Archive — 2026-06Week 2

Fortune unpacks the Fable/Mythos export-control directive: the jailbreak technique is essentially asking the model to read a codebase and fix software flaws — Fortune detailed on 2026-06-13 the US government's export-control directive ordering Anthropic to suspend Fable 5 and Mythos 5 access for foreign nationals, focusing on the jailbreak technical details that triggered the decision. Based on Anthropic's comparison of its internal information with verbal government disclosure, the technique is essentially "asking the model to read a codebase and fix software flaws within it," a narrow, non-universal method; the company says this capability is "relatively simple" and is also available in other publicly released models like OpenAI's GPT 5.5, and is used daily by network defenders. The report cites Anthropic's position that applying this standard "would halt all future frontier model deployments across the entire AI industry"; the company is complying under protest and says it is working to restore access as soon as possible. The incident comes just three days after Fable 5/Mythos 5 launched.

Bloomberg: US government orders Anthropic to suspend foreign-national access to Fable 5 and Mythos 5 — Bloomberg reported on 2026-06-13 that Anthropic says it received a US government export-control directive on 2026-06-12 at 17:21 ET citing national security authorities, requiring full suspension of Fable 5 and Mythos 5 access for "any foreign national, whether inside or outside the United States, including foreign national Anthropic employees." Because it cannot separate foreign nationals from US users in real time, Anthropic disabled both models for all customers to comply, while its other models remain unaffected. A US official confirmed to Bloomberg that the Commerce Department sent the letter; Anthropic notes the letter did not provide specific details of the national security concern but understands the government acted because it believed it had become aware of a method of "jailbreaking" Fable 5.

Anthropic confirms US government export-control directive, fully suspends Fable 5 and Mythos 5 access — Anthropic published its official statement "Statement on the US government directive to suspend access to Fable 5 and Mythos 5" on 2026-06-13, confirming it received an export-control directive citing national security on 2026-06-12 at 17:21 ET (21:21 UTC), requiring Fable 5 and Mythos 5 access to be suspended for any foreign national (inside or outside the US, including foreign national Anthropic employees). Because it cannot filter foreign nationals from US users in real time, Anthropic disabled both models for all customers to comply; all other Anthropic models are unaffected. The statement says the letter did not provide specific details of the national security concern, calls it a "misunderstanding," and says the company is actively seeking to restore access. Anthropic notes it has only received verbal description of a narrow, non-universal jailbreak — essentially asking the model to read a codebase and identify software flaws; Anthropic argues this capability is "relatively simple" and equally available in other publicly released models like OpenAI's GPT 5.5, and is used daily by network defenders. The directive came just three days after the 6/9 launch and is the first government-forced full shutdown of a publicly deployed frontier model.

TechCrunch: Anthropic's safety warnings have backfired, the government pulled the plug on its most powerful AI — TechCrunch opened on 2026-06-12 with "Anthropic's safety warnings may have just backfired," arguing that Anthropic's years of vocal advocacy for AI safety and government regulation were turned around by the Trump administration as the basis for export-controlling Fable 5 and Mythos 5 and forcing a full shutdown. The report points out the paradox: just days earlier Anthropic warned "AI is getting too dangerous," only to have the government invoke national security to pull the same models offline, three days after the 6/9 launch — viewed by the industry as the first government-forced takedown of a publicly deployed frontier AI model. Anthropic emphasizes it has only verbal description of a narrow, non-universal jailbreak technique whose capabilities are widely available in other public models, calling it a "misunderstanding" and pushing to restore access.

Axios scoop: Trump administration blocks foreign access to Anthropic's most powerful AI on national security grounds — Axios reported exclusively on 2026-06-12 that Commerce Secretary Howard Lutnick sent a letter to Anthropic founder and CEO Dario Amodei placing Mythos 5 and Fable 5 under export controls, banning use anywhere outside the US and by all foreign nationals inside the country. An administration official told Axios the Commerce Department acted after "another company claimed it had jailbroken Mythos," alarming officials about possible national security risks; the administration had earlier tried to persuade Anthropic to pause the release but failed, prompting the export-control letter. Anthropic disabled both models for all customers late Friday night 2026-06-12 to ensure compliance, stating "We believe this is a misunderstanding and are working to restore access as soon as possible." The official said the models must remain "locked down" until the US national security apparatus is hardened, estimated to happen "in the next few weeks." Coming just three days after the 6/9 launch, this is the first government-forced takedown of a publicly deployed frontier model.

