ElevenLabs breaks $500M ARR in June, Leanstral 1.5 saturates miniF2F, Fable 5 deployed in Claude Tag

July 4, 2026 is marked by an exceptional review on the generative media side: ElevenLabs crosses $500M in annual recurring revenue in June, while shipping Dubbing v2, Music v2, Ads Engine, and Avatars. In parallel, Mistral releases Leanstral 1.5, which saturates miniF2F at 100% and establishes itself as SOTA in Lean 4 formal verification. On the Anthropic front, Fable 5 is now available in Claude Tag, the internal team tool, and researchers at the University of Waterloo show that a Qwen3-4B distilled on 2.3M Fable 5 traces achieves zero hallucinations.

ElevenLabs — June 2026 review: $500M ARR, Summit in Warsaw, Dubbing v2 and Avatars

July 3 — Mati Staniszewski (CEO of ElevenLabs) publishes the full review of June 2026, which marks several simultaneous records for the company.

Key metrics for June:

Metric	Result
ARR (annual recurring revenue)	$500M crossed
Voice Library payouts	$22M paid out to contributors
ElevenLabs Summit	2,500 executives gathered in Warsaw
Government partnerships	4 new partner governments

Models and research:

Dubbing v2 — The new dubbing model preserves the emotion, performance, and intent of the original content, solving the chronic problem of flat dubbing disconnected from its context.
Music v2 — Significant improvements in vocals, instrumentation, arrangement, and multilingual multi-genre support.

Tools for customer experience teams:

Procedures — Packaged instruction sets defining agent behavior in common scenarios (the equivalent of standard operating procedures for employees).
New modalities — Agents can now handle images, files, audio notes, contacts, and locations in conversations.
Speech Engine — Converts an existing chat agent into a fully voice-enabled agent with a single prompt.

Tools for creators and marketers:

Ads Engine — Automatic localization of ads into 50+ languages, with direct reinjection into advertising platforms.
Avatars — Generation of talking-head videos from a script, a voice, and a photo.
Flows Agent — Building complete creative workflows in natural language inside a node-based canvas.

ElevenLabs also announced new watermarking techniques for content authenticity. Crossing $500M ARR positions ElevenLabs among the few audio generative AI players to have reached this threshold at this stage of market maturity.

🔗 June 2026 review — Mati Staniszewski

Leanstral 1.5 — Lean 4 formal verification SOTA, Apache-2.0 (Mistral)

July 2 — Mistral releases Leanstral 1.5, a formal verification model in Lean 4 under the Apache-2.0 license, with 119 billion total parameters and only 6 billion active parameters (MoE architecture).

Formal verification benchmarks:

Benchmark	Leanstral 1.5 score	Context
miniF2F	100% (saturation)	Validation and test
PutnamBench	587 / 672	+7 vs Seed-Prover 1.5 High, at ~ $4/problem vs ~$ 300
FATE-H	87%	SOTA
FATE-X	34%	SOTA

The model operates in two environments: a multi-turn environment (prove or refute a theorem in a loop with the Lean compiler) and a code agent environment (file system access, bash, Lean language server). Training proceeds in three stages: mid-training, supervised fine-tuning, and reinforcement learning with CISPO.

In practice, an Aeneas pipeline (Rust → Lean) coupled with Leanstral discovered 5 new bugs in 57 open source repositories. Test-time scaling is notable: 44 problems solved at 50,000 tokens → 587 at 4 million tokens.

Leanstral 1.5 is available for free via Hugging Face and via a free API (leanstral-1-5). The recommended usage is with Mistral Vibe:

uv tool install mistral-vibe
vibe --setup
# puis /leanstall dans Vibe, vibe --agent lean

🔗 Official article — Mistral

Fable 5 now available in Claude Tag

July 2 — Anthropic announces that Claude Fable 5 is now available in Claude Tag, the internal tool that enables teams to work with Claude every day. The announcement is accompanied by an interview video between Boris Cherny (Head of Claude Code at Anthropic) and Cat Wu, which retraces the path from Claude Code to Claude Tag and how adoption expanded from engineering to the rest of the company.

A conversation with Boris Cherny and Cat Wu on the path from Claude Code to Claude Tag, and how it spread from engineering to the rest of Anthropic. Claude Fable 5 is now available in Claude Tag. — A conversation with Boris Cherny and Cat Wu about the path from Claude Code to Claude Tag, and how it spread from engineering to the rest of Anthropic. Claude Fable 5 is now available in Claude Tag.

Claude Tag is Anthropic’s product that allows professional teams to interact with Claude in a structured way. This availability of Fable 5 in Claude Tag extends the flagship model’s capabilities to team workflows. The video had gathered 450,800 views and more than 3,000 likes at the time of the scan.

Sakana AI — Ren Ito appointed founding commissioner to the UN “AI for Good” Commission

July 3 — Sakana AI announces that its co-founder Ren Ito has been appointed founding commissioner of the World “AI for Good” Commission of the International Telecommunication Union (ITU), a UN specialized agency. The bilingual article (Japanese/English) published on the Sakana AI blog states that this is the first appointment of a representative from a Japanese AI lab to this level of international AI governance.

Sakana AI, founded by David Ha and Llion Jones (former Google Brain researchers) and based in Tokyo, is known for its research into nature-inspired architectures. This appointment places the lab on the global AI governance stage alongside institutional actors and established technology companies.

