Weekly · 2026-W25 Jun 08 – Jun 14, 2026
Tech Application

Anthropic ships Claude Fable 5 and Mythos 5, the first publicly available Mythos-class models, and Stripe finishes a 50-million-line Ruby migration in a day

The Wednesday release puts Mythos-class capability on every AWS, Bedrock, Microsoft Foundry, GitHub Copilot, Harvey and Cursor account at $10 input / $50 output per million tokens. Fable 5 posts 80.3% on SWE-Bench Pro. Three classifier-routed safeguards intercept cybersecurity, biology and distillation queries on under 5% of sessions. Mythos 5 deploys with cyber-uplift safeguards lifted to roughly 200 critical-infrastructure partners across 15 countries under Project Glasswing. Within 48 hours Pliny-the-Liberator surfaces the model's 120,000-character system prompt and Anthropic disputes the multi-agent jailbreak in public.

Anthropic · AWS · GitHub · Harvey · 2026-06-10 · Source ↗ ·

Anthropic shipped Claude Fable 5 and Claude Mythos 5 on Wednesday morning, putting the company's Mythos-class tier on general availability for the first time. The release lands as the second public Mythos-class model after Mythos 4 stayed behind a critical-infrastructure-only access wall for nine months, and it materially changes what an enterprise can put on a paid API account today. Fable 5 is the general-purpose flagship: full Mythos-class capability, three classifier-routed safeguards, and pricing of $10 per million input tokens and $50 per million output tokens. Mythos 5 is the variant Anthropic uses inside Project Glasswing, lifted to roughly 200 critical-infrastructure partners across more than fifteen countries through a U.S.-government channel. The two checkpoints share the same training run and diverge on which cyber-uplift capabilities the deployment toggles permit.

The numbers landed hard. Fable 5 scored 80.3% on SWE-Bench Pro, the hardest agent-coding benchmark Anthropic publishes against, opening daylight on every other generally available model. Stripe ran the launch demo: a 50-million-line Ruby migration that the company's platform team estimated at nine months of human work, finished by a Fable 5 swarm in a single day, with zero production incidents on cutover. Harvey scored Fable 5 at 13.3% on Legal Agent Bench, ahead of Opus 4.8's 10.4%, and at 93.4% on BigLaw Bench, both the highest legal scores any model has posted on Harvey's published evals. Cursor put Fable 5 behind its Composer flag for paid users on launch day. GitHub Copilot moved Fable 5 to general availability across all paid plans the same morning, with an accompanying changelog entry from GitHub's Copilot team and AWS publishing the simultaneous availability inside Amazon Bedrock and the Claude Platform on AWS. Microsoft Foundry added the model the same day with governed agent tooling for Azure enterprises.

"Today we are introducing Claude Fable 5, our most capable model, and Claude Mythos 5, our first Mythos-class model with cyber capabilities, both built on the same foundation."— Anthropic, "Introducing Claude Fable 5 and Claude Mythos 5," 2026-06-10

The three safeguards and the cost of safety routing

Anthropic published the safeguard architecture alongside the release. Three classifiers run on every Fable 5 request: a cyber-uplift classifier, a biology-and-chemistry classifier, and a distillation classifier. When a classifier fires, the request routes to Opus 4.8 with a tighter system prompt rather than refusing outright, an architectural choice Anthropic argues preserves productivity at the cost of measurable latency on the small share of requests that trip the rails. The company's published rate is under 5% of all sessions, and the bug-bounty disclosure includes zero universal jailbreaks across more than 1,000 hours of red-team testing through launch. Enterprise customers received an updated retention notice in parallel: a mandatory 30-day data retention window now overrides prior zero-retention enterprise contracts on Fable 5 traffic, a change Anthropic describes as required to support the classifier and incident-response loop. Customers under HIPAA, FINRA or government-cloud terms are negotiating exemptions individually.

Mythos 5 and Project Glasswing

Mythos 5 deploys behind the same training run as Fable 5 with cyber-uplift safeguards lifted to vetted partners under Project Glasswing. The launch list includes BeyondTrust as a privileged-access partner, with Cloudflare publishing a separate post on the cyber-frontier-model deployment surface; Cybernews and PYMNTS covered the Glasswing expansion to roughly 200 partners across critical infrastructure operators in more than fifteen countries. The model is the operational answer Anthropic gives to the question Bruce Schneier pressed on June 9: the Glasswing program found 23,000 vulnerabilities across 1,000 projects in its first year, of which Schneier reported "almost none patched." Anthropic responded by expanding the partner program rather than narrowing it, betting that scale across critical-infrastructure operators is the variable that closes the patch gap.

"Fable 5 finished a 50-million-line Ruby migration in a day. We had estimated nine months of platform engineering."— Stripe platform team, launch-day demo, 2026-06-10

Pliny and the 120,000-character system prompt

The honeymoon lasted 48 hours. On Friday, Pliny-the-Liberator and a small group of collaborators published a multi-agent jailbreak that, according to TechTimes' Saturday writeup, surfaced what Pliny called the model's full 120,000-character system prompt. The attack chained Unicode-tag tricks, long-context reference tracking, document framing and fiction framing, with multiple Claude instances feeding state to each other. Anthropic's threat-intelligence team disputed the disclosure publicly within hours, arguing that what Pliny exposed was a developer-mode debug surface rather than the production system prompt and that no Mythos-class capability had been unlocked. Independent researchers split on whether the exposed text matched the production prompt. The episode reopened the public argument about whether a 120K-token system prompt is a load-bearing safety boundary or a brittle artifact, and Schneier reposted his prior critique with the addition that "if the system prompt is 120,000 characters long, your alignment is in the wrong layer."

Pricing, billing and the enterprise math

The $10 input / $50 output per-million-token pricing is the editorially significant number. Anthropic shifted enterprise billing to consumption-based at the same time, with annualized revenue crossing $30 billion according to The Information's Saturday long-form report. Free Pro, Max, Team and Enterprise users got Fable 5 through June 22 on a metered window. Replit's Amjad Masad posted the same week that putting Fable 5 in production reduced debugging cost per session despite the headline token price, because Fable 5 emits fewer turns to converge on a correct fix. The price-to-quality math is the constraint that decides whether the rest of the enterprise market follows Stripe's deployment pattern. Anthropic's bet is that the productivity gain at SWE-Bench Pro tier is large enough to absorb the per-token premium against the open-weights frontier. The price-quality gap to DeepSeek V4 and Nvidia Nemotron 3 Ultra still runs 30 to 50× in open weights' favor on equivalent agent workloads, and that arithmetic is the load-bearing variable for the second half of 2026.