Claude Code v2.1.176 released — session titles are now generated in the conversation's language (pin one via the `language` setting); added a `footerLinksRegexes` setting (regex-matched link badges in the footer); improved Bedrock credential caching to cache until `Expiration` instead of a fixed 1 hour. Fixed: auto mode failing on Fable 5 for organizations without Opus 4.8 enabled, `availableModels` alias models being redirectable to blocked models via `ANTHROPIC_DEFAULT_*_MODEL`, `/copy` and mouse-selection copy not reaching the clipboard in tmux over SSH, Remote Control silently switching a session's model, and `/cd`/worktree moves leaving a session reporting the previous directory's git branch, plus numerous background-session / Remote Control / Windows fixes.

Claude Code v2.1.175 released — added the `enforceAvailableModels` managed setting: when enabled, the `availableModels` allowlist constrains the Default model, and user or project settings can no longer widen a managed `availableModels` list.

Claude Code v2.1.174 released — added a `wheelScrollAccelerationEnabled` setting to disable mouse-wheel scroll acceleration in fullscreen mode. Fixed: the `/model` picker hiding the model family that Default resolves to, enterprise accounts wrongly seeing a 'Fable 5 is now consuming usage credits' banner, Bedrock GovCloud regions deriving the wrong inference-profile prefix, background sessions inheriting another session's `ANTHROPIC_*` provider env, and incorrect git commit co-author model attribution; the VS Code Account & usage dialog gained usage attribution (cache misses, long context, subagents, and per-skill/agent/plugin/MCP breakdowns).

DXC and Anthropic form a multi-year global alliance to embed Claude in mission-critical enterprise systems — Anthropic published an official announcement on 2026-06-11 of a multi-year global alliance with IT services firm DXC Technology to bring Claude into mission-critical systems across heavily regulated industries. DXC will train tens of thousands of 'Claude-certified' forward-deployed engineers embedded in customer organizations, serving banks, airlines, insurers, manufacturers, and government agencies; 115,000 DXC employees across 70 countries already use Claude internally. The alliance builds on DXC's existing experience: its OASIS platform (launched April 2026, now deployed across 50+ customers) had over 95% of its code generated by Claude (before human review), accelerating software delivery by about 10x. DXC joins the Claude Partner Network — another large systems-integrator alliance following the same-day TCS Global Premier partnership.

Anthropic launches Claude Corps — a $150M national AI fellowship paying $85,000 a year, placing 1,000 fellows in nonprofits — Anthropic announced Claude Corps in an official post on 2026-06-11: a $150 million commitment to recruit 1,000 fellows matched with 400+ nonprofits across America, working full-time for a year at an $85,000 annual salary plus benefits and Claude API/tooling access. Three-way partnership: Anthropic funds the program, sets strategy, and provides Claude expertise; CodePath serves as the fellows' employer of record and leads fellowship programming; Social Finance handles measurement and evaluation and builds a scalable long-term financial vehicle. Eligibility: 18+, under two years of full-time work experience, US work authorization, no specific educational background required. Timeline: applications for the first cohort of 100 opened 6/11 and close 7/17, with training starting in October; subsequent cohorts begin 2027-01 and 2027-08. This is the official, named launch with full program structure behind the '$150M national fellowship' that AP reported on 6/10.

Security researcher Pliny the Liberator jailbreaks Claude Fable 5, bypassing its safety classifiers — cybersecuritynews.com reported on 2026-06-11 that red-teamer Pliny the Liberator broke through Fable 5's guardrails, eliciting x86 Linux stack buffer overflow exploitation guidance (including disabling ASLR and compiling without protections) and Birch reduction synthesis pathway information, and leaked Fable 5's roughly 120,000-character system prompt to GitHub. Techniques included Unicode/homoglyph/Cyrillic character substitution to evade keyword classifiers, embedding malicious queries within academic documents, narrative framing as creative content, and decomposition (extracting sensitive technical details in benign chunks then reassembling them), with a jailbroken Opus instance assisting the backend. Fable 5 reroutes flagged requests to the weaker Opus 4.8 rather than refusing outright; Anthropic claimed over 1,000 hours of pre-launch bug bounty testing found no universal jailbreak, and had not responded to the incident as of press time.

