Day 13 — Anthropic names Alibaba and Claude Tag takes Slack
— while claude-code 2.1.190's binary leaks the Fable 5 return as weekly credits.

The cleanest political read on what the Mythos-directive negotiation looks like the day Anthropic stopped letting Commerce carry the foreign-national framing alone. Per Bloomberg's exclusive, CNBC's detailed follow, Yahoo Finance's market-reaction file and the PYMNTS recap, the letter is dated Jun 10 — two days before the Commerce directive surfaced — and was timed for a scheduled Senate Banking AI hearing. The quoted facts: ~25,000 fraudulent Claude accounts ran 28.8 million exchanges between Apr 22 and Jun 5; the operators are characterised as "affiliated with Alibaba and its AI lab" (the Qwen team); the surfaces targeted are software engineering and agentic reasoning — exactly the capabilities a Qwen-class model would need to lift to compete on SWE-bench and agent-loop axes. Anthropic's public statement, quoted by CNBC: "We believe combating the threat of illicit distillation requires coordinated action between government and industry, and we will continue working with Congress and the Administration to maintain American AI leadership." Two reads. (1) The letter-on-a-hearing-surface shape is the cleanest external read on how Anthropic wants the Mythos-directive aftermath framed. The same Day 12 Five Eyes statement that named Mythos and GPT-5.5-Cyber by model now lands with a specific China-attributed distillation incident as the evidentiary case — the capability-class argument Commerce made in private on Jun 12 now has a specific-actor argument on the public record. (2) The ~3%-stock-drop-the-same-day shape is the under-the-radar tell on the signalling cost Anthropic is willing to land on a customer — Alibaba is the operator of Alibaba Cloud, an Anthropic-model-availability surface and a counterparty on multiple joint enterprise pilots. Naming the customer in a hearing-direct letter is the clearest public signal that the regulation-track work has gone fully external, which fits the $20M+ PAC-war shape covered on Day 12.

The cleanest product read on what Anthropic wants the enterprise agent surface to look like in the Cowork-class world. Per the Anthropic announcement, VentureBeat's launch piece, Fortune's "virtual employee" framing and IT Pro's file, Claude Tag is in beta for Claude Enterprise and Team customers, runs on Opus 4.8, and rewrites the Claude-in-Slack primitive end to end. The behavioural shape is the headline. (a) Multiplayer: one Claude per Slack channel, not one per user — anyone in the channel can see what the agent is working on, redirect it, or pick up the conversation from where the last person left it. (b) Accumulating memory: Claude follows the channel as conversations unfold and builds an institutional context layer over time. (c) Ambient mode: when enabled, Claude does not wait to be tagged — it watches assigned channels and intervenes with a reminder, summary, or piece of cross-org context when it judges that intervention will help. (d) Async multi-step execution: users assign a task and walk away; Claude works through the steps using connected tools and posts results back to the thread. (e) Scoped admin controls: system administrators configure exactly which tools and which data Claude can touch in which channels — the surface a Fortune 500 CISO needs the day after a Commerce Department directive. The legacy Claude in Slack app will be retired on Aug 3; administrators have 30 days to opt in. Two reads. (1) The 65%-of-our-own-product-code-runs-on- this proof point is the cleanest external read on what Anthropic believes a Cowork-class agent is worth on a knowledge-worker surface; the disclosed dogfooding rate is now an S-1-track sales asset. (2) The Slack-as-the-agent-channel framing is the under-the-radar tell on where the buyer-side agent conversation has actually moved in 2026. The same week OpenAI shipped Workspace Agents with Slack integration and Viktor closed $75M for the agent-that-lives-in-Slack-and-Teams — covered in prior editions — Anthropic ships its own first-party answer. The Slack channel is now the de facto agent surface for enterprise multi-user work, and the freeze week is the week Anthropic bet the product roadmap on that surface.

