The frontier changed hands again this week. Anthropic's Claude Fable 5.2 moved past Google's Gemini 4 Carbon on the Artificial Analysis Intelligence Index, the independent scorecard both companies cite. Meta went the other direction with a model priced at $0.10 per million input tokens, built to do the grunt work while a bigger model plans. OpenAI kept its most capable security model behind a vetting process.

Anthropic and Meta Split the Frontier Into Planners and Workers

Anthropic released Claude Fable 5.2 on Monday (Nov. 9), alongside Claude Mythos 5.2, a restricted sibling for vetted cybersecurity and life sciences organizations. Fable 5.2 scored 65 on the Artificial Analysis Intelligence Index, ahead of Gemini 4 Carbon at 60 and Anthropic's own Claude Opus 5.5 at 59. Anthropic says the model generates output up to 30% faster than Fable 5.1 and that its safety filters block about 55% fewer legitimate requests. Pricing is $8 per million input tokens and $40 per million output tokens. Tokens are the small chunks of text AI companies use to meter usage. Anthropic named Jump Trading, Balyasny Asset Management and XTX Markets as users for quantitative research. It said Box used the model to review loan application files and flag a mismatch between salary deposits and a bank statement balance.
Meta released Muse Ember, its cheapest model yet, on Nov. 9 at $0.10 per million input tokens. Meta built it to pair with its larger Muse Wonder: Wonder plans a task and Ember executes the repetitive steps, such as tool calls, data extraction and long-document parsing. It accepts text, images and audio, so an agent can read a scanned statement or a voice memo without a second model. Meta says it costs less per task than Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6.1 Luna.

OpenAI Locks Down Its Cyber Model as Rivals Go Cheaper

OpenAI released GPT-6 Cyber on Nov. 5 to approved customers in its Daybreak Red security program, with no public access. OpenAI says the model found 81 previously unknown high-severity vulnerabilities in Firefox and built 948 working exploits against the Linux kernel in testing. OpenAI rated the model Critical for cybersecurity under its Preparedness Framework, the first OpenAI model to reach that level. That rating is why access is gated. OpenAI says the model stayed within researcher-set boundaries in 99.93% of tests.
Thinking Machines Lab released Inkling 2 Small on Nov. 5, the first model in its second generation. The startup, founded by former OpenAI Chief Technology Officer Mira Murati, says the model has 594 billion parameters but uses only 17 billion on any request. Parameters are the internal settings a model learns in training. It costs $0.30 per million input tokens and $1 per million output tokens. Thinking Machines says it will release the weights under an Apache 2.0 license, which allows free commercial use, after outside safety testing.
Baidu released ERNIE 5.5 Flash Base on Nov. 5 with open weights. The 280-billion-parameter model uses 14 billion per request and supports a 1-million-token context window, meaning it can take in roughly 1,500 pages of text at once. Baidu positions it as a base for companies to fine-tune on their own data rather than a finished assistant.

Google Brings AI to the Factory Floor; Ricursive Targets Coding

Google DeepMind released Gemini Robotics 2.5 to robotics partners on Tuesday (Nov. 10). In a pilot on live Mercedes-Benz production lines in Hungary and Germany, Apptronik's Apollo 2 humanoid robots running the model sorted metal brackets into trays in vehicle order with a 96% success rate. Google says the model resists jailbreak attempts 99.5% of the time, up from 82% for its predecessor, Gemini Robotics ER 2. Boston Dynamics and Agile Robots are also deploying it.
Ricursive Intelligence released Ripple-65B Preview on Nov. 4 with open weights. The Palo Alto startup was founded by former Google researchers to design AI chips with AI. It says the 65-billion-parameter model is the strongest U.S. open model for agentic coding, in which AI writes and runs code on its own. Ripple scored 80.5 on Terminal-Bench 2.1 against 80.1 for Reflection AI's Beam. Ricursive says the full project cost $32 million, which it estimates is about a tenth of what Reflection spent on Beam.

Meta's Muse pair makes the split explicit: one model to plan, one to execute. For financial firms, the pricing question is shifting from which model to buy to how to route each task to the cheapest one that can handle it.

Anthropic's Claude Tops AI Rankings as OpenAI Limits Cyber Model | PYMNTS.com