Day 15 — GPT-5.6 Sol ships into the Fable regime, Liccardo's deadline goes silent
— Mythos's NSA red-team surfaces, and Codex + Claude Code cut at the same minute.

The largest single-day move on the model-availability ledger of the year and the cleanest signal yet that the Fable-shaped regime is a frontier-launch default now, not a one-lab emergency. Per OpenAI's launch post, the CNBC, TechCrunch, VentureBeat and TheNextWeb files: Sol is the new flagship (with Terra as the balanced tier and Luna as the fast/affordable tier), evaluations highlight improved agentic capabilities in coding, biology and cybersecurity, and the new ultra mode goes "beyond the capabilities of a single agent by leveraging subagents to accelerate complex work" — the first production multi-agent surface shipped by a frontier lab at launch. Two reads. (1) The government-managed access list shape is the regime change: this is the Anthropic-Jun-12 directive's technical mechanism — a curated set of approved users — implemented as a company-side launch policy at a different lab, with no 14 C.F.R. § 744.22(b) "is informed" letter required. The line between export-control and responsible-deployment at the frontier collapses in this single paragraph. (2) OpenAI's "not the long-term default" disclaimer is the tell: a company that voluntarily accepted the regime is also publicly resisting it on the record — exactly what Anthropic's Jun-10 Senate Banking letter argued for the other lab two weeks earlier. The two frontier labs now share a public position on USG-mediated launch: comply where required, reject it as the default. The third lab — Google, whose Gemini 3.5 Flash native computer-use tool shipped on Day 14 with no analogous process — is the variable being watched.

The first congressional deadline of the freeze passes without a written Commerce filing on any of the four items the Jun-18 bipartisan letter named: the legal hook (formally 14 C.F.R. § 744.22(b)'s "is informed" mechanism under the Export Control Reform Act of 2018's emerging-and-foundational-technologies clause), the technical evaluation that drew the line, the review process, and the restoration / licensing criteria. Two reads. (1) The Polymarket-Jul-1-mid-sixties close at the deadline-day book is the cleanest external read on whether the market is pricing Tom-Brown-replaces-Dario as working: it is — the ~67% tag is the highest the Jul-1 contract has traded since the freeze opened on Jun 12, and the Jun-26 book itself closed in the low-teens, a clean resolves-NO. The market is betting the negotiation finishes on Jul 1's side of the calendar, not Jun 26's. (2) The no-Commerce-filing shape forces the discoverable-object question on the other side: with no written justification on the record, the 14 C.F.R. § 744.22(b) statutory hook stays a private-channel directive — exactly the surface Anthropic's Jun-10 Senate Banking letter argued Congress should not allow to set frontier-AI export precedent. The Jul-1 Polymarket book is now the next dated read.

The first time the actual technical finding behind the directive surfaces in attributable form. Per the AP file carried by CNBC, SecurityWeek, Tom's Hardware, TechSpot and Cybersecurity News, the Project Glasswing exercise was an Anthropic-led joint test with U.S. intelligence agencies that used Mythos 5 to look for weaknesses in classified federal computer systems; the model identified vulnerabilities in nearly all of them within hours, though identification is not exploitation. Two reads. (1) The Project Glasswing framing is the surface that complicates the jailbreak-was-the-problem narrative David Sacks ran on Day 1 of the freeze and revisited on Day 7: this is Anthropic-initiated, not adversary-found, and was designed to surface exactly the capability the directive treats as disqualifying. The conflict is now framed as capability, not misalignment — a much harder argument to settle with a safeguards-fix commit. (2) The Gen. Rudd-via-Sen. Warner attribution chain is the discoverable-object the Liccardo letter wanted on the Commerce side: the technical evaluation now exists in the Senate record at a hearing already covered by the Anthropic Jun-10 letter that opened the political phase of the freeze. The Polymarket Jul-1 contract must now price both the Tom-Brown-led negotiation and the fact that the technical-disqualification argument is public.

The industry-counter-pressure surface stays loud as the Project Glasswing disclosure lands. Per freefable.org, the Eastern Herald tracking piece and the Dark Reading security-community-slams file, the letter's argument is narrowly technical, not political: a Mythos-class model finding-nearly-all-NSA-classified in hours is not a unique-capability claim if other foundation and open-source models routinely do the same work in security audits and red-team training — the Glasswing result, on this read, demonstrates frontier-AI-as-tooling, not frontier-AI-as-weapon. The signature list compounds the U.S.-industry-capability argument that Anthropic's Jun-10 Senate Banking letter and the Liccardo deadline letter both lean on: the same companies the directive is meant to protect are publicly arguing the protection is misaimed.

The nothing-here stable cut is its own signal. Per the openai/codex release pages, rust-v0.142.3's notes carry exactly one line — "Maintenance-only patch release with no user-facing changes since 0.142.2" — and the tag is authored by github-actions, a pattern that on the public Codex ledger has historically only appeared on defensive patches (CVE-class fixes, dependency-bump triggers, supply-chain rotations). Two reads. (1) The no-changes-since-0.142.2 framing is a release-engineering tell: OpenAI rolled a tag through CI on a Friday at 21:29 UTC without any user-facing changelog — exactly the shape a silent-supply-chain rotation takes, and exactly what v0.143-alpha.26's 20:08 UTC pre-release on the same day suggests the team would otherwise have gone via the alpha train. (2) The same-minute-as-Claude-Code coincidence is the under-the-radar tell on harness-policy coordination surface: two competing frontier-AI shops shipping independent harnesses cut releases at the same minute on a Friday evening UTC, with the Claude Code tag fixing a security-class hyphenated-hook- matcher bug — the read is scheduled-coordination more naturally than coincidence.