🔗 Sakana AI — AI for Good Commission

HuggingFace Diffusers — New release with Ideogram4, MotifVideo, DiffusionGemma

July 3 — Sayak Paul (ML engineer at Hugging Face) announces a new version of the Diffusers library, reposted by the official @huggingface account. This release integrates several new image and video pipelines:

Pipeline	Type
Ideogram4	Image generation
MotifVideo	Video generation
DiffusionGemma	Gemma-based diffusion (popular)

The Diffusers library is the open source reference for using diffusion models — Stable Diffusion, FLUX, and their variants. Each release expands the catalog of pipelines available directly in Python without manual weight configuration.

🔗 Announcement @RisingSayak

Distillation of 2.3M Fable 5 traces into Qwen3-4B — zero hallucination, open source

July 4 — Researchers at the University of Waterloo have distilled 2.3 million reasoning traces from Claude Fable 5 into Qwen3-4B, and are releasing the weights as open source. The results are remarkable:

Metric	Result
Self-consistency at 512 samples	100%
Output entropy	0.00 bit
Hallucination variance	Zero

“we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B — 100% self-consistency @ 512 samples — 0.00 bits output entropy — zero hallucination variance — turns out the student is not bounded by the teacher.” — We distilled 2.3M reasoning traces from Claude Fable 5 into Qwen3-4B — 100% consistency, zero output entropy, zero hallucination variance — the student is not limited by the teacher.

The tweet, reposted by @huggingface, gathered 1.6 million views and 8,000 likes. The phrase “the student surpasses the teacher” summarizes a phenomenon documented in the distillation literature: under certain training conditions, a distilled model can converge toward a more deterministic solution than the source model.

ElevenMusic — Launch of Tools: Voice to Song

July 2 — ElevenMusic, ElevenLabs’ music product, launches a suite of creative tools called Tools, designed to create, reshape, and evolve music.

Introducing Tools on ElevenMusic, purpose-built ways to create, reshape, and evolve your music. — We are launching Tools on ElevenMusic, tools designed to create, reshape, and evolve your music.

The first highlighted feature is Voice to Song: the user sings and receives a faithful, professional-quality studio version. A Loop Studio is also visible in the tweet’s demo video.

Briefs

Claude Code v2.1.201 — Single fix targeting Claude Sonnet 5 sessions: harness reminders no longer use the current mid-conversation system role. Maintenance update with no new exposed behavior. 🔗 CHANGELOG
HuggingFace CEO — 250 open US AI milestones for July 4 — Clément Delangue publishes a collection of 250 open source AI achievements from the United States (Attention is All You Need, PyTorch, GPT-2, LLaMA, ImageNet, LoRA…) to mark the 250th American anniversary. 🔗 @ClementDelangue tweet
Gemini CLI v0.51.0-nightly.20260702 — Security fix: an attacker could use a symbolic link to escape the allowed directory in the memory import processor (PR #28233). 🔗 GitHub release
Runway Engineering — Nighttime recovery of idle GPUs — The Runway platform team details “deckard”, their controller that reallocates idle production GPUs at night to research, relying on Erlang-C queueing theory and 5 predefined time windows. 🔗 Runway article
GLM-5.2 in Claude Code via HuggingFace Inference Providers — GLM-5.2 (Z.ai) is now selectable in Claude Code via Hugging Face’s hf-claude mechanism, confirmed by an official repost from @Zai_org. 🔗 Announcement

What this means

The audio AI economy is reaching maturity. ElevenLabs’ crossing of $500M ARR in June, combined with the$ 22M paid back to Voice Library contributors, gives a concrete picture of what monetization at scale means in generative audio. Dubbing v2, Music v2, Ads Engine, and Avatars are not experimental features: they are products aimed at marketing and creative teams with specific use cases (ad localization, music production, automated videos). ElevenMusic Tools, with Voice to Song, extends this logic toward individual creators. The convergence of these announcements suggests a market that is no longer in the exploration phase but in the industrialization phase.

Formal verification is becoming accessible. Mistral’s Leanstral 1.5 represents a turning point for formal verification: saturating miniF2F at 100%, solving 587 PutnamBench problems at a cost one hundred times lower than Seed-Prover 1.5 High, and automatically discovering new bugs in open source repositories — all under Apache-2.0 and via a free API. Formal verification used to be reserved for specialists equipped with expensive systems. A model accessible via uv tool install mistral-vibe changes the profile of the teams that can benefit from it.

Distillation goes beyond its own theoretical limits. The University of Waterloo work — 2.3M Fable 5 traces in Qwen3-4B, zero hallucination, open source weights — confirms a documented but still surprising phenomenon: a distilled model can converge toward a more deterministic solution than the teacher model. Combined with Fable 5’s availability in Claude Tag (integration into Anthropic team workflows) and GLM-5.2 support in Claude Code via HuggingFace, it is the tooling ecosystem that is expanding: frontier models are less and less locked inside their proprietary environments.

AI governance is also being shaped at the institutional level. The appointment of Ren Ito (Sakana AI) to the World “AI for Good” Commission of the ITU is the first concrete signal of a Japanese AI lab’s presence in UN AI governance bodies. In a context where most industry representatives in these forums are American or European, this appointment broadens the geography of deliberation on global AI standards.