AI News: September 23, 2026

1. Xiaomi’s MiMo-V2.6-Pro Tops the Open Models, and Anthropic Says Claude Paid for It

Xiaomi. MiMo-V2.6-Pro scores 46 on Artificial Analysis’s Intelligence Index, the highest among openly available models, ahead of Kimi K3 and Qwen, at $0.435 per million input tokens and $0.87 per million output tokens, which works out to roughly $0.13 per test task. Pro is a 1.02T parameter mixture-of-experts with 42B active per request, alongside a smaller Flash. Xiaomi credits a heavily scaled reinforcement learning phase that ran under six days for about $2.62 million on Pro and $0.85 million on Flash, lifting DeepSWE coding scores from 58.4 to 72.6 on Pro and 48.8 to 65.7 on Flash. The catch: Anthropic accuses Xiaomi of improperly siphoning training data through Claude. Source

2. Snorkel AI Raised $350 Million at $3.5 Billion as Training Data Becomes the Bottleneck

Snorkel AI. The Series E, led by Insight Partners and S32, values the seven-year-old company at nearly triple the $1.3 billion it carried 17 months ago. Annualized revenue run rate is $375 million, an eighteenfold increase over twelve months. Snorkel pivoted last year from data-labeling automation software to selling completed datasets and RL environments as a service, generating data synthetically alongside subject matter experts. Because it sells environments rather than human labor, expert payments sit in cost of goods sold rather than inflating the headline revenue figure the way they do at Mercor or Handshake. Source

3. Snapdragon 8 Elite Extreme Gen 6 Runs a 30B MoE Model on the Phone

Qualcomm. The two new flagship processors announced at Snapdragon Summit put local inference at the center. Both add sensing hubs that run models up to 200 million parameters, enough for an always-on local scribe with speaker differentiation and usage-based memory for task automation, and Qualcomm says a full voice-in, voice-out agent can run on the chip. The Extreme variant runs a 30-billion-parameter mixture-of-experts model locally, against the 20B MoE Apple shipped as its third-generation foundation model at WWDC in June. Motorola announced the Signature 27 on the Extreme part. Source

4. Toyota Has Workers Teaching the Robots That Will Staff 400,000 Positions

Toyota. From 2028 Toyota plans to invest $6.42 billion annually in factory robotics, targeting 150,000 robots in its own automotive plants and 250,000 more across group component and materials facilities. Wheeled ELEY humanoids are already on assembly lines learning from human workers, who wear jigs modeled on the robots’ fingers to teach precise hand movements. Toyota executive vice president Hiroki Nakajima says the goal is coexistence rather than replacement. Toyota employs 18,000 veteran workers across 60 factories worldwide. Source

5. British Columbia Is Suing OpenAI for the Cost of Rebuilding a Demolished School

British Columbia. The province filed a complaint Monday seeking the cost of replacing Tumbler Ridge Secondary School, demolished in August after a February shooting in which eight people died, and demanding the shooter’s ChatGPT logs. Victims’ families had already sued over OpenAI’s failure to warn law enforcement after discovering in advance that the shooter was using ChatGPT to plan the attack. The province’s filing is the first accounting of the disaster response costs, and asks for changes to prevent ChatGPT-assisted violence in addition to damages. Source

6. AstroForge Is Putting a Transformer in Command of a Deep Space Spacecraft

AstroForge. The asteroid mining startup built Solo, an in-house transformer-based autonomous control stack, and plans to fly it in 2027 on Stoke Space’s first launch, on a NASA-backed solar science mission. Most spacecraft autonomy still uses traditional control algorithms because neural networks are considered unreliable, and the first neural network control of satellite positioning happened only last year. The motivation is economic: CEO Matthew Gialich frames it as choosing between roughly $200 million for five ground dishes plus operations, or putting enough intelligence onboard that the spacecraft recovers itself. AstroForge lost contact with its Odin probe in 2025. Source

7. A Startup Says Measurements From Lab-Grown Neurons Make Video Models 5x Faster

The Biological Computing Co. TBC announced a partnership with AWS to sell a text-to-video model it claims runs five times faster at 80 percent lower inference cost than the open-source base model it builds on, which the company declines to name. No biological hardware is involved: the product is a proprietary adapter layer adding under 0.1 percent to model size, derived from growing cortical neurons on 4,096-electrode chips and measuring how activity spreads and decays. The plan is Trainium for training, SageMaker AI for serving, and AWS Marketplace for distribution, but only an early access signup exists today. TBC raised $25 million in February. Source

8. GitLab Duo Self-Hosted Can Now Point at Models in Your Own Azure Tenant

GitLab. GitLab Duo Self-Hosted added support for models deployed through Microsoft Foundry, covering OpenAI GPT, Anthropic Claude, Meta Llama, and Mistral families. The architecture is a self-managed GitLab instance, a self-hosted GitLab AI Gateway, and one or more Foundry model endpoints, with the gateway sitting between Duo features and models rather than binding features to a provider. Feature-level model selection means Code Suggestions, agentic workloads, and high-volume tasks can each run a different model. The tradeoff is that capacity, networking, credentials, and model lifecycle become the platform team’s problem. Source

9. Nscale’s IPO Puts 85 Percent Customer Concentration in Front of Public Investors

Nscale. The British neocloud has amassed over $103 billion in contracts since spinning out of Arkon Energy two years ago, but $43.8 billion comes from a Microsoft compute deal through 2033 and $44.6 billion from an Anthropic supply agreement that is contingent on Nscale securing financing and cancellable if it misses milestones the filing calls “stringent.” Nscale plans to list on the NYSE at an expected $35 billion valuation, seeking $3 billion. Revenue for the six months to June 30 was $140.6 million against $10.4 million a year earlier, with net losses rising to $1.02 billion from $369 million. Source