AI Briefing

Anthropic Q3 revenue hits $16.8B, growth slows before IPO

Lead · The Decoder · 2026-10-15

Anthropic achieved $16.8 billion in third-quarter revenue representing 45 percent growth over the previous quarter. This marks a significant deceleration from Q2's 140 percent surge. With a planned November IPO approaching, some investors now expect the company's valuation to drop below $1.5 trillion, a substantial decline from earlier projections. The slowdown reflects broader market cooling around AI-related valuations despite sustained demand for frontier models.

Revenue growth deceleration signals potential challenges in AI market expansion and will shape investor expectations for large-scale model developers heading into public markets.

• Q3 revenue: $16.8B, 45% growth vs Q2's 140%
• November IPO planned with valuation expectations below $1.5T
• Slowing growth reflects AI boom cooling in investor markets

Meta releases Muse Spark Open 816B multimodal model

Hacker News · 2026-10-15

Meta announced Muse Spark Open, claiming it as the most capable US open-weight model to date. The model features 816 billion parameters and processes text, images, video, and audio with support for up to one million tokens of context. Meta and Apple optimized the model to run on Apple M5 Ultra Mac Studios using the latest MLX framework, marking the first model in its class capable of running on desktop machines. Performance on benchmarks is roughly on par with leading open-weight Chinese AI models.

The release intensifies open-weight AI competition and demonstrates that frontier-scale multimodal models can now run on consumer hardware.

• 816B parameters, processes text/image/video/audio
• Runs on Apple M5 Ultra with MLX framework
• Matches Chinese open-weight leaders on benchmarks

Black Forest Labs releases FLUX 3 Image Klein model

The Decoder · 2026-10-15

Black Forest Labs released FLUX 3 Image Klein, an open-weights image generation model with 5.6 billion parameters. The model supports native 4K generation, pixel-precise editing, and generation with up to ten reference images. With 38 percent fewer parameters than its predecessor, it outperforms proprietary competitors in internal testing. The model can run on consumer GPUs with just 8 GB of VRAM.

Improved efficiency and quality enable practical deployment of high-capability image generation on standard hardware, expanding accessible AI tooling.

• 5.6B parameters, 38% fewer than predecessor
• Runs on 8GB VRAM consumer GPUs
• Native 4K generation with pixel-precise editing

Google releases Gemini 4 Argon for robotics and general tasks

The Decoder · 2026-10-15

Google released Gemini 4 Argon, a model designed to control robots, generate simulations from photos or sketches, and evaluate training videos. The model scores 87 percent on Google's own ERQA benchmark. Google positions Argon as potentially the only AI model robots need for diverse tasks. Published example prompts show that the model's demos still require substantial human guidance for optimal results.

A unified robotics-focused model could streamline deployment across autonomous systems, though demonstrated examples suggest real-world reliability remains limited.

• 87% score on Google's ERQA robotics benchmark
• Controls robots, generates simulations, evaluates training
• Requires significant human guidance in practice

Meta and Apple optimize Muse for M5 Ultra Macs

The Decoder · 2026-10-15

Meta and Apple announced a partnership to optimize Muse Spark Open for Apple M5 Ultra Mac Studios using the MLX framework. This marks the first model of its class capable of running locally on desktop machines. The optimization makes a 816-billion-parameter multimodal model viable on high-end consumer hardware, processing text, images, video, and audio with up to one million tokens of context.

Desktop deployment of frontier models shifts AI capabilities from cloud dependency toward local processing, affecting privacy and inference economics.

• First billion-parameter model running on M5 Ultra
• Uses MLX framework for optimization
• Handles text, image, video, audio locally

Nvidia releases Nemotron 3.5 Super VL for enterprise agents

The Decoder · 2026-10-15

Nvidia released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters designed for enterprise agents operating continuously. Nvidia also released two compact Bolt models for local machine deployment. Dozens of organizations are already fine-tuning Super VL for their use cases. However, maximizing the model's performance typically requires Nvidia hardware infrastructure.

Enterprise-specific agents mark a shift toward long-running autonomous systems, though hardware lock-in limits adoption flexibility.