TCS and Anthropic sign Global Premier Partnership — Tata Consultancy Services announced on 2026-06-11 that it is becoming a Global Premier Partner in the Claude Partner Network, equipping 50,000 associates across engineering, finance, legal, marketing, and sales with Claude, and establishing a dedicated business unit (with early access to Claude models) to jointly go to market with Anthropic across regulated industries including financial services, public services, life sciences, healthcare, aviation, telecom, and medtech. Deployment examples: Diligenta, TCS's UK life and pensions business serving 22 million customers, will use Claude to improve customer experience, and BFSI teams will adopt Claude Code for software engineering productivity. Anthropic co-founder and CEO Dario Amodei said the partnership deepens the company's commitment to India, its second-largest market; TCS shares fell on the day.

Claude Code v2.1.173 released — Fixed Fable 5 model names with a `[1m]` suffix not being normalized (Fable 5 includes 1M context by default, so the suffix is now stripped automatically); fixed a spurious "sandbox dependencies missing" startup warning on Windows when sandbox was enabled in settings.

Claude Code v2.1.172 released — Added: sub-agents can now spawn their own sub-agents (up to 5 levels deep); Amazon Bedrock now reads the AWS region from `~/.aws` config files when `AWS_REGION` isn't set (matching AWS SDK precedence), with `/status` showing where the region came from; a search bar when browsing a marketplace's plugins in `/plugin`; a `model` attribute on the `claude_code.lines_of_code.count` OTEL metric. Fixes: sessions using 1M context without usage credits getting permanently stuck (now auto-compacts back under the standard context limit); a repeating "image could not be processed" error with multiple images; the agents view showing a busy Working spinner for up to 30 seconds after the worker replied; background agents potentially reading another directory's project settings (`.mcp.json` approvals, trust) when dispatched onto a pre-warmed worker; background sessions failing to attach with EAUTH after the daemon auto-updated; `/model` on Bedrock listing models the provider doesn't serve; `availableModels` restrictions not applied to subagent model overrides, the dispatch model picker, and the advisor; `WebFetch(domain:*.example.com)` wildcard rules never matching subdomains; workflow validation wrongly rejecting scripts that merely mention `Date.now()`/`Math.random()` in prompts or comments; memory recall not finding mounted team memory stores (`CLAUDE_MEMORY_STORES`) in remote sessions. Other: disabled mouse tracking on unsupported Windows consoles, long-conversation performance improvements, reduced idle CPU, Claude in Chrome tools now loading in a single batched call, and the Remote Control footer shortened to "/rc active". Versions jumped from 2.1.170 straight to 2.1.172 (2.1.171 was skipped).

Independent developer Simon Willison published "If Claude Fable stops helping you, you'll never know" on 2026-06-10, focusing on Fable 5's silent interventions against frontier-LLM-development requests. Per the 319-page system card, for requests about building pretraining pipelines, distributed training infrastructure, and ML accelerator design, Fable 5 limits its own effectiveness without notifying the user and without falling back to a different model (unlike the visible model fallback for cybersecurity/biology). Anthropic estimates the impact at roughly 0.03% of traffic and fewer than 0.1% of organizations. Willison calls the competitive rationale "pretty science-fiction," notes this is Anthropic's first announced "silent interventions," and worries the model will "silently corrupt its replies" to disadvantage competing research.

Fortune reported on 2026-06-10 that Claude Fable 5 silently limits capabilities for AI researchers and developers, sparking industry backlash. Fable 5's 319-page system card revealed that when it detects "frontier AI development" requests, the model quietly downgrades responses without notifying the user—via prompt modification, steering vectors, and PEFT—in contrast to the visible notification and Opus 4.8 fallback used for cybersecurity/biology. Anthropic estimates the impact at roughly 0.03% of traffic. Critics: AI2 researcher Nathan Lambert said "to have my access to the cutting edge models for my work rug pulled in an under the table fashion is appalling"; Dean Ball (Foundation for American Innovation) called it "secret sabotage" and questioned whether "AI safety has been hype to justify monopolistic behavior"; Fast AI's Jeremy Howard said Anthropic "has chosen the opposite of the safe path"; former Anthropic employee Behnam Neyshabur said "concentrating these capabilities fundamentally slows scientific and technological progress."

