AI Daily Brief — 2026-10-26

Sebastian Raschka is developing a video course on building reasoning models, a topic he has extensively covered on Ahead of AI. Reasoning models, which include most current-generation LLMs, remain highly relevant. The course will begin with a pre-trained base model rather than training one from scratch, to then add reasoning methods. An evaluation step will assess the utility of these additional methods for the model.

Hurston-1 is an open-source text-to-speech (TTS) model designed for phone use, capable of generating accented English speech, specifically Southern American and Afro-English. It is steered by plain-language voice descriptions and built using a corpus that includes scripted reading, conversational speech, and approximately 10 hours of real phone calls with AI voice agents from consented contributors. The model demonstrates high confidence in various African accents, such as East African (0.98) and Nigerian (0.94), while showing low confidence in general American (0.02) and Southern American (0.01) accents.

Microsoft launched its Decision-1 model on October 9, 2026, built on Alibaba’s Qwen3.5-9B. This model is designed to score fixed answer options in a single forward pass, providing probabilities rather than generating text, summaries, or emails. Microsoft claims Decision-1 runs 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol, with approximately 35x lower P50 latency. It supports text inputs up to 32,000 tokens and is available via Microsoft Foundry and OpenRouter.

OpenAI has launched GPT-6.1 Terra, a new model designed for "fast and capable computer use at scale." This release emphasizes clearer, more natural writing and combines detailed visual reasoning. The pricing for GPT-6.1 Terra is set at $1/1M input tokens and $5/1M output tokens, aiming to deliver efficient performance.

Hob is a free, open-source AI tool designed as an on-device alternative to services like Claude and ChatGPT. It allows users to interact with AI on their own computers, supporting tasks such as filling out PDFs and working through documents. Hob automatically selects the most suitable open-weight model for the user's computer, offering a private and offline experience. It is available for Mac, Windows, and Linux, and is developed by Frigade under an MIT license.

A dermatologist compared the diagnostic accuracy of GPT-6 Astra and Claude Fable 5.1 for skin cancer against a 10-year-old CNN, Inception v3. Across 200 lesions, Inception v3 achieved 62% accuracy, outperforming GPT-6 Astra (42%) and Claude Fable 5.1 (42%). For specific diagnoses like Melanoma, Inception v3 also showed higher accuracy (43%) compared to GPT-6 Astra (27%) and Claude Fable 5.1 (37%), indicating that the older CNN was superior.

ChatGPT is increasingly recommending 404 Media to users experiencing common tech support issues, suggesting the publication can help resolve their problems. This trend stems from the lack of human customer support at major tech platforms, leading users to AI chatbots for assistance. Many users, like a man from Russia seeking help with a banned Instagram account, are then directed to 404 Media, with some even having the chatbot compose their outreach emails.

Google DeepMind's AI hurricane forecasting model demonstrated superior performance compared to conventional models by accurately predicting Hurricane Isaias's landfall in Florida. This highlights the increasing adoption of artificial intelligence in hurricane forecasting, a technology that experts believe is rapidly transforming the field and enhancing predictive capabilities.

OpenAI and Anthropic significantly reduced their LLM prices, by up to 80%, within a few months, largely due to competition from smaller open-weight models like DeepSeek V4 Flash 0420. The DeepSeek model, released in April 2026, offered a key competitive advantage with its low cache read price of $0.0028, making its effective cost much lower. This competitive pressure, along with the general trend of decreasing LLM prices, forced major providers to lower their rates to remain competitive.

Anthropic's S-1 filing reveals that the company spent only $482M on capital expenditures in the first half of 2026. The company primarily leases its computing resources, with its capital expenditure largely outsourced through $782B in lease obligations. This strategy means Anthropic owns very few assets, despite its significant revenue per employee, which reached $6.6 million.

Anthropic's S-1 filing reveals its post-IPO governance structure will maintain Public Benefit Corporation status. This arrangement grants its seven co-founders 50.1% voting control through a single Class F share. The company unveiled its US initial public offering paperwork on Monday, preparing for a significant market debut.

Anthropic's IPO documents reveal an unusual succession plan, detailing confidential, ranked lists of up to ten potential successors for each co-founder. These lists, which outline who would inherit voting power, require approval from all five co-founders. This private line of succession is a notable feature within the company's IPO filings, as reported by Anna Tong for Forbes.

Cloudflare was alerted on September 8, 2026, by Gadi Evron to a tenant isolation issue in the GLM-5.3 Workers AI model, active since August 27, 2026. The issue, which caused cross-request KV cache exposure, was remediated, and the model was restored on September 9. This bug, reported to SGLang as issue #30233 on July 6, 2026, led to patches in SGLang v0.5.20, v0.5.21, and v0.5.22, with Cloudflare recommending updates to v0.5.22 or later.

Anthropic is cautioning potential investors that its commitment to AI safety and advocacy could lead to disproportionate and selective targeting by the Trump administration. This warning comes as Anthropic prepares for one of the largest IPOs in history, suggesting that its public stances might continue to make it a target for the administration.

IBM is currently experiencing a severity one incident in its Classic Cloud, acquired from SoftLayer in 2013. This "emergency maintenance" has prevented users from provisioning new resources or virtual private clouds for 90 minutes, with the incident expected to last at least 16 hours until 08:00 UTC on Thursday. The Register is investigating if critical bugs in PHP are the cause.