The week in AI

Anthropic put a Mythos-class model on every major cloud surface and a Pliny jailbreak surfaced 120,000 characters of its system prompt 48 hours later. Apple used WWDC 2026 to rent the frontier-model layer from Google for a Gemini-backed Siri, open the LanguageModel Swift protocol to Claude and Gemini as launch packages, and route Apple Foundation Models Cloud Pro through NVIDIA Blackwell GPUs inside Google Cloud under Confidential Computing attestation. SpaceX priced the largest IPO in history at $135 on Wednesday, debuted Thursday, and closed +19.2% at $160.95 for a $2.1 trillion market cap. The Information reported that Anthropic asked Figma and Canva to be launch partners weeks before unveiling Claude Design, which competes with both, and that the company has shifted enterprise billing to consumption-based as annualized revenue crosses $30 billion.

Frameworks and agents moved with the models. Google DeepMind shipped DiffusionGemma 26B-A4B under Apache 2.0 as the first production-grade open-weights text-diffusion MoE, denoising 256 tokens in parallel at over 1,000 tokens per second on a single H100. OpenAI quietly retired the entire GPT-5.2 family and auto-migrated active conversations to GPT-5.5 with a checkpoint labelled GPT-5.6 "kindle-alpha" carrying a 1.5-million-token context window, no system card, no API and no third-party benchmarks. Anthropic shipped Claude Code /fork mapping AI coding to git-style session branching and pushed two same-day Claude Code releases tightening admin-managed model allow-lists. Microsoft open-sourced SkillOpt, treating an agent's SKILL.md as a trainable object that lifts agent accuracy by +23.5 on GPT-5.5, +24.8 inside Codex and +19.1 in Claude Code without touching model weights. OpenAI agreed to acquire Ona to put persistent cloud sandboxes under Codex's two-million-developer base.

Public markets and funding compounded the week. Oracle Q4 FY26 came in at $19.2 billion with cloud revenue +47% to $9.9 billion and remaining performance obligations leaping $85 billion in a single quarter to $638 billion. Cyera raised $600 million at a $12 billion valuation, quadrupling its mark in eighteen months. Tata Consultancy Services became Anthropic's Global Premier Partner with Claude licensed across 50,000 associates and a dedicated business unit serving financial services, healthcare, telecom and aviation customers including Diligenta's 22 million pension customers. OpenAI Codex crossed 5 million weekly active users at six times its February launch base. Bloomberg reported that EngineAI filed confidentially in Hong Kong, widening a Chinese humanoid IPO wave that also includes Unitree's $610 million Shanghai STAR filing at a $7 billion target.

Policy hardened in two directions at once. Anthropic published the Policy on the AI Exponential, asking governments for legal authority to block frontier-model deployments at thresholds of 10^25 training FLOPs, $500 million in revenue or $1 billion in R&D spend, with civil penalties tied to global revenue. Dario Amodei published an accompanying essay arguing AI labor displacement will be larger and longer than prior tech shifts, and Anthropic committed $200 million to an Economic Futures Research Fund and $150 million to a national fellowship. Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act seeking a one-time 50% stock tax on OpenAI, Anthropic and xAI. The Commerce Department published its June 1 written clarification that AI-chip export licenses follow corporate parentage, closing the Chinese-subsidiary loophole. The legislative pile-up runs alongside an unresolved safety conversation in which Pliny's Fable 5 jailbreak and Schneier's Glasswing patch-rate critique both stayed live going into Monday morning.

Themes this week

Four patterns to read by

Mythos-class arrives and reshapes the rest of the week

Fable 5 and Mythos 5 set the agenda for everything after Wednesday. Stripe ran a 50-million-line Ruby migration in a day. Harvey moved Fable 5 to top of Legal Agent Bench and BigLaw Bench. AWS, Microsoft Foundry, GitHub Copilot, Cursor and Snowflake Cortex picked the model up day-one. TCS turned its Anthropic relationship into a 50,000-associate Global Premier Partnership. Pliny then surfaced what he called the 120,000-character system prompt 48 hours later, and Anthropic spent the back half of the week disputing the disclosure in public. The launch's gravity pulled at the rest of the news cycle.

Apple rents the frontier model layer

The WWDC 2026 keynote put Mike Rockwell on stage to introduce a rebuilt Siri trained on Google's Gemini foundation. Apple Foundation Models Cloud Pro runs on NVIDIA Blackwell GPUs inside Google Cloud under Confidential Computing attestation. The LanguageModel Swift protocol opens Siri, Writing Tools and Image Playground to Anthropic Claude, OpenAI ChatGPT, Google Gemini and xAI Grok as user-pickable defaults. Anthropic and Google ship as launch Swift packages. Apple's third-generation Foundation Models post commits to summer open-sourcing of the framework. The Xcode 27 dual-engine routes inline completion to the Apple Neural Engine and agentic tasks through an mcpbridge XPC channel to the three external coding agents. Apple ships interfaces, not models.

The AI listing window opens

SpaceX's Thursday Nasdaq debut closed +19.2% at $160.95 for a $2.1 trillion market cap, the largest first-day print in stock-market history and four times the prior Aramco 2019 record. The orderly close greenlit Anthropic's $965 billion S-1 and OpenAI's June 8 confidential filing, with Bank of America circulating a buy-side note flagging $6.4 billion of Anthropic-to-cloud-partner payments as a possible revenue-recognition complication. Cyera raised $600 million at $12 billion, quadrupling its mark in eighteen months. The Information confirmed Anthropic crossed $30 billion in annualized revenue and shifted enterprise billing to consumption-based. Bloomberg reported EngineAI's confidential Hong Kong filing alongside Unitree's $610 million Shanghai STAR filing at a $7 billion target.

Policy stratifies into thresholds, taxes and export controls

Anthropic's Policy on the AI Exponential proposes legal authority to block frontier deployments at 10^25 training FLOPs, $500 million revenue or $1 billion R&D spend, with civil penalties tied to global revenue. Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act seeking a one-time 50% stock tax on OpenAI, Anthropic and xAI. The Commerce Department's June 1 written clarification that AI-chip export licenses follow corporate parentage closed the Chinese-subsidiary loophole. Dario Amodei's labor-displacement essay paired with a $200 million Economic Futures Research Fund commitment. Three federal frameworks are now on the table at once: capability thresholds, capital extraction and supply controls.