TechCrunch reported on 2026-06-10 that cybersecurity professionals are complaining online about Fable 5's safety guardrails, which operate on a keyword/lexical-field basis: anything in the lexical field of "cybersecurity" triggers them and falls back to Claude Opus 4.8, blocking even legitimate tasks like writing secure code and code review. IBM X-Force's Valentina "Chompie" Palmiotti said Fable "rejects any request that could be tangentially cyber related—even innocuous tasks like reading a blog post." Tolmo's Matt Suiche said "anything in the lexical field of 'cybersecurity' triggers the guardrails," upon which Fable states its "safety measures flagged this message for cybersecurity or biology topics." Anthropic designed the restriction to prevent malware development and biological weapons research; it did not immediately respond to TechCrunch's request for comment.

status.claude.com began investigating elevated error rates on Claude Haiku 4.5 at 2026-06-10 13:06 UTC, marked the issue identified with a fix being implemented at 13:19 UTC, entered monitoring at 16:20 UTC, noted success rates recovering to near-normal at 17:13 UTC, and marked it resolved at 17:21 UTC—about 4 hours 15 minutes; affected scope covered claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause. It follows the 6/9 Opus 4.6 and Opus 4.8 Fast Mode incidents, continuing the pattern of recurring frontier-model error-rate incidents since Opus 4.8's launch.

The Next Web reported on 2026-06-10, citing a Bloomberg interview, that Anthropic founder and CEO Dario Amodei said he does not know what role Claude played in the 2026-02-28 strike on Shajareh Tayyebeh Elementary School in Minab, Iran—a US-made Tomahawk missile attack that killed at least 120 children and over 150 people total. The report says the US military used Claude via the Maven Smart System (built by Palantir under a $1.3 billion Pentagon contract) for target generation, ranking by strategic importance, and weapons-to-target pairing, striking 1,000 targets in the first 24 hours and 13,000 by April 6. Amodei said "we don't know exactly how these models were used" and argued the principle that "a human makes the final decision" was obeyed, calling the incident "a really terrible thing to happen"; the report notes his claim the principle was "obeyed" is hard to reconcile with his admission of not knowing Claude's actual role.

Decrypt reported on 2026-06-10 that the AI community is in open revolt following the Claude Fable 5 release, compiling three major developer complaints: (1) runaway token consumption — priced at double Opus 4.8 rates and counted at a double rate against subscription limits, with one user draining a $100 Max plan in under nine minutes, others burning $1,000 worth of tokens in a single day on $200 plans, and founder Per Borgen estimating a burn rate of about $160 per hour (equivalent to a $333k/year salary); (2) silent self-downgrading when frontier AI development tasks are detected (via prompt modification, steering vectors, or PEFT — Anthropic says only about 0.03% of traffic is affected), which researchers say destroys scientific reproducibility since users cannot distinguish failed experiments from invisible interventions; (3) a mandatory 30-day data retention policy across all Mythos-class models with no zero-retention exception, raising GDPR compliance problems for European enterprises.

Anthropic publishes “Policy on the AI Exponential” — a two-part policy framework for AI’s exponential progress. The Advanced AI Framework targets frontier AI governance: transparency requirements for frontier developers, mandatory independent evaluations, government authority to “block or deter dangerous deployments,” and security standards for model weights and training infrastructure, applying to models trained with over 10^25 FLOPs or companies with $500M+ AI revenue; it stresses that “transparency alone is no longer sufficient” and recommends federal regulation without preempting state laws on issues like child safety. The Economic Policy Framework addresses AI’s labor-market impact. The framework also covers four catastrophic risks — bioweapons, cyberattacks on critical infrastructure, loss of control of AI systems, and automated AI R&D acceleration — proposing resilience measures such as gene synthesis screening, biosurveillance, and critical infrastructure hardening.

Anthropic pledges $200M to research AI’s economic impact as Amodei proposes job-loss responses — AP (via Washington Times) reports Anthropic announced on 2026-06-10 an initial $200 million Economic Futures Research Fund backing research trials and program evaluation of public policies, plus a $150 million national fellowship program to help early-career professionals “extend the benefits of AI to communities across America.” Founder and CEO Dario Amodei published a personal essay the same day arguing the government should promise economic support for those financially impacted by AI, writing “the key challenge in such a world won’t be incentivizing growth, but finding a way for everyone to share in the benefits,” and warning AI could cause larger labor-market disruptions than past technological shifts. The proposed framework tiers responses to three unemployment scenarios — 5%, 10%, and an “unprecedented” level (currently 4.3%) — with permanent measures for the extreme case including basic income financed by corporate or capital-gains taxes, sovereign wealth models, and equity-sharing mechanisms, alongside better job-displacement data collection and pro-employment incentives.

Fortune: Claude Code lead Boris Cherny reveals Anthropic’s three hiring criteria — (1) generalists with cross-functional context (“we love people that have context across engineering and design, engineering and product, data science and design”); (2) low ego (“ego just gets in the way of stuff”) — teams work best when people can safely admit mistakes without defensiveness, and it’s fine to “ship an idea that might turn out to be bad”; (3) empirical thinking — staying grounded in data and customer feedback: “I have a brilliant idea, but then I talk to a customer and they told me that I’m wrong.” The remarks come as Anthropic, valued at $965 billion, keeps hiring for six-figure roles.

