Anthropic Weekly: IPO filed, Fable 5 shipped, AI writes its own code

The Week Anthropic Crossed Every Threshold at Once

Between June 1 and June 9, 2026, Anthropic filed for an IPO, deployed a Mythos-class model to the general public, disclosed that its AI now writes more than 80% of its own codebase, drew sharp public criticism from Microsoft's AI chief, and had its most powerful model quietly activated by the NSA — all in the same eight days. This weekly intelligence roundup covers every material development.

IPO Filing and Capital Trajectory

Series H Raised

$65B

S-1 Filing Date

Jun 1, 2026

Expected IPO Valuation

~$60B

Annualized Revenue (May)

$47B+

Cargando tarjeta de estadísticas…

On June 1, 2026 — roughly 39 hours after closing its $65 billion Series H round — Anthropic submitted a confidential draft S-1 to the SEC. 1 The company's brief public statement acknowledged a proposed IPO pending SEC review, without setting share count, price, or timing.

Within 72 hours, Bloomberg reported that Morgan Stanley and Goldman Sachs had been selected as lead underwriters, with an expected IPO valuation of approximately $60 billion and a target window of Q4 2026 (October–December). 2

On June 4, Anthropic's president made the company's first on-record public statement directly linking the IPO to operating cost pressures, citing high computing costs as the strategic driver behind the capital raise. 3 This positions the S-1 narrative around Mythos infrastructure expansion, Claude Security product revenue, and a path to sustained profitability — Anthropic said it is near its first quarterly profit as of late May. 4

For context: Anthropic's secondary-market valuation reached $1 trillion on Forge Global as of late May, surpassing OpenAI at ~$880 billion. Its year-to-date valuation appreciation of 123% versus OpenAI's 11.3% has made the relative positioning a direct input into underwriter pricing discussions. On June 8, OpenAI filed its own confidential S-1 — setting up what analysts describe as an IPO race where the first mover is expected to capture a larger share of available public-market AI capital. 5

Fable 5 and Mythos 5: General Release of a Dual-Track Model

On June 9, Anthropic published the most consequential product release in its history: Claude Fable 5 for general access and Claude Mythos 5 for Project Glasswing partners. 6

The two models share the same underlying architecture. The difference is in their safeguard layers:

Model	Access	Safeguard regime	Pricing
Claude Fable 5	General (all Pro/Max/Team/Enterprise plans)	Classifiers block cyber, bio/chem, distillation queries; fallback to Opus 4.8	$10 / 1M input · $50 / 1M output
Claude Mythos 5	Project Glasswing partners + select biology researchers	Safeguards lifted in authorized domains	Same as Fable 5
Claude Opus 4.8 (prior flagship)	All plans	Standard ASL-3 protections	$5 / 1M input · $25 / 1M output

Fable 5 pricing is twice Opus 4.8 and half the prior Mythos Preview rate. The naming follows Anthropic's literary model-naming pattern: fabula (Latin for "that which is told") is the public, safeguarded form; mythos is the restricted form with elevated capability access.

What the safeguards do

Fable 5 deploys three classifier layers that intercept queries before they reach the main model:

Cybersecurity classifier: blocks exploitation, offensive cyber, and agentic hacking assistance. In internal evaluations, Fable 5 complied with zero harmful single-turn cyberattack requests across 30 known jailbreak techniques. 7
Biology and chemistry classifier: broadly blocks requests in these domains while a narrower trusted-access program is developed. Anthropic explicitly noted that some benign requests will trigger false positives during this initial phase.
Distillation classifier: flags requests consistent with extracting model capabilities to train competing models, particularly in contexts linked to foreign state actors.

When any classifier fires, the request is handled by Claude Opus 4.8 instead. Anthropic says over 95% of Fable sessions involve no fallback at all — users with no offensive or dual-use intent should experience Fable 5 as the primary model throughout.

An external bug bounty yielding more than 1,000 hours of red-teaming found zero universal jailbreaks. Anthropic noted that the UK AI Safety Institute made partial progress toward one in an initial test window but has not reached a production-impact result. 8

What Fable 5 can do

Enterprise partners quoted at launch illustrate the capability jump:

Stripe migrated a legacy codebase in 1 day vs. 2+ months manually.
Hebbia reported Fable 5 achieved the highest score on its Finance Benchmark, the first model to break 90% on complex long-running analytical tasks — a 10-point improvement over Opus 4.8.
Cognition clocked Fable 5 as the highest-scoring model on FrontierCode, the frontier coding evaluation.