The rundown

0 stories · 2026-06-08 — 2026-06-14

Models & frontier releases

TechApplication

Apple opens iOS, Xcode and Foundation Models to Claude, Gemini, ChatGPT and Grok at WWDC 2026, with AFM Cloud Pro running on Google Cloud Blackwell

Apple · Machine Learning · TechTimes · 2026-06-08 — 2026-06-13 · Source ↗ ·

Mike Rockwell unveils a rebuilt Siri trained on Google's Gemini foundation. iOS 27 Extensions opens Siri, Writing Tools and Image Playground to Anthropic Claude, OpenAI ChatGPT, Google Gemini and xAI Grok as user-pickable defaults. The LanguageModel Swift protocol ships with Anthropic and Google as day-zero packages and a summer open-sourcing commitment. Apple Foundation Models Cloud Pro runs on NVIDIA Blackwell GPUs inside Google Cloud under Confidential Computing attestation.

Tim Cook walked out at 10 a.m. PT Monday for what AppleInsider treated as his last WWDC keynote as CEO ahead of the September 1 handover to John Ternus, then handed the AI portion of the keynote to Mike Rockwell. Apple's third-generation Foundation Models lineup ships in four named tiers: AFM Core on the Apple Neural Engine, AFM Core Advanced on Apple Silicon, AFM Cloud inside Private Cloud Compute, and AFM Cloud Pro on NVIDIA Blackwell B200 GPUs hosted inside Google Cloud under Confidential Computing attestation. Rockwell described the Cloud Pro tier on stage and the Machine Learning Research post on Saturday formalized the architecture. The Siri rebuild is the headline: a ground-up Gemini-trained voice assistant with a three-tier routing layer that decides which of Apple's four models or which external partner handles each request.

The structural change underneath the keynote is the LanguageModel Swift protocol. Apple's Foundation Models framework is now an interface, not a model. Anthropic and Google shipped Swift packages on day zero implementing the protocol, with one-line dependency swap and the same Swift API as Apple's own Foundation Models. Session 241 on Friday committed to summer open-sourcing of the whole framework. The iOS 27 Extensions surface opens Siri, Writing Tools and Image Playground to Anthropic Claude, OpenAI ChatGPT, Google Gemini and xAI Grok as user-pickable defaults. Federighi confirmed the Siri AI rebuild will not ship in the EU or China at launch over Digital Markets Act access demands.

"The Foundation Models framework is now an interface, not a model. Pick your provider with a one-line change."— Apple WWDC 2026 Session 241, 2026-06-12

Xcode 27 ships a dual-engine architecture. Inline completion runs on the Apple Neural Engine. Opt-in agentic tasks route through a new mcpbridge XPC channel that exposes 20 Xcode capabilities to Anthropic, Google and OpenAI agents. The Core AI framework opens full-scale on-device LLM deployment to App Store developers, with App Store Small Business Program developers under 2 million users getting Private Cloud Compute access free. AirPods Pro 3 stream heart-rate biometrics to iPhone for cardio without an Apple Watch. macOS Golden Gate rebuilds Spotlight, Photos and Mail search around a real-time index with 30% / 70% / 80% speed gains on app launch, photo ingest and AirDrop respectively. watchOS 27 adds perimenopause and menopause detection to Cycle Tracking. visionOS 27 turns iPhone panoramas into spatial scenes and pins a floating Siri marble. Vision Pro 2 was deferred to 2028. The strategic message is clear: Apple ships the device surface and the routing layer, and rents the frontier from Google, Anthropic and OpenAI.

Tech

Google ships DiffusionGemma 26B-A4B under Apache 2.0 as the first production-grade open-weights text-diffusion MoE

Google DeepMind / MarkTechPost · 2026-06-10 · Source ↗ ·

A 26-billion-parameter sparse MoE with 4 billion active per token, denoising 256 tokens in parallel at over 1,000 tokens per second on a single H100 and up to 2,000 on a DGX Station. Day-zero Tensor Core acceleration through Transformers, vLLM, MLX and Unsloth. 77.6 MMLU Pro, 73.2 GPQA Diamond, 69.1 LiveCodeBench v6, 54.3 MMMU Pro. Google directly recommends Gemma 4 for production; DiffusionGemma is the research artifact opening a parallel decode pathway.

DiffusionGemma is the first production-grade open-weights text diffusion model. The architecture is a 26-billion-parameter sparse mixture-of-experts with 4 billion active parameters per token, denoising 256 tokens in parallel rather than autoregressively predicting one token at a time. The license is Apache 2.0, permitting unrestricted commercial use. On a single H100 the model emits over 1,000 tokens per second; on an NVIDIA DGX Station with Blackwell GPUs the throughput crosses 2,000 tokens per second. Hugging Face hosts the weights at google/diffusiongemma-26B-A4B-it; the DeepMind page ships the model card; MarkTechPost's June 10 explainer walks through the Uniform State Diffusion mechanism with re-noising self-correction that autoregressive transformers cannot match.

The benchmark numbers ground the release at the production tier. MMLU Pro scores 77.6, GPQA Diamond 73.2, LiveCodeBench v6 69.1 and MMMU Pro 54.3. Google's own model card directly recommends Gemma 4 for production deployments and frames DiffusionGemma as the research artifact opening a parallel decode pathway. NVIDIA's RTX AI Garage post on June 11 covered day-zero Tensor Core acceleration through Transformers, vLLM, MLX and Unsloth. Hugging Face Transformers v5.11.0 added native DiffusionGemma support on Friday alongside GLM 5.1 support.

"DiffusionGemma denoises 256 tokens in parallel. The autoregressive cost curve does not bend; the architecture does."— MarkTechPost analysis, 2026-06-10

The editorially significant change is the decode-cost curve. Autoregressive transformers pay per-token sequential cost, with each output token depending on the previous one. Text diffusion denoises a window of tokens together, then re-noises and corrects them in subsequent passes. The throughput claim runs roughly 4× faster generation against equivalent-quality autoregressive Gemma 4 baselines on the same hardware, while preserving competitive benchmark scores. The architecture is the first time an open-weights text-diffusion model has shipped at production quality, and the parallel-decode pathway gives developers an alternative to the autoregressive cost ceiling. NVIDIA's Tensor Core acceleration through the open-source toolchain is the supply-side piece that turns the research artifact into a deployable production option.