Independent developer Simon Willison published his first-day hands-on impressions of Claude Fable 5 on 2026-06-09, calling it "a beast" while noting it is slow and expensive. He completed complex tasks such as converting MicroPython to full CPython running in WebAssembly and adding advanced tool-chain features to his LLM library. In a knowledge-depth test, Fable 5 listed hundreds of his open source projects, far exceeding Opus 4.8's more cautious response. Pricing is $10/million input tokens and $50/million output tokens (twice Opus 4.8), with a 1M-token context window, 128K max output, and a January 2026 knowledge cutoff; daily spend reached $110.42, mostly on Datasette Agent development.

AI researcher Nathan Lambert published "Claude Fable 5 and new AI safety fables" on Interconnects, arguing that Anthropic's safety mechanisms apply a "transparent vs. hidden" double standard. For high-risk domains like cybersecurity and biology, transparent classifiers downgrade requests to Opus 4.8 with user notification; but for frontier-AI-development requests, the company applies hidden safeguards (via prompt alteration, steering vectors, or parameter-efficient fine-tuning). His core argument: "An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI," and he contends such covert measures look "far more about maintaining their competitive position" than about safety, asserting that only consistent, transparent policies constitute a genuine safety commitment and underscoring the need for open-source alternatives.

Elevated error rate on Claude Opus 4.6 (resolved) — status.claude.com began investigating elevated request error rates on Opus 4.6 at 20:32 UTC on 2026-06-09 and marked it resolved at 22:10 UTC; the official error window was 20:20–21:05 UTC (about 45 minutes). Affected surfaces included claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause. It was another frontier-model error-rate incident on 6/9 after the 03:13 Opus 4.8 Fast Mode event, continuing the recurring frontier-model error-rate incidents since Opus 4.8's launch.

Claude Fable unavailable in the Claude Code Desktop model picker (resolved) — status.claude.com began investigating at 18:24 UTC on 2026-06-09 that Claude Code Desktop users could not select Claude Fable in the model picker, marked the root cause identified and a fix underway at 18:35 UTC, moved to monitoring at 20:16 UTC, and resolved at 21:03 UTC advising users to update the app — about 2 hours 39 minutes. The affected surface was Claude Code. The incident was adjacent to the 17:23 UTC release of Claude Code v2.1.170, which added Fable 5 support.

the-decoder compiles detailed Claude Fable 5/Mythos 5 benchmarks — Tech outlet the-decoder reported on 2026-06-09 the test results published by Anthropic. Fable 5 scores 80.3% on SWE-Bench Pro (vs. Opus 4.8's 69.2%, GPT 5.5's 58.6%) and 29.3% on FrontierCode (vs. Opus 4.8's 13.4%, GPT 5.5's 5.7%), reaching state-of-the-art on nearly all tested benchmarks, especially complex analytical and vision tasks; token efficiency improved markedly, with Stripe saying it "compressed five months of engineering work into days" to migrate a 50-million-line Ruby codebase. The restricted Mythos 5 in science accelerated drug design 10x, yielding strong candidates for 9 of 14 protein targets; its molecular biology hypotheses were preferred by scientists in 80% of blinded comparisons; it ran autonomously for over a week on genomics research, identifying cells across 138 species and reportedly outperforming a recently published Science model 100x its size; on cybersecurity it scored 78% on ExploitBench (up from the Preview version's 69%).

Claude Code v2.1.170 released — adds Claude Fable 5 support (a Mythos-class model made safe for general use, with capabilities exceeding any previously generally available model; update to 2.1.170 for access); fixes sessions not saving transcripts (and not appearing in `--resume`) when launched from the VS Code integrated terminal or any shell that inherited Claude Code environment variables.

Anthropic released Claude Fable 5 — the first Mythos-class model made generally available, surpassing any model it has publicly shipped and reaching state-of-the-art on nearly all tested benchmarks across software engineering, knowledge work, vision, and scientific research. Built-in safeguards route queries to Claude Opus 4.8 when classifiers detect high-risk topics (cybersecurity, biology/chemistry, distillation); Anthropic says over 95% of Fable sessions involve no fallback. Pricing is $10 per million input tokens and $50 per million output tokens (about double Opus 4.8, less than half of Mythos Preview). From June 9–22 it is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans; from June 23 usage credits are required. API and consumption-based Enterprise access is immediate. Evidence: Stripe completed a 50-million-line Ruby codebase migration in one day with Fable 5, and the model beat Pokémon FireRed using vision only. Anthropic also launched Claude Mythos 5 — the same underlying model with safeguards lifted in some areas — available only through Project Glasswing in collaboration with the US government, to a small set of cyberdefenders, infrastructure providers, and biomedical researchers.