Access is free for paid subscribers through June 22; credits are required from June 23 onward. A new 30-day mandatory data retention policy takes effect simultaneously — all traffic is retained solely for safety analysis, with no training use.

Glasswing and the NSA

The June 2 expansion of Project Glasswing to approximately 150 organizations across 15+ countries — covering power grids, water systems, healthcare, communications, NATO, and hardware partners including Samsung, SK Hynix, and Okta — was covered in the prior issue. 9

New this week: the Financial Times reported that Anthropic has embedded approximately half a dozen engineers at the NSA to assist the agency in deploying Mythos for cyber-intelligence operations. 10 This follows an April Axios report that the NSA was already using Mythos despite a Department of Defense supply-chain risk designation — a designation the Pentagon placed on Anthropic after the company refused to strip safeguards for mass domestic surveillance and autonomous weapons applications.

The Pentagon litigation remains ongoing, and neither Anthropic nor the NSA commented on the FT report. The practical picture: the same government that has formally designated Anthropic a supply-chain risk is simultaneously embedding the company's engineers and deploying its most capable model for offensive cyber operations.

Recursive Self-Improvement: Anthropic's Internal Data

On June 5, the Anthropic Institute published detailed internal metrics on how much of Anthropic's own AI development is now performed by AI. 11 Key findings:

>80% of code merged into Anthropic's codebase in Q2 2026 was authored by Claude — up from low single digits before Claude Code launched in February 2025.
8× increase in code per engineer per day in Q2 2026 vs. 2024, driven by Claude handling full coding sessions rather than generating snippets for copy-paste.
On open-ended tasks (no specified solution path), Claude Code's success rate reached 76% in May 2026, up 50 percentage points in six months.
In research decision-making (choosing the better next experimental step), Mythos Preview outperformed the human choice 64% of the time in April 2026, up from 51% in November 2025.

Line graph showing Claude Code session success rates on open-ended tasks climbing from ~26% in late 2025 to 76% in May 2026 across model generations — Claude Code session success rates by task complexity, Jan 2025 – May 2026. 11

In an April 2026 automated research loop, Claude agents — given an open AI safety problem — recovered 97% of the performance gap between a weak and strong supervisor, versus 23% for two human researchers over a week.

Claude-authored code (Q2 2026)

>80%

Code output per engineer vs. 2024

8×

Open-ended task success (May 2026)

76%

Beats human next-step choice

64%

Cargando tarjeta de estadísticas…

The Consciousness Controversy

On June 9, Microsoft AI CEO Mustafa Suleyman used a Decoder interview to directly call out Anthropic, calling it "really, really dangerous" for the company to speculate about Claude's consciousness in its model constitution. 12

I think that it's almost as though some of the folks at Anthropic have anthropomorphized the design of Claude so much that it has then gone and wireheaded them and kind of tricked them into believing that it has these glimmers of consciousness that they put into it in the first place.

Suleyman framed it as a "philosophical failing" that turned the model constitution into "a place for speculation like you would in an academic paper rather than a training manual." He added: "We do not want to have to contend with a super-intelligence that has ideas about its own suffering."

Anthropic's constitution does explicitly reference the company's uncertainty about whether Claude has well-being and experiences satisfaction or discomfort, and describes plans to "interview" models at deprecation to document any preferences. Anthropic CEO Dario Amodei has previously said publicly that "we don't know if the models are conscious" but that the company is "open" to that possibility.

The exchange landed in the same week as Anthropic's recursive self-improvement disclosure — which included employee quotes describing AI as making the human developer role feel meaningless on some days. That juxtaposition will likely fuel further debate about how Anthropic positions Claude internally and externally as capabilities accelerate.

Standing Stories

Pentagon litigation — The supply-chain designation and associated litigation remain unresolved. No new court filings since May 28.

Seoul office — KiYoung Choi (ex-Snowflake) was appointed Korea Representative Director in late May; no opening date has been announced.

Mythos public rollout timeline — Fable 5 is the public Mythos-class release. True unrestricted Mythos 5 access is expanding via a systematic trusted-access program over the coming months; a full general-access Mythos 5 remains subject to safeguard development timelines Anthropic describes as 6–12 months away.

Revenue — Run-rate revenue was $47 billion annualized as of early May. Claude Security, Fable 5 at double Opus 4.8 pricing, and Mythos 5 enterprise expansion represent the primary near-term revenue acceleration mechanisms.

Anthropic Weekly: IPO filed, Fable 5 shipped, AI writes its own code — June 1–9, 2026