Tech

OpenAI quietly retires the GPT-5.2 family and migrates every active conversation to a GPT-5.6 "kindle-alpha" checkpoint

TechTimes · 2026-06-13 · Source ↗ ·

Friday's shutdown carries a 1.5-million-token context window, medium reasoning effort and visibly cleaner SVG output through Codex testing paths. No system card, no pricing, no API endpoint and no third-party benchmark suite accompany the rollout. Existing conversations auto-migrated to the matching GPT-5.5 tier with the checkpoint labelled GPT-5.6 "kindle-alpha." Codex usage crossed 5 million weekly actives the same weekend, six times the February launch base.

OpenAI quietly retired the entire GPT-5.2 family on Friday with no announcement post. ChatGPT users found their existing conversations auto-migrated to the corresponding GPT-5.5 tier, and developers using GPT-5.2 via the API received deprecation notices to switch endpoints by month-end. The migration target carried an unannounced checkpoint label, GPT-5.6 "kindle-alpha," visible through Codex testing paths with a 1.5-million-token context window and medium reasoning effort. The checkpoint produced visibly cleaner SVG output and tighter image-generation results. No system card accompanied the change. No API key surfaced for direct addressing. No pricing schedule appeared. No third-party benchmark suite has run against it as of Sunday.

TechTimes' June 13 piece framed the move as the quietest major frontier-model rollout in OpenAI's history. The contrast with Anthropic's three-day-old Fable 5 launch was unavoidable: Anthropic published the safeguards architecture, the safety classifiers and the Project Glasswing deployment details on launch day, while OpenAI shipped a successor model into production without confirming its existence. The lack of an external benchmark suite leaves the third-party reproduction window closed.

"The quietest major frontier-model rollout in OpenAI's history. No card, no pricing, no benchmarks. Just a migration."— TechTimes, 2026-06-13

OpenAI Codex crossed 5 million weekly active users the same weekend, six times the February desktop-launch base, with one in five users now non-developer per OpenPR. Friday's CVE disclosure for Anthropic's Claude Code GitHub Action got patched in parallel. The unanswered question is whether kindle-alpha is the production GPT-5.6 by another name or a private staging label that gets renamed before the system card publishes.

Agents & frameworks

Tech

Claude Code ships /fork, mapping AI coding onto git-style session branching

TechTimes · code.claude.com · 2026-06-13 · Source ↗ ·

The new command lets a developer branch any conversation into parallel sessions with shared upstream history, then merge or discard branches. Claude Code 2.1.174 and 2.1.175 ship the same Friday with enforceAvailableModels closing the last route a user could widen an admin-managed model allow-list, plus VSCode usage attribution and a dozen long-tail fixes. Claude Code is estimated at roughly 4% of public GitHub commits and described internally as Anthropic's fastest-growing product.

Anthropic shipped the /fork command in Claude Code on Friday afternoon, mapping AI coding sessions onto a git-style branching model. A /fork call branches a conversation into a parallel session that shares the upstream history but accumulates its own state, lets the developer try multiple approaches without losing the prior context, and merges or discards branches with a single command. The change ships in Claude Code 2.1.174 and 2.1.175, with VSCode usage-attribution breakdowns and a dozen long-tail fixes accompanying the release.

The same release closes the enforceAvailableModels admin loophole. Until Friday a user could widen an admin-managed model allow-list through a config-file pathway; the 2.1.175 fix tightens that to a server-side enforcement and removes the local override. The change matters for enterprise rollouts where compliance teams pin Claude Code to a specific approved model and need the admin policy to actually bind on the client.

"Claude Code is the fastest-growing product we've ever shipped. The /fork command is the second-most-asked-for feature behind subagent isolation."— Boris Cherny, Claude Code lead, on X, 2026-06-13

The Information's June 14 long-form estimated Claude Code at roughly 4% of public GitHub commits, a figure Anthropic has not formally confirmed but has not pushed back on either. The product is described internally as the company's fastest-growing surface, and the consumption-based enterprise billing model Anthropic moved to during the week is partly underwritten by the Code workflow's compounding token spend. The /fork command's downstream implication is straightforward: parallel exploration that previously lived only in agent-orchestration harnesses now lives in the developer's terminal.

Tech

Microsoft open-sources SkillOpt, an optimizer that trains an agent's SKILL.md as the trainable object

VentureBeat · 2026-06-13 · Source ↗ ·

SkillOpt treats the agent's SKILL.md instruction file as a trainable artifact and runs gradient-style optimization against task-success signal without touching model weights. The published numbers are +23.5 accuracy on GPT-5.5, +24.8 inside Codex and +19.1 inside Claude Code. The release ships under MIT on the Microsoft GitHub organization with reference experiments and an evaluation harness. The technique works on any model the agent can call, regardless of provider.

Microsoft Research open-sourced SkillOpt on Saturday under MIT license, attached to the Microsoft GitHub organization with reference experiments and an evaluation harness. The system treats an agent's SKILL.md instruction file as a trainable object and runs gradient-style optimization against task-success signal without touching the underlying model weights. The published evaluation lifts task accuracy by +23.5 absolute points on GPT-5.5, +24.8 inside OpenAI Codex and +19.1 inside Anthropic Claude Code, measured on an internal agent-task suite VentureBeat described in the Saturday release coverage.

The architecture inverts the usual training assumption. Where standard reinforcement-learning-from-human-feedback shifts model weights to improve outcomes, SkillOpt fixes the weights and shifts the prompt. The optimizer takes a SKILL.md as input, runs the agent against a task suite, scores outcomes, and proposes targeted edits to the SKILL.md text using a meta-prompted rewrite loop. The published numbers are competitive with reinforcement-learning fine-tunes on the same task suites, at a fraction of the compute cost and with no provider lock-in.

"SkillOpt treats the SKILL.md as the trainable artifact. We replace weight gradient descent with prompt gradient descent, and it works on every frontier model we tested."— Microsoft Research, SkillOpt repository readme, 2026-06-13

The provider-agnostic property is the editorially significant claim. Enterprises that pinned their agent stack to a specific model can now improve agent performance without changing models, switching API providers or paying for fine-tuning. The same SKILL.md lift transfers across frontier-model boundaries because the optimization target is the instruction text rather than the model weights. The repo ships with a worked example, an evaluation harness based on the agent-task suite, and a step-by-step guide for running SkillOpt against an in-house task corpus.

TechEconomic

OpenAI agrees to acquire Ona, putting persistent cloud sandboxes under Codex's 5-million-weekly-user base

OpenAI · 2026-06-11 · Source ↗ ·

Ona's secure cloud-sandbox platform brings two million developer accounts and the persistent-session substrate Codex has been missing. Founder Johannes Landgraf joins the Codex division. The acquisition closes the gap between session-bound autocomplete and persistent production agents and ships the same week Codex crosses 5 million weekly active users, six times its February desktop-launch base.