The hyphenated-hook-matcher fix is the headline. Per the anthropics/claude-code changelog: hook matchers like code-reviewer or mcp__brave-search were silently doing a substring match rather than an exact match — meaning a policy that intended to scope mcp__brave-search would also have matched mcp__brave-search-pro, mcp__brave-search-internal and anything else with that prefix in its name, in either direction. The fix requires policies to switch to mcp__brave-search__.* to keep covering all tools from a hyphenated MCP server. Two reads. (1) This is a silent-overreach class of bug — the kind a Mythos-class auditor would flag as capability-leakage: policies were doing more than the operator thought they were, not less. Shipping the fix on the same day the Project Glasswing disclosure goes wide and on the same minute as the Codex maintenance-only patch reads as a coordinated-hardening cadence, not a routine ship. (2) The voice-dictation-auto-submit fix for Japanese, Chinese and Thai is the under-the-radar tell on Asia-tier usage: those three writing systems share the no-inter-word-space property the underlying audio model was gating auto-submit on. Shipping the fix the same week Anthropic's Seoul office continues onboarding (Day 12) is a coverage signal the harness team is investing in non-Latin-script input surfaces.

The instruction-following chapter closes on the consumer surface. Per OpenAI's ChatGPT release notes and the Jun 3 sunset announcement, GPT-4.5 was the bridge model between the instruction-following lineage (GPT-4o → GPT-4-turbo → GPT-4.5) and the reasoning-as-default generation that started with GPT-5.0; its exit from the consumer picker is the last non-reasoning GPT-4-series option going away on ChatGPT. Two reads. (1) Sub-routing-to-GPT-5.x-default lands the same week GPT-5.6 Sol is in restricted preview, which makes Sun Jun 28 the first day the entire ChatGPT consumer base is on reasoning-default models — with the next-flagship surface gated behind a USG-trusted-partners list. The contrast is the read. (2) The o3-Aug-26 retirement on the same template is the release-engineering tell on how OpenAI is sequencing consumer-side cleanup ahead of the EU AI Act Aug-2 GPAI enforcement cutover and the GPT-5.6 family GA window.

The single largest AI-infrastructure round of the week. Per the BusinessWire release, the TechCrunch file and the Verdict / Startup Fortune coverage, the two-tranche-at-$13B-and-$11B structure is the growth-equity signature on a round that priced upmarket mid-process — Sands Capital and Wellington Management co-leading is a public-market-tier ticket on what is still a private-stage infrastructure company. Two reads. (1) The 1B-inference-calls-per-day volume is the cleanest production-scale reference for multi-model inference serving in U.S. enterprise — and lands the same week the OpenAI/Broadcom Jalapeño tape-out tells the market OpenAI is building first-party inference silicon. Baseten is being valued as the multi-tenant alternative to the lab-owned inference path OpenAI just declared on. (2) The 20×-revenue-growth shape is the justify-the-multiple data point — at $13B, the round prices Baseten as a Stripe-of-inference bet, the middleware layer that converts model weights into tokens-per-second-per-dollar at production grade.

The spatial-temporal-reasoning bet of the week. Per the Axios, GamesBeat, PitchBook and The AI Insider files, the thesis is narrowly contrarian: the action-labeled-gameplay dataset — where each frame carries the player input that produced the next frame — is the cheapest large-volume source of action-conditioned video on the public record, and Fortnite-to-quadruped transfer on the same weights is the existence proof that action-conditioning generalises across embodiment classes. Two reads. (1) The Khosla-leads / Bezos-Schmidt-participate signature on a $320M-at-$2.3B-Series-A is the frontier-bet shape: the same investor-class concentration Cognition and Sierra drew in 2026, on a different-thesis founder. (2) The CoreWeave-handles-pre-train detail is the infrastructure tell that pairs with the Baseten-$1.5B inference round: the foundation-model tier is not what's being commoditised this year — what's being verticalised is the data-source and the action-conditioning objective, and the infra stack underneath them.

The simulation-of-the-world-the- agent-will-act-in bet, closing the week's evaluation-tier capital. Per the PRNewswire release, the TechCrunch file and the SiliconANGLE production read, Patronus's pitch is a category step up from Day 14's Coval bet: where Coval simulates voice-agent-call-branches, Patronus's Digital World Models simulate the entire web app the agent acts on — a replica of the corporate surface, generated by a language diffusion world model, that an agent can hit with RL-during-training instead of only at eval. Two reads. (1) The 15×-revenue-growth shape on a $50M-out-of-$70M-total financing puts Patronus at the defining-capability moment of the evals tier — the moment a category shifts from "we wrap your model with assertions" to "we build the environment your agent learns in". (2) The most-frontier-labs-as-customers disclosure pairs naturally with the week's GPT-5.6-Sol and Project-Glasswing headlines: the labs themselves are moving the evaluation surface upstream of release, and Patronus is one of the two or three vendors that surface routes through.