Claude Cowork stay on the existing subscription limit. Optional usage-credits opt-in bills overflow at API rates rather than rejecting calls. Two reads. (1) The programmatic-is-now-a-meter separation is the structural shift. Until today, agent developers could ride a flat $200/month Max 20x ceiling for long-horizon Agent SDK work; from today, the same long-horizon run draws on a token-priced credit that exhausts inside hours on Mythos-class output if a workload is bursty. CI pipelines triggered by multiple committers are also worse off — credit can't aggregate across the team, so each user's slice applies only to their calls. (2) The timing next to the Fable-5-shutdown is brutal. The same day Anthropic loses its top model worldwide, the company also raises the effective price for the agent builders most exposed to Claude — and does it at a moment when Codex CLI's nineteenth alpha is buffering features for a stable cut. Migration pressure on the OpenAI Developer Account tier is higher this week than at any point since GPT-5.5 shipped.

The largest mission-critical-systems integrator on the planet pledging a Global Premier-tier commitment to Claude two days before the US government pulled the model — the kind of contract-cycle inertia that gives the ecosystem its real read on the next year of enterprise Claude spending. Jun 11: DXC Technology — $13B+ revenue, 140k+ employees, runs the mission-critical IT for many of the world's largest banks, airlines, insurers, manufacturers and government agencies — joined the Claude Partner Network as a Global Premier partner. The operational commitments: train tens of thousands of Claude-certified engineers through a 90-day Anthropic Partner Academy programme; deepen Claude integration inside DXC OASIS, an AI-native managed- services platform DXC says was built with more than 95% of its code generated by Claude before human review, accelerating delivery ~10×, currently deployed across 50+ customers and rolling out globally. Two reads. (1) The integrator-as-distribution shape is the actual enterprise lift. OpenAI sells seats; DXC sells multi-year managed- services contracts that the buying organisation cannot easily back out of mid-flight — every DXC client now in the pipeline for an OASIS deployment becomes a Claude end-user customer on a contract horizon that vastly outlasts an individual model version going offline. The Fable 5 shutdown does not move OASIS deployments that contracted on Claude-the-family rather than Claude-Fable-5. (2) The DXC-built-OASIS-on-Claude detail is the self-evidencing proof-point — DXC effectively published a case study where Claude wrote almost the entire codebase of a flagship managed-services product. That's a stronger sales artefact than any horizontal benchmark, and one DXC's competitors (Accenture, Cognizant, Infosys) will have to match before the next Premier announcement cycle.

The second Global Premier partner to land inside the same week as DXC, and the largest single-customer rollout of Claude on the operational side ever announced. Jun 11: Tata Consultancy Services — $30B+ in revenue, 600k+ employees, the largest Indian IT services firm and one of the largest global SI firms — announced a strategic global partnership with Anthropic, becoming the second Global Premier Partner in the Claude Partner Network and committing to roll Claude out to 50,000 associates across engineering, finance, legal, marketing and sales as the internal-experience baseline, then extend the model into jointly-developed AI solutions and services across financial services, public services, life sciences, healthcare, aviation, telecom and medtech via a dedicated Claude-powered solutions business unit. TCS shares dipped on the announcement, suggesting the market is pricing the operating-margin impact of large Claude usage rather than the top-line opportunity. Two reads. (1) The regulated-industries vertical naming is the structural tell. TCS handles core-banking modernisation for multiple top-10 global banks, insurance policy systems for several FTSE100 carriers and the back-office for two of the four largest sovereign-wealth-fund-aligned healthcare systems in Asia; surfacing Claude as the AI layer in those contracts — specifically in industries the EU AI Act and the Fed Reserve treat as high-risk — gives Anthropic regulator-recognised deployment scale that $30B+ of revenue alone does not. (2) The two-Premier-partners-in-72-hours pattern is the under-the-radar political shape. DXC + TCS together touch most of the world's major-bank-and-insurer IT estate; Anthropic just locked in integrator coverage for the exact set of buyers most exposed to Carney's "systemic-AI-dependence" read. The enterprise stack is signing up faster than the policy stack is pulling away.

The first time a major SCM has pulled agentic-coding-agent orchestration into its core CI surface as a first-class primitive, and the structural answer to "where does the agent run inside an enterprise" for the 180M-developer-account GitHub base. Jun 11: GitHub moved Agentic Workflows from February's tech preview to public preview. The product shape: author an automation in natural-language Markdown ("On every CI failure, triage the logs, draft a fix, open a PR"); Agentic Workflows compiles it to a standard Actions YAML; the workflow runs on the org's existing runner groups under its existing policy constraints. The same day, GitHub removed the PAT requirement — Agentic Workflows now uses the built-in GITHUB_TOKEN, removing the long-lived-PAT operational risk that has been the single largest objection from security teams. Two reads. (1) The compiles-to-Actions shape is the under-the-radar enterprise sell — every governance rail that today applies to Actions (env protection rules, OIDC federation, deploy environments, audit logs, restricted runner labels) applies automatically to agentic workflows. The "agent platform" tools (LangChain, CrewAI, Agno, OpenHands) all need a governance retrofit to clear an enterprise-security review; Agentic Workflows ships with the retrofit done. (2) The competitive read against the Linear / Jira agent integrations that crowded the brief last month is blunt — for most enterprises the code already lives in GitHub and Actions is already a budgeted line item; pulling issue triage, CI-failure analysis and doc updates into the same runner means the agent platform layer just lost a slice of the "build vs. buy" total addressable spend.

The Évian-week pairing that lets Macron walk into the AI lunch with a homegrown "the diversification already exists, we are funding it" answer to Carney's model-risk frame. Bloomberg reported on Jun 12 that French frontier-lab Mistral AI is in talks to raise approximately €3B at a valuation of around €20B — roughly 2.5× the €7.6B post-money from its October 2025 Series C. The positioning is unmissable: the round lands in the same window as the Anthropic shutdown and the EU Commission's "should not be discriminatory" warning, and Mistral is the EU's only full-stack frontier lab with model-and- deployment-and-data-residency control inside European jurisdictions. The ~$400M annualised run-rate from end-2025 plus a stated $1B ARR target for end-2026 sit underneath the valuation, suggesting public-market comparables rather than pure strategic-premium pricing. Two reads. (1) The capital-stack-as-sovereignty- hedge argument has been theoretical since Le Chat launched; the Anthropic shutdown made it concrete in a single weekend, and a fresh ~€20B mark turns the European tech-sovereignty pitch from a policy-paper into a valuation-comparable one investors can plug into a memo. Carney's "never a good idea to have one option" gets a homegrown G7-friendly second option publicly priced two days before the lunch. (2) The open-weights-as-export-control-immune read is the under-the-radar shape. Lutnick can write a letter to Amodei. He cannot write a letter to HuggingFace. Mistral's ongoing open-weight release line is now also a regulatory-availability hedge — the model Brussels and Ottawa keep access to even if Washington pulls the next closed-model export-control lever. That changes the strategic case for open weights independent of the technical case.