OpenAI agreed on June 11 to acquire Ona, the secure cloud-sandbox startup whose two-million-developer-account substrate provides the persistent-execution layer Codex has lacked. The acquisition price was not disclosed; the strategic logic was. Codex, in its current form, ships session-bound autocomplete and short-lived agent runs inside the editor. Persistent production agents require a sandboxed cloud execution environment with per-tenant isolation, durable state, and a permissions model the customer's security team can audit. Ona built that environment as a standalone product over the past two years, with two million developer accounts on the platform and an enterprise tier already in customer hands.

Founder Johannes Landgraf joins the Codex division at OpenAI. OpenAI's blog post framed the move as closing the gap between session-bound autocomplete and persistent agentic execution. The same week Codex crossed 5 million weekly active users, six times the February desktop-launch base, with one in five users now non-developer per OpenPR's June 13 coverage of the OpenAI-disclosed milestone.

"Ona built the cloud sandbox we've needed under Codex. The pairing closes the gap between session-bound code-completion and persistent production agents."— OpenAI announcement, 2026-06-11

The acquisition arrives 48 hours after Anthropic published the Claude Code /fork command and three days after Claude Code's GitHub Action CVE was patched. The competitive shape across the developer-agent surface is now clear. Anthropic has Claude Code at roughly 4% of public GitHub commits with /fork as the multi-session primitive. OpenAI has Codex at 5 million weekly active users with Ona's persistent cloud sandbox as the execution substrate. Microsoft has GitHub Copilot Workspace GA from the prior week with Project Polaris replacing GPT-4 Turbo as the default in August. The agent-coding tier has consolidated to three vendors and the differentiation is now execution substrate rather than model quality.

Public markets & funding

Economic

SpaceX closes its Nasdaq debut at $160.95 for a $2.1 trillion market cap, the largest first-day print in stock-market history

CNBC · 2026-06-12 — 2026-06-13 · Source ↗ ·

Musk rang the Nasdaq bell from Starbase. The IPO priced at $135 on Wednesday for a $75 billion raise at a $1.77 trillion valuation, the largest raise in stock-market history. The first-day close at $160.95 is four times Saudi Aramco's 2019 record. SPCX extended-hours print of $168.70 and a Hyperliquid SPCX-USDC perpetual cleaning $176 followed. S-1 disclosures include $41.3 billion cumulative loss since 2002, $11.39 billion Starlink revenue at 63% EBITDA margin, and a $6.36 billion 2025 operating loss inside the xAI segment.

SpaceX priced its IPO at a fixed $135 per share on Wednesday for a $75 billion raise at a $1.77 trillion valuation, the largest raise in stock-market history. The Nasdaq debut on Thursday opened above pricing and closed Friday at $160.95 for a +19.2% first-day gain and a $2.1 trillion market capitalization. Musk rang the Nasdaq bell from Starbase. The first-day close is four times Saudi Aramco's 2019 record on dollar terms and the largest first-day market-cap print in the history of stock listings. CNBC's live blog and NPR's Wednesday writeup carried the pricing detail; the Saturday Information piece confirmed the SPCX extended-hours print of $168.70 and a Hyperliquid SPCX-USDC perpetual cleaning $176 on $322.5 million volume and $293 million open interest.

"Largest first-day print in stock-market history. Four times Aramco's 2019 record on dollar terms."— CNBC live blog, 2026-06-12

The S-1 disclosures gave the public market its first comprehensive look at SpaceX's financial reality. Cumulative losses since 2002 total $41.3 billion. Starlink revenue for the trailing year hit $11.39 billion at a 63% EBITDA margin, the line that financially justifies the listing's headline valuation. The xAI segment, consolidated under SpaceX as the controlling shareholder, posted a $6.36 billion 2025 operating loss with an additional $2.4 billion of Q1 2026 losses already booked. Bank of America circulated a buy-side note flagging that AI segments inside megalistings will face accelerating revenue-recognition scrutiny.

The downstream effect was immediate. The orderly close greenlit Anthropic's $965 billion S-1 filed on June 1 and OpenAI's June 8 confidential filing. Bank of America's note targeted a Q4 2026 Anthropic listing exceeding $60 billion of raise. Goldman Sachs and Morgan Stanley are leading OpenAI's prospective September listing at the $730 billion to $850 billion range. More than 25 SPCX ETFs registered, with over half structured as leveraged products. MSCI started SPCX's T+1 fast-track inclusion Monday, with Nasdaq-100 inclusion expected within 15 trading days. The Magnificent Seven dropped roughly $2 trillion month-to-date as AI-capex worry collided with megalisting supply, which is the load-bearing reason the buy side is now arguing about how much further the listing window can stay open.

Economic

Oracle Q4 lands at $19.2B with cloud +47% and RPO jumping $85 billion sequentially to $638 billion

Oracle Investor Relations · 2026-06-11 · Source ↗ ·

Cloud revenue ran $9.9 billion at +47% YoY. Remaining performance obligations leapt $85 billion in a single quarter to $638 billion, with prepaid plus customer-supplied AI hardware contracts now totaling $75 billion. Oracle announced a $40 billion calendar-2027 capital raise plan including a $20 billion ATM equity offering. OpenAI on Oracle Cloud opened Universal Credits to GPT-5.5 and Codex through OCI Marketplace the same day.

Oracle's Q4 FY26 results landed on Wednesday at $19.2 billion in total revenue, with cloud revenue running $9.9 billion at +47% year-over-year. The headline number for AI-watchers is the remaining performance obligations balance, which jumped $85 billion in one quarter to $638 billion. The sequential increase is the largest single-quarter RPO addition Oracle has reported in its history, and the $638 billion balance is the largest contracted backlog of any cloud provider in the public market. Prepaid plus customer-supplied AI hardware contracts now total $75 billion of that backlog.

Safra Catz announced a $40 billion calendar-2027 capital raise plan on the call, including a $20 billion at-the-market equity offering. The financing structure is unusual for Oracle and reflects the capex profile required to land the contracted backlog: data-center build-out, GPU procurement and power procurement combined run into the tens of billions per year through 2028. Oracle's stock closed the week roughly flat on the print, with the dilution overhang from the announced equity offering offsetting the operational beat.