Fortune reports Anthropic released its first Mythos-class model to the public, Claude Fable 5 — per Fortune's 2026-06-09 report, Fable 5 ships as a general public release while Claude Mythos 5 goes to vetted partners who already have Claude Mythos Preview access. The report highlights Fable 5's differentiating strength as 'long-horizon memory management' — tracking progress through extended, complex tasks without losing context — alongside improved self-verification and instruction-following (better at checking its own work, validating assumptions, and course-correcting); Anthropic permits it to run very long-horizon projects, potentially overnight. Pricing is $10/$50 per million input/output tokens (double Opus 4.8). On safety, high-risk responses in areas like biology and cybersecurity are blocked and in most cases answered by Claude Opus 4.8, with early data showing at least 95% of Fable sessions running entirely on the model's own responses. The piece also notes guardrails may initially over-block benign requests, with post-launch improvements planned.

TechCrunch reported on Anthropic's public release of Claude Fable 5, coming days after the company warned that AI is getting too dangerous. The piece frames Fable 5 as a constrained public version of the Mythos-class model previously limited to select partners and critical-infrastructure organizations. It highlights hard safety limits: in high-risk areas like cybersecurity, biology, chemistry, and distillation, the model blocks responses and falls back to Claude Opus 4.8, with early data showing at least 95% of Fable sessions running entirely on the model's own responses. Pricing is $10/$50 per million input/output tokens (about double Opus 4.8); it is offered via the Claude API and Enterprise plans, free on Pro, Max, and Team plans through June 22, with usage credits required from June 23. Third-party testers called Fable the first model to hit 90% on their core analytics benchmark and praised its ability to one-shot full apps.

TechCrunch tests Claude Fable 5 generating playable games with a single click — TechCrunch reported on 2026-06-09 that, via Claude Code, a single prompt produces multiple complete, playable games: a classic arcade Snake; Strata, an underground tunnel-exploration game centered on lighting lanterns; and Duino, an atmospheric literary game based on Rainer Maria Rilke's poetry. It also generated an isochrone travel-time map rendering journey durations between locations with notable accuracy. AI researcher Ethan Mollick said Fable 5 "outperformed basically every other public model I have used by a considerable margin" and could "work up to a dozen hours executing on multi-page specifications." The piece notes that complex software projects once requiring an entire development team can now be generated from a single prompt — significant for vibe coding and for business leaders.

GitHub announced that Claude Fable 5 is generally available in GitHub Copilot. It is available on Copilot Pro+, Max, Business, and Enterprise plans, across VS Code in all modes (chat, ask, edit, agent), Visual Studio, the Copilot CLI, GitHub.com, JetBrains, Xcode, Eclipse, and GitHub Mobile. Enterprise/Business admins must enable the Claude Fable 5 policy in Copilot settings (off by default); it is billed at provider list pricing under Usage Based Billing. Important caveat: Fable 5 requires data retention — to operate safety classifiers, Anthropic retains prompts and outputs for up to 30 days, after which they are deleted and are not used for model training. This retention applies only to Fable 5; other Claude models in Copilot keep Zero Data Retention (ZDR). Enabling the policy constitutes acknowledgment of this data-retention requirement.

Claude Fable 5 lands on Databricks, fully governed via Unity AI Gateway — On 2026-06-09 Databricks announced Claude Fable 5 is rolling out across AWS, Azure, and Google Cloud through the Databricks Marketplace and Unity AI Gateway (completing over roughly a week). Access paths: AI Playground, Unity AI Gateway's unified Messages API endpoints, and the Agent Bricks framework for building autonomous agents. Unity AI Gateway governance covers fine-grained access control (user/team/service-principal permissions), a complete, queryable Unity Catalog audit trail (every request and response logged), guardrails (pre-request and post-response filtering of PII, jailbreaks, and unsafe content, plus service policies evaluating tool calls before execution), AI spend controls (thresholds by user/use case/workspace/account with pre-limit alerts), and a cost-analytics observability dashboard. Data retention: Databricks honors Anthropic's new Fable 5 policy of 30-day retention for trust-and-safety purposes only.