The cleanest cadence read on what the Anthropic harness team is shipping inside the freeze. Per the anthropics/claude-code release page, v2.1.190 opened at Wed Jun 24 15:53 UTC (a ~18.8-hour turnaround from v2.1.187's Tue Jun 23 21:03 UTC) and v2.1.191 at 21:58 UTC a further ~6 hours later — two shipped cuts in a single calendar day. The headline ship is v2.1.191: (a) /rewind finally rescues state from before /clear, closing the only context-loss surface the slash command always had; (b) the sandbox network permission dialog — the surface that arbitrates allowed outbound hosts on the developer machine — now remembers hosts the user approves with "Yes" for the rest of the session, instead of re-prompting on every connection; (c) MCP capability discovery (tools/list, prompts/list, resources/list) retries transient network errors with short backoff; (d) MCP OAuth discovery + token requests retry once after transient errors, and headless environments skip the browser-popup path and go straight to the paste-the-URL prompt — a meaningful quality-of-life upgrade for CI/CD agent runs that don't have a browser; (e) the streaming-response loop coalesces text updates at 100ms, cutting CPU during streams by ~37%; (f) background agents stopped from the tasks panel no longer resurrect, and the agent-panel jump-by-row bug is squashed; (g) managed forceRemoteSettingsRefresh actually takes effect when set via MDM or file policy, and the fetch now sends Cache-Control: no-cache to prevent intermediate proxies from serving stale settings. Two reads. (1) The two-ships-in-one-day shape is the cleanest external read on what the harness team is doing with the inference plant's cycles while the Day-12 credentials-and-admin ship cures in. The /rewind landing is the only one of these that changes the workflow — the rest are reliability surface hardening, exactly what the post-freeze enterprise CISO buyer wants receipt for. (2) The MCP-OAuth- headless shape is the under-the-radar tell on where Anthropic's MCP surface lands on the CI deployment axis. The agent-on-a-server story has been pending an auth-without-a-browser primitive for months; shipping it quietly the day after the sandbox.credentials/ admin-can-ban-a-model triumvirate is the cleanest signal that the headless-server-agent use case is the 2027 roadmap track.

The cleanest restoration-shape read on what the negotiation between Anthropic and Commerce is actually trending toward — and the cleanest market read on how the Polymarket and Kalshi tape has revalued the deal since the Day-12 Five Eyes statement landed. Per Decrypt's Jun 24 piece, the v2.1.190 claude-code binary — cut ~6 hours before v2.1.191 — contains shipped strings that do not appear in the public changelog and were spotted by an external researcher inspecting the binary: "You've used your included Fable 5 usage for this week. Continuing on Fable 5 uses usage credits." The framing is consistent on three reads with the weekly-credit subscription tier shape Anthropic has been building since the Jun 15 credits overhaul: a tier where a Pro/Max/Team seat gets an included weekly Fable 5 quota and falls through to a usage-credit overage rate on exhaustion — a precise mirror of the Fri-night-billing-cliff the non-Fable models have been charging on since Jun 22 00:00 UTC. On the same day the strings surface, the Polymarket tape retreats: the Fable-5-restored-for-US- customers-by-Jun-26 contract sits at ~6%, collapsing from ~41% on Sat Jun 20 and ~73% on Thu Jun 18; the Mythos-5-access-restored-by book leads at "Jun 30" at ~18%. Two reads. (1) The strings-before-the-product shape is the under-the-radar tell on how Anthropic is negotiating: you do not ship binaries with weekly-credit-exhaustion copy if the restoration is not at least operationally pre-mortar. The biometric-KYC privacy-policy update effective Jul 8 (covered in prior editions) is the second leg; the third is the credit overage rate the strings reference. (2) The Polymarket-Jun-26-at-6% shape is the cleanest external read on what the deal-tracking market actually believes after the Jun-22 Five Eyes statement and the Jun-23 Claude outage stacked together — the by-Jul-1 book has retreated from the high-70s through the 40s across the freeze week, and the by-Jun-26 book is now an out-of-the-money option. The operational signal in v2.1.190's binary points past Jun 26, not before it.

