
Fable 5.1 Leads Real-SWE at 38.8%; Code Benchmark Halves Scores
Specific Labs' new benchmark tests code agents on licensed private corpora. No frontier model exceeds 40%, with a drop of over 42 points against SWE-Bench Pro.
Strategy
222 analyses

Specific Labs' new benchmark tests code agents on licensed private corpora. No frontier model exceeds 40%, with a drop of over 42 points against SWE-Bench Pro.

Allianz cuts 1,800 in call centers, Mizuho eliminates 5,000 clerical jobs, Ergo reduces 1,000 positions in Germany. Meta asks ICs to return to management roles.

Workgroups from the three companies have met since July to establish shared safety benchmarks, transparency protocols, and external auditing.

After reallocating 7,000 engineers to Applied AI and facing morale issues, Meta allows return to original roles; Zuckerberg acknowledges mistakes and rules out more cuts in 2026.

OpenAI released the Agents API on September 10, featuring a managed sandbox and no extra charges beyond consumption. Cloudflare, DigitalOcean, and Oracle are environment partners.

DeepSeek announced V4.1 Flash on September 10, with 552B MoE, 1M context tokens, and automatic routing to Flash from Pro starting September 14.

Google secured a 22-year PPA with Fortum for half of Loviisa's output and announced a €13 billion AI infrastructure initiative in Finland, its largest European investment to date.
Announcement in Shanghai places Alibaba Cloud and Cambricon on the board of the PyTorch Foundation with Ant Group as Gold, a coordinated move by China's biggest names in AI within the framework born at Meta.

The consulting firm confirmed on September 7th the hiring of 1,500 new graduates in the US, scaling the Frontier team to 15,000 and doubling the Synapse initiative target from 1 million to 2 million. Ravi Kumar S. refers to the model as player coaches.

The three companies released AI models for security in restricted access programs in the same week. This marks the monetization of a vertical that was confined to research reports until 2025.

Anthropic, Google, Meta, and OpenAI stacked models in five business days. Corporate buyers can no longer complete a comparison before the next generation.

Broadcom's Q3 report saw AI chip revenue reach $16.7 billion. On the other side of the ledger are banks like MUFG, Mizuho, and SMBC, contracting OpenAI and Sakana to rewrite the back office.