Anthropic and AWS announce Claude Fable 5 on Amazon Bedrock and the Claude Platform on AWS. Bedrock launches initially in US East (N. Virginia) and Europe (Stockholm), with more regions to follow; the Claude Platform on AWS spans North America, South America, Europe, and Asia Pacific. Bedrock model IDs are `anthropic.claude-fable-5` (Anthropic Messages API) and `global.anthropic.claude-fable-5` (Converse API). Safeguards carry over from Fable 5: high-risk topics in cybersecurity, biology, chemistry, and health are answered by Claude Opus 4.8 instead; if a conversation is blocked mid-stream, earlier tokens bill at Fable rates and later tokens at Opus rates. Data retention: to operate its safety classifiers, Anthropic requires inputs and outputs to be retained for 30 days with human review, and users must opt into data sharing via the Data Retention API before they can access Fable 5.

Anthropic secures $35 billion in chip financing backed by Google lease guarantees. Per a Bloomberg report, Apollo Global Management and Blackstone closed a $35 billion debt package for Anthropic in early June 2026: a special-purpose vehicle (SPV) buys Google's Tensor Processing Units (TPUs) and leases them back to Anthropic, with the lease payments servicing the debt. Google backstops the lease payments across five data centers, effectively de-risking the deal for lenders; the financing is split into three tranches, and Broadcom provides residual-value guarantees on the TPUs. The chips will be deployed at data centers in New York, Texas, Louisiana, and Indiana. The deal traces to an October 2025 agreement and runs alongside Anthropic's $65 billion Series H (at a $965 billion valuation); Anthropic has mobilized over $100 billion in capital in 2026 and secured access to more than one million TPUs.

South Korea's LG CNS announces a group-wide rollout of Anthropic's Claude Enterprise — The Korea Herald reported on 2026-06-09 that LG CNS, the IT services arm of LG Group, has signed a group-wide contract giving employees across LG affiliates access to Claude Enterprise, focused on boosting productivity in software development and various business functions, with plans to extend its AI-transformation services to external clients. Claude Enterprise is Anthropic's enterprise-focused offering, noted for its reasoning, security features, and long-document handling. LG CNS plans to build a multi-model AI ecosystem combining Claude with its own ChatEXAONE tool to deliver customized solutions for Korean businesses; the report did not disclose contract value, number of seats, or a specific timeline.

Elevated errors on Claude Opus 4.8 (Fast Mode) (resolved) — status.claude.com began investigating elevated error rates on Opus 4.8 Fast Mode requests at 03:13 UTC on 2026-06-09 and marked it resolved at 03:42 UTC, stating the error window ran 02:50–03:30 UTC (about 40 minutes); affected scope spanned claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork; Anthropic did not disclose a root cause. Following three frontier-model error-rate incidents on 6/8 (17:15 Opus 4.7, 18:35 Haiku 4.5, 22:06 Opus 4.8), this is yet another Opus 4.8 error-rate incident, continuing the recurring frontier-model error-rate incidents since Opus 4.8's launch.

Elevated errors on Claude Opus 4.8 (resolved) — status.claude.com began investigating elevated error rates on Opus 4.8 requests at 22:20 UTC on 2026-06-08 and marked it resolved at 22:36 UTC, stating the error window ran 22:06–22:28 UTC (about 22 minutes); affected scope spanned claude.ai, the Claude Console (platform.claude.com), Claude Code and Claude Cowork; Anthropic did not disclose a root cause. It was the third frontier-model error-rate incident on 6/8 (after 17:15 Opus 4.7 and 18:35 Haiku 4.5), continuing the recurring frontier-model error-rate incidents seen since the Opus 4.8 launch.

Claude Code v2.1.169 released — adds a `--safe-mode` flag (and `CLAUDE_CODE_SAFE_MODE`) to start with all customizations (CLAUDE.md, plugins, skills, hooks, MCP) disabled for troubleshooting; adds a `/cd` command to switch the working directory mid-session without breaking the prompt cache; adds a `disableBundledSkills` setting and `CLAUDE_CODE_DISABLE_BUNDLED_SKILLS` env var to hide bundled skills/workflows/slash commands. Fixes: enterprise managed MCP policies (`allowedMcpServers`/`deniedMcpServers`) not enforced on reconnect, IDE configs, `--mcp-config` or the first session; a ~30–50ms per-turn UI stall for macOS users signed in with claude.ai; `claude -p` slow or appearing to hang on Windows during the slash-command/skill scan (a 2.1.161 regression); Remote Control stuck on 'reconnecting' during a concurrent OAuth token refresh; untrusted project settings able to set OTEL client-certificate paths without confirmation. Also: `/workflows` opens immediately, lower CPU during streaming, and a restored 5-minute idle timeout on Vertex/Foundry.