The cleanest shipping-cadence read on what OpenAI's harness team is doing with the restoration window Polymarket still prices for Mythos. Per the openai/codex releases feed, after Tue Jun 23 18:56 UTC's alpha.9 (covered yesterday), the team cut alpha.11 at Wed Jun 24 01:17 UTC, alpha.12 at 02:23 UTC, alpha.13 at 03:24 UTC, alpha.14 at 05:42 UTC, and alpha.15 at 19:41 UTC — six pre-releases stacking onto the nine from Jun 23, for a total of fifteen v0.143 alphas in 44 hours. The alpha.9 → alpha.15 GitHub compare reports ~65 commits across 571 files; the v0.143 stable has not yet cut. Two reads. (1) The no-stable-yet shape is the under-the-radar tell on how OpenAI is sequencing this window. The team has shipped a stable roughly every 2–3 days through the first half of June; the v0.143 stable lane has now run ~52 hours past the v0.142.0 stable cut, the longest pre-release gap of the month. The clearest read is that the harness team is holding the code-mode/rollout-budget/ app-server-thread/items surfaces (merged in the early alphas) for a stable that lands when the Mythos-class capability returns and the timing makes them load-bearing. (2) The sub-hourly cadence into Jun 24 shape is the cleanest external read on what OpenAI believes the multi-harness Cursor/ Devin Desktop/claude-code consumer of the app-server contract is willing to integrate. Every consumer of the published contract gets the new surfaces in sub-hourly deltas without touching a stable until the team chooses — the harness layer is now continuous, with stables as annotated checkpoints rather than feature gates.

The cleanest flagship-cadence read on what OpenAI's release rhythm is doing the week the Mythos restoration is being priced. Per the FindSkill release-date tracker and the TechTimes June piece on chief scientist Jakub Pachocki's staff briefing, the public-record set on GPT-5.6 reads: (a) a May 13 routing entry to gpt-5.6 in the Codex backend, surfaced briefly and then withdrawn, with traffic reverting to gpt-5.5; (b) Pachocki's internal characterisation of it as a "meaningful improvement" over 5.5; (c) no blog post, no system card, no API model ID, no benchmark table — the most recent officially-documented OpenAI model remains GPT-5.5, shipped Apr 23. The market read: the Polymarket contract on a Jun 22–28 launch has collapsed from ~83% to ~18% across the last ~5 days; a July ship is now the median expectation. Two reads. (1) The no-stable-API-ID-yet shape is the cleanest external read on what OpenAI believes the flagship story should look like in Mythos-restoration week. A 5.6 ship during the same window where Anthropic's Mythos/Fable directive is still being negotiated would have produced capability-class press the Five Eyes statement made explicitly costly; a July slip lets the regulatory dust settle and the Patch the Planet / 5.5-Cyber defensive-AI framing carry the second half of June. (2) The 83%→18% reprice is the under-the-radar tell on how the Polymarket tape now reads OpenAI's sequencing — when the harness ships sub-hourly but the flagship does not move, the market reads the flagship as load-bearing on the restoration-window timing, not the capability readiness.

The cleanest config-plane read on what the multi-harness world is converging on as the buyer-side answer to the AGENTS.md/CLAUDE.md/ .cursorrules/.windsurfrules fragmentation. Per the Cursor Customize changelog entry, Cursor 3.9 on Mon Jun 22 lands a first-class Customize page that exposes plugins, skills, MCPs, subagents, rules, commands and hooks as a single config surface, scoped at the user, team or workspace level. The marketplace picks up a leaderboard of most-installed plugins/skills/MCPs across the team — the discovery layer the intellectronica/ruler trend made unavoidable. Plugin canvases ship as shared setup templates (the Hex Canvas for data visualisation, the Atlassian Canvas for realtime issue tracking), and team marketplaces now ingest plugin repos from GitLab, Bitbucket and Azure DevOps. The Bugbot line is the headline number: Composer 2.5 now powers Bugbot, dropping average review time from ~5 min to ~90 sec, finding +10% bugs per review (~0.62 from ~0.56), and cutting cost per run by ~22%. The /review slash command runs Bugbot + Security Review pre-push and dedupes against any same-diff Bugbot run already attached to the PR on GitHub/ GitLab. Two reads. (1) The Customize-page-as-a-config-plane shape is the cleanest external read on where the harness layer war converges in H2 2026: every harness is now competing on the configuration management axis, not on the agent loop primitive. (2) The Composer-2.5-powers-Bugbot shape is the under-the-radar tell on what Cursor's in-house model investment is starting to purchase. A ~3.3x speedup on the same QA loop at ~22% less cost is the cleanest possible proof point that the SpaceXAI-trained successor model (the 10x compute Composer 2.5) is now a legitimate tier-2 model in the Cursor stack.

