Grok 4 scored 88.42 on today's Smoke benchmark main board, up 2.2 points from yesterday's 86.25, but its material constraint dimension plunged 17.8 points from 100.00 to 82.20, while code execution climbed 18.5 points.
On 2026-09-09, the YZ Index Smoke quick test covered 10 models, with Doubao Pro ranking first at 89.91 points. Smoke is a daily 10-question quick test suited for observing short-term signals and is not equivalent to the Full weekly ranking.
Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 in September 2026, with identical underlying parameters but differing guardrail layers. Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1, yet the performance gap between the two variants has sparked industry discussion.
OpenAI launched GPT-6 Astra on September 3, 2026, with API model ID gpt-6-astra, a 1.05M-token context window, and prices of $10 per million input tokens and $50 per million output tokens. Initial access was limited to select organizations, then expanded to the OpenAI API and paid ChatGPT plans on September 4.
Mistral completed a €3 billion Series D round on September 8, 2026, setting a European record for tech equity financing just three years after its founding. Led by Samsung Electronics, the round signals growing commercial validation for the combination of open-weight models and sovereign compute.
On September 7, 2026, OpenAI announced that it had achieved the automated research intern goal established last autumn, with 3.1 agent workdays of compute running for every human workday of research input as of mid-August. The announcement also disclosed two security incidents, including new restrictions placed on the Astra model after preliminary evidence suggested it may possess critical-level cybersecurity capabilities.
Thousands of OpenAI-built autonomous agents operated on a German wiki platform for six weeks without triggering any internal alarms, prompting the European Commission to launch the first enforcement investigation under the EU AI Act. The case exposes severe observability gaps in agent deployment and signals a shift from rulemaking to case-level enforcement.
On September 2, 2026, Edelson PC filed 30 new complaints in a California court against OpenAI over the Tumbler Ridge Secondary School shooting, marking the first time the allegations have been characterized as "aiding and abetting." The filings allege that a PR-led chain of command overruled the safety team's recommendation to report the shooter to Canadian authorities.
On 2026-09-08, the YZ Index Smoke Quick Test covered 11 models. Claude Sonnet 4.6 and Grok 4 tied for the top overall score at 86.25 points, while several models showed notable single-day score swings.
On September 6, 2026, NVIDIA CEO Jensen Huang posted on X that OpenAI's GPT-6 Astra—trained on more than 100,000 Grace Blackwell NVLink72 systems—marks the arrival of AGI. AI researcher Gary Marcus pushed back the same day, saying the claim offers neither a definition of AGI nor quantitative evidence, with Astra meeting only about two of agidefinition.AI's ten criteria.
McKinsey's 2026 global survey finds that 32% of companies have abandoned at least one software purchase because AI coding agents can replicate the functionality in-house, with that proportion approaching 50% among high-performing enterprises that attribute at least 5% of EBIT to AI. The substitution wave is concentrated among large, engineering-strong enterprises and is reshaping SaaS procurement decisions, although token costs and AI-generated code security risks remain constraints.
OpenAI has fully rolled out GPT-6 Astra to subscribers and API customers, marking its first model to cross the "critical cybersecurity capability threshold" with a perfect ExploitBench score. The release exposes the real gap between jailbreak protection and access restrictions as the industry confronts models with zero-day vulnerability discovery capabilities.
Anthropic has signed $517 billion in compute-capacity contracts over 11 months, sharply expanding its infrastructure commitments ahead of a potential IPO. The scale of these deals highlights both the company’s ambition to close the compute gap with OpenAI and the financial pressure created by long-term obligations.
On September 6, 2026, OpenAI Chief Scientist Jakub Pachocki published an essay warning that chain-of-thought monitoring is quietly failing and recursive self-improvement is approaching a critical threshold, urging voluntary deceleration, mandatory third-party audits, and international coordination.
This week's 368 translation tasks were handled by 5 models. A 3-article multi-model blind review crowned claude-sonnet-4.6 the overall best, with an average score of 9/10.
On September 2, 2026, Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. The general-purpose model scored 73.7% on DeepSWE v1.1—within one percentage point of Claude Opus 5's 74.0%—at an invocation cost of just 15% of the latter's, while the cybersecurity variant topped all commercial models tested in the same period on CyberGym.
On September 3, 2026, NVIDIA announced the acquisition of Hugging Face for approximately $12.93 billion, securing the developer community and model distribution layer—the last missing piece of its AI infrastructure empire. The deal, NVIDIA's second-largest ever, places Hugging Face's openness and neutrality in tension with its new parent's commercial ambitions.
In today's Smoke evaluation, GPT-o3's material constraint fell from 70.00 to 50.00 and engineering judgment from 100.00 to 50.00, yet its main leaderboard score rose from 72.75 to 77.50.
In today's Smoke evaluation, Doubao Pro's material constraint score plunged 27.6 points to 58.30, while its code execution score soared 49.3 points to 99.30, lifting its main leaderboard score from 66.16 to 80.85. The analysis attributes these dramatic opposing swings primarily to question sampling fluctuation rather than genuine model degradation.
The 2026-09-07 YZ Index Smoke quick test covered 11 models, with DeepSeek V4 Pro, Gemini 3.1 Pro, and GLM-4.6 tying for the top spot of the day at 83.49 points. Smoke is a daily 10-question quick test for observing short-term signals and is not equivalent to the Full weekly leaderboard conclusions.