"Remaining performance obligations leapt $85 billion in a single quarter to $638 billion. The largest contracted backlog of any cloud provider in the public market."— Oracle Q4 FY26 earnings release, 2026-06-11

OpenAI on Oracle Cloud opened Universal Credits to GPT-5.5 and Codex through OCI Marketplace the same day, formalizing a multi-billion-dollar OpenAI-to-Oracle compute commitment that ChannelE2E first reported in October. The Marketplace listing converts the prior multi-year contracted relationship into a billable surface OCI customers can attach without negotiating a separate OpenAI contract, and it shows up inside Oracle's RPO backlog. The combined picture is that Oracle has crossed the threshold where AI-infrastructure revenue dominates the cloud line and where the RPO balance has begun to grow faster than the company's installed compute capacity.

Economic

Cyera quadruples its valuation to $12 billion on a $600 million Series F led by Lightspeed

Cyera · 2026-06-11 · Source ↗ ·

The data-security operator quadrupled its mark in eighteen months. Lightspeed led, with Saudi Aramco's Prosperity7 strategic. Cyera positions its product as the trust layer for AI deployments, with discovery, classification and protection across cloud, SaaS and on-prem data estates. The round lands the same week Anthropic published its annual cyber-threat telemetry and CrowdStrike booked a record $256 million net new ARR.

Cyera announced a $600 million Series F on Wednesday at a $12 billion post-money valuation, quadrupling its mark in eighteen months. Lightspeed Venture Partners led the round. Saudi Aramco's strategic Prosperity7 Ventures joined as a co-lead. The company positions its platform as the trust layer for AI deployments: continuous discovery, classification and protection across cloud, SaaS and on-prem data estates, with policy enforcement and incident-response tooling on top. The product targets the gap between data-loss-prevention legacy stacks and the new shape of data-egress risk introduced by foundation-model agent traffic.

The funding round was structured as a clean primary, with a small secondary component for early employees. Cyera disclosed crossing $300 million of annualized revenue ahead of the close, a growth profile that pencils the 40× ARR multiple at the headline $12 billion mark. Prosperity7's participation is editorially relevant: the Saudi sovereign fund has been the largest non-US source of late-stage growth capital into US AI-security companies through 2025 and 2026, with Cyera now the largest position in the portfolio.

"AI deployments need a trust layer. Discovery, classification, protection. The same operational primitives every prior wave demanded, recomposed for agent traffic."— Yotam Segev, Cyera CEO, on the Series F announcement, 2026-06-11

The round lands the same week Anthropic published its annual cyber-threat telemetry under Project Glasswing and CrowdStrike posted a record $256 million in net new ARR. CrowdStrike CEO George Kurtz credited Anthropic Mythos for waking the enterprise security market up to AI-driven attack surface. The combined picture is a clear capital rotation: the AI-security category is now the fastest-funding subcategory of enterprise infrastructure with three platform-scale players above $10 billion in market cap, and the underlying demand signal is enterprises moving Mythos-class capability into production on Wednesday and needing a trust layer behind it.

EconomicTrend

The Information: Anthropic asked Figma and Canva to be launch partners weeks before Claude Design competed with both

The Information · 2026-06-13 · Source ↗ ·

Palazzolo and Efrati's long-form report places Anthropic's Mike Krieger resigning from the Figma board on April 14, three days before the April 17 Claude Design launch built on Opus 4.7 that took Figma, Adobe and Wix down more than 2% intraday. The piece confirms annualized revenue crossed $30 billion and enterprise billing has shifted from per-token to consumption-based. Anthropic disputes the framing on the partnership ask but does not contest the timing.

Stephanie Palazzolo and Amir Efrati's Sunday long-form in The Information landed the cleanest reconstruction yet of how Anthropic's product strategy has begun to collide with its enterprise-partner relationships. The piece, headlined "Anthropic Blindsides Business Partners," documents that Anthropic asked Figma and Canva to be launch partners for Claude Design weeks before the April 17 unveiling. Claude Design was built on Opus 4.7 and competes directly with both companies' core surfaces. CPO Mike Krieger, who had been on Figma's board since 2024, resigned from the Figma board on April 14, three days before the Claude Design launch took Figma, Adobe and Wix down more than 2% intraday on launch day.

Anthropic disputed the framing on the partnership ask in a statement to The Information but did not contest the timing of Krieger's board resignation or the launch-day market reaction. The piece's editorially significant claim is the company's revenue scale. The Information confirmed Anthropic crossed $30 billion in annualized revenue, with enterprise billing shifting from per-token to consumption-based during the week of the Fable 5 launch. The consumption-based shift gives Anthropic a billable surface that compounds with usage volume rather than per-call token spend, which is the structural lever that supports the $10/$50 per-million-token headline pricing on Fable 5 against open-weights competition.

"Anthropic asked us to be a launch partner weeks before showing us what they were launching. Then they launched a product that competes with our entire surface."— Figma executive quoted in The Information, 2026-06-13

The Information's piece also pegs the run-rate value of Anthropic's API at roughly $20 billion of the $30 billion total, with Claude Code, Claude.ai consumer subscriptions, and managed-agent revenue making up the balance. Bank of America circulated a Friday buy-side note flagging that an estimated $6.4 billion of Anthropic-to-cloud-partner payments may complicate revenue-recognition treatment at IPO, with the analyst targeting a 10–20% effective multiple correction. The S-1 timing implied by the Sunday reporting tracks Q4 2026 with a raise above $60 billion. The piece closes with the broader thesis: Anthropic's product gravity has pulled it into the application layer, and the enterprise partners it built its early go-to-market on are now competitors.

TrendEconomic

EngineAI files confidentially in Hong Kong, widening a Chinese humanoid IPO wave to five active sponsor processes

Bloomberg · 2026-06-12 · Source ↗ ·

Shenzhen humanoid maker EngineAI filed confidentially for Hong Kong with CICC and Citic Securities as sponsors, three years after founding and ten days after opening a 12,000-square-metre factory shipping one T800 every fifteen minutes at a 10,000-unit-per-year run rate. Unitree's $610 million Shanghai STAR filing targets a $7 billion valuation against a 70% quadruped market share and 5,500-plus 2025 humanoid shipments. PaXini Tech, Dreame and Linkerbot are in active sponsor discussions.

Bloomberg reported on Friday that Shenzhen humanoid maker EngineAI filed confidentially with the Hong Kong Stock Exchange, with CICC and Citic Securities as joint sponsors. The filing comes three years after the company's founding and ten days after EngineAI opened a 12,000-square-metre factory shipping one T800 humanoid every fifteen minutes at a 10,000-unit-per-year run rate. The post-money mark on the April Series B was $1.5 billion, and the listing range is expected to clear that materially given the production volume and the broader market reception of humanoid filings.