• 120B multimodal model for enterprise agents
• Compact Bolt models for local deployment
• Dozens of organizations already fine-tuning

Anthropic grants METR independent evaluator access

TechCrunch · 2026-10-15

Anthropic announced that METR (Frontier Model Testing) researchers will have employee-level access to evaluate the company's models. METR will publish its own independent assessments of Anthropic's risks, incidents, and practices. Significantly, Anthropic has relinquished the right to block or suppress these published evaluations.

Third-party access and publication rights create external accountability for frontier model safety and practices.

• METR gains employee-level system access
• METR can publish risk and incident assessments
• Anthropic waived right to block publication

Anthropic offers Claude for $1 across all US government branches

Anthropic · 2026-10-15

Anthropic announced it is removing barriers to government AI adoption by offering Claude for Enterprise and Claude for Government to all three branches of government. This includes federal civilian executive branch agencies as well as legislative and judiciary branches at $1 per year. The move represents a substantial pricing concession to accelerate federal adoption.

Government access to frontier models at nominal cost shapes AI policy development and increases federal dependency on Anthropic's systems.

• $1/year pricing for all three branches
• Includes executive, legislative, judiciary access
• Removes adoption barriers for federal agencies

FTC probes Anthropic and OpenAI after Trump dismisses AI safety

Ars Technica · 2026-10-15

The FTC is probing both Anthropic and OpenAI following political statements characterizing AI safety concerns as exaggerated. The agency may assess fines comparable to penalties levied against companies that delay reporting data breaches. The investigation coincides with regulatory scrutiny of frontier model developers and their disclosure practices around safety incidents and risks.

Regulatory enforcement against frontier labs signals tightening government oversight of AI safety disclosures and incident reporting.

• FTC investigating both Anthropic and OpenAI
• Potential fines comparable to breach notification penalties
• Follows Trump's dismissal of AI safety concerns

Bank of America deploys 4,000 staff to fix AI-discovered vulnerabilities

The Decoder · 2026-10-15

Bank of America has deployed 4,000 people to fix software vulnerabilities uncovered by AI models, including Anthropic's Mythos, according to CNBC reporting. The effort includes 1,000 outside contractors working alongside internal teams. The scale of the remediation effort reflects both the volume of vulnerabilities discovered and the operational impact of AI-driven vulnerability identification across enterprise infrastructure.

Mass vulnerability discovery by AI models creates new operational burdens for enterprises but also accelerates security posture improvements.

• 4,000 staff including 1,000 contractors assigned
• Fixing vulnerabilities from AI models like Mythos
• Reflects scale of AI-discovered security issues

GPT-6-Astra solves 49 of 50 theoretical physics problems

The Decoder · 2026-10-15

On the TheoPhysEval benchmark, GPT-6-Astra solved 49 of 50 unpublished theoretical physics calculations with accuracy matching leading specialists. In several cases where the model appeared to fail, researchers discovered the human reference solutions contained errors. The results cost $2.40 per problem solved and suggest AI agents can match or exceed specialized human expertise on frontier research tasks.

AI solving theoretical physics problems at scale and cost indicates frontier models may accelerate discovery across highly specialized domains.

• Solved 49 of 50 unpublished physics problems
• $2.40 per problem solved
• Matched or exceeded human specialist solutions

Nscale's $35B IPO plans delayed by AI data center setbacks

Bloomberg · 2026-10-15

A flagship AI data center project in Norway is six months behind schedule as part of Nscale's planned $35 billion IPO. To manage the delay, the company is arranging floating hotel accommodations for workers to accelerate completion. The infrastructure setback represents a significant challenge to the company's growth projections and IPO timeline.

AI infrastructure delays directly impact the capital deployment capacity and revenue forecasts of publicly planned AI-focused companies.

• Norway data center six months behind schedule
• Workers housed in floating hotel to accelerate
• $35B IPO plans clouded by delays

Amazon Bee launches new personal AI devices starting at $75

TechCrunch · 2026-10-15

Amazon Bee launched its first new AI devices since acquisition, including an updated wearable and a credit-card-sized recorder that attaches to smartphones. The devices can summarize conversations, create to-do lists, and integrate with Alexa and Amazon shopping. Pricing starts at $74.99 for the entry model.

Hardware-based personal AI assistants expand the market for always-on AI beyond phones and computers into wearables.

• Wearable and card-sized recorder in new line
• Conversation summarization and list creation
• Pricing from $74.99