A Free Mystery Model Just Beat GPT-5.6 and Claude at Coding, Google's Old Guard Is Walking Out the Door, and Two Labs Moved Pricing in Opposite Directions

A free, anonymous model beat GPT-5.6 and Claude Fable 5 at coding. It's Zhipu, and it has a name now

On August 20, OpenRouter quietly listed a new model under the identifier stealth/ox-alpha, no company name, no parameter count, no technical report, priced at zero. The listing described it only as "a reasoning model designed for coding, sustained agentic work, and production workloads." Within 48 hours, independent testing on the DeepSWE coding benchmark had it completing 80% of tasks correctly, ahead of Claude Fable 5 at 65% and GPT-5.6 Sol at 52%. It carries a 1,048,576-token context window and accepts text, images, and video. Stripe's CEO reportedly tested it directly and summed it up in two words: "quite impressive."

The guessing game that followed had real technical teeth, not just speculation. Developers ran tokenizer comparisons and found a constant 75-token difference from Zhipu AI's GLM-5.3. A separate video-encoder test matched Zhipu's GLM-5V-Turbo token-by-token. An API error format echoed Zhipu's services, consistent with a pattern where every prior anonymous "Alpha" model on OpenRouter turned out to be a Chinese lab testing under a cover name: Pony Alpha was Zhipu's GLM-5, Hunter Alpha was Xiaomi's MiMo-V2-Pro, Elephant Alpha was Ant's Inclusion AI, Owl Alpha was Meituan's LongCat-2.0.

On August 26, the guessing ended. Z.ai (Zhipu AI's international brand) confirmed to Bloomberg that Ox Alpha was a new entry in its GLM series, and released model weights the same evening. The official name is GLM-5.3-Flash, a 320-billion-parameter model with 18 billion active parameters, natively multimodal across text, images, video, and visual documents, released under an MIT license on Hugging Face. The free preview window closed with the reveal; the named API now runs at $0.075 per million input tokens and $0.25 per million output at a 50% launch promotion, rising to $0.15 and $0.50 after September 9. Ox Alpha reportedly hit #1 on OpenRouter's usage leaderboard before the reveal, more than doubling DeepSeek's usage, and China-based labs have now led global model usage on the platform for 18 straight weeks.

Google DeepMind's founding generation is heading for the exits

Demis Hassabis is stepping back from day-to-day leadership of DeepMind to become Alphabet's Chief Scientist, a real title change Google has confirmed. Alongside that shift, several of Google's most senior technical veterans, more than 20 years at the company in most cases, are reported to be departing: Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals among them, names attached to MapReduce, Bigtable, Spanner, TensorFlow, and much of what became Gemini. Reporting points to a new startup focused on automating scientific research as the destination for at least some of the group, though the company's structure, funding, and full roster remain unconfirmed as of this writing. A talent exodus at this level, from people who built foundational infrastructure the rest of the industry still runs on, is the kind of story worth watching closely regardless of where it lands.

OpenAI slashed prices 80% the same fortnight DeepSeek raised them 93%

Within roughly two weeks in mid-August, GPT-5.6 Luna became ChatGPT's free default after an 80% price cut, from $1 to $0.20 per million tokens, while DeepSeek raised its V4 Flash pricing 93%, from $0.14 to $0.27 per million, in the same window Gemini reportedly crossed 1 billion monthly active users. Neither company has framed the move as a response to the other, and pricing shifts this size happen for plenty of reasons that have nothing to do with rivalry, infrastructure cost, margin targets, promotional windows expiring. Still, two frontier-adjacent labs moving list price in opposite directions inside the same two weeks is a genuine signal that the pricing floor under this market hasn't settled, whatever the individual reasons turn out to be.