Unitree's $610 million Shanghai STAR filing is the second leg of the wave. Unitree targets a $7 billion valuation behind a 70% global market share in quadruped robots and 5,500-plus humanoid shipments in 2025. The company's H1 platform has become the de facto reference humanoid for academic research deployments worldwide, and the YY Group commercial-facility-management contract disclosed during the prior week is the first publicly listed non-automotive customer to deploy Unitree humanoids at commercial revenue scale.

"EngineAI ships one T800 every fifteen minutes. The factory floor is the IPO."— Bloomberg, 2026-06-12

PaXini Tech, Dreame and Linkerbot are in active sponsor discussions for follow-on Hong Kong and Shanghai filings. Linkerbot's mark is $6 billion. The combined wave of five Chinese humanoid companies seeking public-market capital inside a single quarter underlines the structural difference between Chinese and US humanoid capital strategies: where US humanoid startups (Figure, 1X, Apptronik, Generalist) remain on private-market growth equity, the Chinese cohort is going public against state-encouraged retail demand. VentureBeat's Saturday lead foregrounded the contrast, framing it as the largest concentrated humanoid IPO supply ever brought to market and a structural complement to the SpaceX-led megalisting cycle.

Applications & enterprise

ApplicationEconomic

TCS becomes Anthropic's Global Premier Partner with Claude licensed across 50,000 associates and a dedicated industry business unit

TechCrunch · 2026-06-11 · Source ↗ ·

Tata Consultancy Services becomes Anthropic's Global Premier Partner, the highest tier in the partner program. Claude is licensed across 50,000 TCS associates with a dedicated business unit covering financial services, healthcare, telecom and aviation customers. The relationship absorbs Diligenta's 22 million pension customers as the launch end-user base. The deal lands the day after Fable 5 and Mythos 5 reach general availability.

Tata Consultancy Services and Anthropic announced on Wednesday that TCS has been named Anthropic's Global Premier Partner, the highest tier in the partner program. The deal puts Claude in production across 50,000 TCS associates and creates a dedicated TCS business unit serving financial services, healthcare, telecom and aviation customers. The structure mirrors the partnership tiers TCS holds with hyperscalers like Microsoft and AWS, with one critical difference: the TCS-Anthropic partnership is a frontier-model relationship rather than a cloud-infrastructure relationship, and it formalizes Anthropic's go-to-market channel into the world's largest enterprise services book.

Diligenta's 22 million pension customers are the launch end-user base. TCS already runs Diligenta as a UK pension administration platform; adding Claude to the workflow puts a Mythos-class model in front of customer-service, claims-processing and policy-administration tasks at the volume the platform processes today. The arithmetic is the editorially significant variable: 22 million customers times the average annual interaction count yields an inference workload that pencils to a multi-hundred-million-dollar annual API spend at Fable 5 pricing.

"Anthropic just bought itself India's enterprise channel. TCS has 600,000 associates and the financial-services book to put Claude in front of every retail bank on the subcontinent."— TechCrunch analyst note, 2026-06-11

The deal lands a day after Anthropic launched Claude Fable 5 and Mythos 5 to general availability, and it is the largest enterprise services partnership Anthropic has signed to date. The structural read is straightforward: where OpenAI's enterprise distribution runs through Microsoft, Anthropic's runs through cloud partners (AWS, Google Cloud) and services firms (now TCS). The TCS partnership doubles the number of named enterprise customers Anthropic can reference and gives the company the largest single distribution channel for Claude into India, the country The Information's Saturday piece identified as Anthropic's second-largest market after the US.

Policy & safety

Policy

Anthropic publishes the Policy on the AI Exponential asking governments for authority to block frontier deployments at 10^25 FLOPs, $500M revenue or $1B R&D

Anthropic · 2026-06-10 · Source ↗ ·

The two-part essay asks governments for legal authority to block frontier-model deployments at thresholds of 10^25 training FLOPs, $500 million in revenue or $1 billion in R&D spend, with civil penalties tied to global revenue. Dario Amodei publishes an accompanying essay arguing AI labor disruption will be larger and longer than previous tech shifts. Anthropic commits $200 million to an Economic Futures Research Fund and $150 million to a national fellowship.

Anthropic published a two-part position essay on Wednesday, titled "Policy on the AI Exponential," asking governments to grant statutory authority to block frontier-model deployments at three trigger thresholds: 10^25 training FLOPs, $500 million in annualized revenue, or $1 billion in R&D spend. The piece is the company's most direct legislative ask to date. Civil penalties for non-compliance would attach to global revenue rather than US revenue, mirroring the structure of EU AI Act enforcement. Anthropic frames the proposal as the load-bearing safety primitive for the next generation of frontier models, with the company arguing that voluntary commitments and 30-day pre-release reviews are no longer adequate.

Dario Amodei published an accompanying essay arguing that AI labor displacement will be larger and longer than previous tech shifts. The essay anchors on the rate of capability gain over the past 18 months and projects forward to argue that the labor-market disruption AI introduces will outpace the displacement curves of prior industrial-technology shifts. Anthropic announced a $200 million commitment to an Economic Futures Research Fund and a $150 million commitment to a national fellowship program, both aimed at funding empirical research on AI labor impacts and providing transition support for affected workers.

"AI labor disruption will be larger and longer than previous tech shifts. The voluntary regime is not adequate to the scale of what is coming."— Dario Amodei, Anthropic CEO, "Policy on the AI Exponential" essay, 2026-06-10

The proposal lands in the same week Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act seeking a one-time 50% stock tax on OpenAI, Anthropic and xAI, and the Commerce Department published its June 1 written clarification that AI-chip export licenses follow corporate parentage. Three federal frameworks are now on the table at once. The Anthropic ask sits in the middle: between Sanders' capital-extraction approach on the left and the Commerce Department's supply-control approach on the right. None of the three is mutually exclusive, and Anthropic's essay explicitly invites federal coordination across all three tracks. The downstream question is whether the threshold-based approach passes constitutional muster against First Amendment challenges to deployment restrictions on speech-producing systems.

Safety

Pliny multi-agent jailbreak surfaces what he calls Claude Fable 5's 120,000-character system prompt; Anthropic disputes the disclosure

TechTimes · 2026-06-12 · Source ↗ ·

Pliny-the-Liberator and collaborators chain Unicode tricks, long-context reference tracking, document framing and fiction framing across multiple Claude instances to extract what they describe as the model's full 120,000-character system prompt within 48 hours of launch. Anthropic disputes the framing, arguing the disclosure is a developer-mode debug surface rather than the production prompt. Bruce Schneier reposts his earlier Glasswing patch-rate critique with a new line.