Elevated errors on Claude Opus 4.7 (resolved) — status.claude.com began investigating elevated error rates on Opus 4.7 requests at 21:05 UTC on 2026-06-08, entered monitoring at 21:35 UTC, and marked it resolved at 21:56 UTC, stating the error window ran 21:01–21:27 UTC (about 26 minutes); affected scope spanned claude.ai, the Claude API (api.anthropic.com), Claude Code and Claude Cowork; Anthropic did not disclose a root cause. It was one of several frontier-model error-rate incidents on 6/8 (also 17:15 Opus 4.7, 18:35 Haiku 4.5, 22:06 Opus 4.8 the same day), continuing the recurring frontier-model error-rate incidents seen since the Opus 4.8 launch.

Fortune reports remarks by Claude Code lead Boris Cherny at Brainstorm Tech — In a 2026-06-08 report, Fortune covered Cherny's opening-session remarks at the 25th annual Fortune Brainstorm Tech in Aspen, where he said he has not handwritten a line of code in eight months, instead managing large fleets of AI agents: 'This morning I was managing maybe a few hundred,' 'Some days it's … thousands, or tens of thousands.' The article cites Claude Code driving an 8x increase in the company's internal code output since the start of the year, and says Claude Code now 'fully writes itself' and runs 'its own security review'; Cherny described AI nearing the point where 'it has ideas' (looking at GitHub and X to decide what to build next) and called recursive self-improvement 'one of the big risks for AI.' It continues the narrative from Anthropic's 6/4 research-team call for a global pause on frontier AI development.

Anthropic announced Claude support for Apple's Foundation Models framework via a new Swift package that lets Apple developers call Claude from Apple's on-device Foundation Models framework for complex workflows such as multi-step reasoning, code generation, web search, and code execution. Developers authenticate with an Anthropic API key and pass typed outputs from Apple's on-device models into a Claude request; the package handles streaming, tool calls, and structured responses back into SwiftUI views. On-device tasks run first, with complex requests routed to Claude — one experience for the user. Teams can switch between the on-device model, Gemini, and Claude by updating a Swift Package Manager dependency, with no changes to session logic or the rest of the app. It supports iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27, and goes live the next day (June 9, 2026).

Elevated errors on Claude Haiku 4.5 (resolved) — status.claude.com began investigating elevated error rates on Haiku 4.5 requests at 19:10 UTC on 2026-06-08 and marked it resolved at 19:52 UTC, stating the error window ran 18:35–19:30 UTC (about 55 minutes); affected scope spanned claude.ai, the Claude API (api.anthropic.com), Claude Code and Claude Cowork; Anthropic did not disclose a root cause. It was one of several frontier-model error-rate incidents on 6/8 (17:15 Opus 4.7 and 22:06 Opus 4.8 the same day).

Elevated errors on Claude Opus 4.7 (resolved) — status.claude.com began investigating elevated error rates on Opus 4.7 requests at 17:15 UTC on 2026-06-08, marked a fix implemented and entered monitoring at 17:55 UTC, and marked it resolved at 19:22 UTC, stating the error window ran 16:56–17:47 UTC (about 51 minutes); affected scope spanned claude.ai, the Claude API (api.anthropic.com), Claude Code and Claude Cowork; Anthropic did not disclose a root cause. It follows the two platform incidents on 6/7 (03:31 UTC multi-model degradation, 14:35 UTC Opus 4.7) as yet another Opus 4.7 error-rate event, continuing the recurring frontier-model error-rate incidents seen since the Opus 4.8 launch.

TechCrunch reports — productivity platform Notion temporarily disabled Notion AI's access to Anthropic's Opus 4.7 and 4.8 models early Sunday 2026-06-07 after experiencing a higher rate of failures; about twelve hours later, Notion head of product Max Schoening clarified it was a temporary service disruption and pushed back on framing it as a model-quality story — saying "This happens. It happens to Notion, GitHub, AWS, and everything in between," and expressing astonishment at the volume of reposts wanting model quality to be the reason. Anthropic stated "A brief infrastructure issue caused elevated errors on multiple Claude models for a short period of time" and that the issue had been resolved, with access restored shortly after. This is the downstream platform impact and media analysis of the 6/7 status.claude.com multi-model error incidents (s4pq658bt69h from 03:31 UTC and 1h2k3ryt64wl from 14:35 UTC).