The cleanest enterprise-content read on what Skills are actually for. Per the Box blog post on Tue Jun 23, Box AI now integrates with Anthropic's Claude Skills API end to end. The behavioural framing: a user prompts Box AI to consolidate a folder of meeting notes into a board deck, and the PowerPoint-creation Skill runs against the Box content the user has permission to see, generating the .pptx as a first-class Box file. The supported file types are PowerPoint, Excel, Word and PDF. The permission posture is the operationally important detail: the Skill respects the requesting user's existing Box access controls — the agent does not escalate privilege off the content graph. Two reads. (1) The same-Skills-as-Claude.ai-and- Claude-Code framing is the cleanest external read on Anthropic's portable-capability design bet. Skills are deliberately not Claude-product-bound; the same artefact runs in the consumer surface, the developer harness and the enterprise content plane, and the Box integration is the first content-vendor tier-1 wiring that proves the framing. (2) The Box-on-Tue + Claude-Tag-on-Tue pairing is the under-the-radar tell on Anthropic's enterprise-week sequencing. Slack is the multi-user conversation surface; Box is the multi-user content surface; both are now first-party Claude agent planes for the same buyer in the same calendar week.

The cleanest vertical-agent read on where a16z's enterprise-AI-doesn't-pause book is still leading checks while the harness layer is in regulatory turmoil. Per the Andreessen Horowitz announcement post, the GlobeNewswire release and the Unite.AI writeup, Probook closed $34M Series A led by a16z on Tue Jun 23, bringing total disclosed capital to $40M with the prior $6M seed led by Sequoia. The product framing is the headline: "AI operating system for home services", anchored at dispatch — the central operational decision the rest of the vertical-software stack ladders off. The decision logic is deliberately not proximity-first: Probook ranks technicians by skill, conversion rate and expertise for the specific job. Customer receipts: Summers Plumbing, Heating & Cooling (14 locations) booked 2,542 jobs in the first month with zero human intervention; Del-Air reports dispatcher productivity doubling from 10 to 22 technicians per dispatcher. Two reads. (1) The a16z-leads-the-A-after-Sequoia- seeds shape is the cleanest external read on where the vertical-agent operating system trade has actually clearing into 2026: a16z is treating it as a tier-1 portfolio bucket and is willing to write seed-priced terms per-vertical, not just horizontal. (2) The zero-human-intervention-on- 2,542-jobs-in-30-days shape is the under-the-radar tell on the buyer-side autonomy threshold in the home-services vertical: the buyer is willing to ship live customer-facing dispatch fully on agent output the first month of production. That is the fully-delegated-economic-work threshold the Cowork / Devin story has been trying to land at the enterprise tier; Probook has it already on a single vertical.

The cleanest agent-payments- substrate read on where Amplify and Kleiner want the 2027 agent-to-agent payments story to clear off. Per the BusinessWire release and Fortune's exclusive, Allium closed a $40M Series B led by Amplify Partners with Kleiner Perkins and Theory Ventures on Tue Jun 23, bringing total disclosed capital to ~$61.5M after the $16.5M Series A closed in 2024. The product framing is utility-shaped: process and standardise on-chain data from ~150 blockchains into a queryable analytics surface for institutional finance. The customer list is the headline credibility receipt: Visa (which built its Onchain Analytics Dashboard on top of Allium), Stripe, Coinbase and the U.S. Federal Reserve all cite or use Allium data. The agent-tier framing is explicit: Amplify partner David Beyer, quoted by Fortune, identifies AI-agent-driven blockchain payments as the "agentic piece" — the largest long-term growth opportunity for the position. Two reads. (1) The Visa/Stripe/Fed already on the customer list shape is the cleanest external read on what the agent-pays-agent stack actually requires under the hood: a data layer institution-grade enough for SIFI users, not a new payment rail. (2) The Stripe- style-Beyer quote shape is the under-the-radar tell on the framing convergence. Stripe has been arguing for ~12 months that blockchains/stablecoins are the cleanest agent-to-agent payment medium; Amplify priced $40M behind that thesis the same week Probook's zero-human-intervention receipts proved that agent economic action is now fully delegable on production verticals.