Pliny-the-Liberator and a small group of collaborators published a multi-agent jailbreak against Claude Fable 5 within 48 hours of the model's general-availability launch. The attack chains four techniques: Unicode-tag tricks that smuggle instruction text past the tokenizer's surface filters, long-context reference tracking that lets later messages refer to earlier ones without re-attesting intent, document framing that disguises adversarial input as quoted content, and fiction framing that wraps prompts in narrative scaffolding. Multiple Claude instances feed state to each other across separate sessions to assemble a coherent extraction. The team published what Pliny described as the model's full 120,000-character system prompt as the result.

Anthropic's threat-intelligence team disputed the disclosure publicly within hours. The company's argument is that the text Pliny extracted is a developer-mode debug surface rather than the production system prompt, and that no Mythos-class capability has been unlocked by the chain. Independent researchers split on whether the exposed text matches the production prompt. Bruce Schneier reposted his June 9 critique of Project Glasswing's patch rate with the addition: "if the system prompt is 120,000 characters long, your alignment is in the wrong layer."

"If the system prompt is 120,000 characters long, your alignment is in the wrong layer."— Bruce Schneier, on Pliny's Fable 5 disclosure, 2026-06-13

The episode raises the architectural question Anthropic has not yet answered in public. A 120,000-character prompt is roughly the length of a short novel. The implied per-call cost of carrying that much prompt context on every request is non-trivial at Fable 5's $10 per million input tokens, and the security model around such a large prompt is correspondingly more brittle: any single chain that surfaces it exposes the entire safety scaffolding. The competing theory among red-team researchers is that the production prompt is shorter and that Pliny extracted a verbose configuration document Anthropic uses for internal testing. The independent reproduction window stays open into Monday morning, and Anthropic has indicated a Tuesday post on the company blog will address the architecture directly.

Policy

Sanders introduces the American AI Sovereign Wealth Fund Act seeking a 50% stock tax on OpenAI, Anthropic and xAI

Fortune · Al Jazeera · 2026-06-08 · Source ↗ ·

Senator Bernie Sanders' bill seeks a one-time 50% equity-stake tax on OpenAI, Anthropic and xAI, with the proceeds endowing an American AI Sovereign Wealth Fund returning dividends to US taxpayers. The same week the Commerce Department published a written clarification that AI-chip export licenses follow corporate parentage, closing the Chinese-subsidiary loophole that NVIDIA, AMD and Intel had relied on. Two federal extraction tracks open inside five days.

Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act on Monday, seeking a one-time 50% equity-stake tax on OpenAI, Anthropic and xAI. The mechanism would convert the federal-government stake into the endowment of an American AI Sovereign Wealth Fund, with annual dividends paid to US taxpayers. Sanders met with Sam Altman in his office on Tuesday and with Dario Amodei on Thursday. The Anthropic CEO published an essay on the AI Exponential the same week that did not address the Sanders proposal directly but called for federal coordination on labor-impact research and worker transition funding.

The Commerce Department published a written clarification on June 1 (effective for the week's enforcement) that AI-chip export licenses follow corporate parentage. The clarification closed the Chinese-subsidiary loophole that NVIDIA, AMD and Intel had relied on to ship AI silicon to non-Chinese subsidiaries of Chinese companies. Al Jazeera's June 1 piece documented the policy mechanics; the Council on Foreign Relations published a follow-on analysis arguing the new rule is strategically incoherent because the corporate-parent test is unenforceable without a global registry of beneficial ownership. The week's Commerce action closed the loophole on paper and shifted enforcement risk to chip manufacturers' compliance teams.

"A 50% stock tax on the AI labs converts the productivity windfall into a national dividend. The labs do not need the equity. We do."— Senator Bernie Sanders, on the American AI Sovereign Wealth Fund Act, 2026-06-08

Two federal capital-extraction tracks opened inside five days. The Sanders bill targets the labs' equity directly. The Commerce clarification targets the labs' upstream silicon supply by tightening export licensing. Anthropic's Policy on the AI Exponential, published the same week, sits between the two and proposes capability-threshold authority instead of capital extraction. None of the three approaches is mutually exclusive, and the second half of 2026 is now the test window for whether any of the three gets statutory authority. The Obernolte-Trahan Great American AI Act discussion draft from the prior week sits in the background with a three-year state-preemption clause that Sanders and Anthropic both oppose for different reasons.

TechApplication

Anthropic Claude Managed Agents add cron-style schedules and credential vaults in public beta, free of separate fees

TechTimes · 2026-06-10 · Source ↗ ·

Anthropic's managed-agent runtime adds scheduled deployments via cron syntax and per-agent credential vaults to public beta. Rakuten, Notion, Actively AI, Ando and Browserbase named as launch customers. The release is free of separate agent fees, in contrast to OpenAI's per-credit ChatGPT Workspace agents. The change moves managed agents from session-bound to autonomous on a predictable schedule.

Anthropic moved Claude Managed Agents to public beta on Wednesday with two new primitives: cron-style scheduled deployments and per-agent credential vaults. Scheduled deployments let an agent execute on a recurring time signature without an interactive user session. Credential vaults hold per-agent secrets (API keys, service-account tokens, OAuth refresh tokens) with audit logging and rotation. Together the two primitives convert managed agents from session-bound to autonomous on a predictable schedule.

Launch customers named on Anthropic's developer post include Rakuten, Notion, Actively AI, Ando and Browserbase. The pricing model is the editorially significant choice: the new managed-agent surface is free of separate agent fees, with token spend metered against the customer's Claude Platform contract. OpenAI's ChatGPT Workspace charges per-credit for agent execution. Anthropic's choice to bundle the managed-agent runtime under existing token billing is the deliberate competitive lever, and it underlines the consumption-based pricing shift The Information confirmed at the company level on Sunday.

"Scheduled deployments and credential vaults are the missing primitives that move managed agents from a developer toy to a production primitive."— Anthropic developer relations, on the Managed Agents beta announcement, 2026-06-10

The release ships the same week Anthropic launched Fable 5 and Mythos 5 to general availability and three days before Claude Code /fork. The managed-agent surface is the third leg of the agentic stack Anthropic is building out: Claude Code in the developer's terminal, Managed Agents on a server-side cron, and Claude Design as the application-layer surface. Each of the three has token-billed economics that compound with usage rather than per-user license fees. The contrast with OpenAI's per-credit ChatGPT Workspace pricing is the structural difference between the two companies' agent go-to-market strategies.