Gemini 4 Argon goes GA with a 1M-token output limit

Overview

10 stories in this issue. The first 3 are today's priorities.

Hot model updates

1. Top · Gemini 4 Argon goes GA with up to 1M output tokens
2. Top · TokenSpeed reports 685 tok/s per user on Qwen3.8-Flash-Next
3. Top · Singapore signs an AI safety MOU with Anthropic as Claude use grows there
4. Anthropic asks the full D.C. Circuit to rehear the Pentagon's Claude blacklist

Global AI news

5. X Subscription bundles Grok, Cursor and X Premium at $8–$200 a month
6. NVIDIA commits $250M in compute over five years to UK agentic safety research
7. Scimplify raises $25M for AI-driven specialty chemicals and critical materials

Regional and early signals

8. Dobot's VEGA 7 seven-axis force-controlled arm claims 50,000-hour MTBF
9. StepFun's STEPX Neo agent phone goes on sale at 3,999 yuan
10. Qwen Office opens an invite-only e-commerce edition

Jiufeng graphic based on the sources cited in this issue.

Hot model updates

On September 30, Google opened Gemini 4 Argon only to cyber-defense partners. It is now available on the paid Gemini API, AI Studio and the Gemini app at $2 input and $10 output per million tokens.

• Architecture: sparse mixture-of-experts. Parameter count not disclosed, trained on TPUs, no open weights
• Context: a 1M-token window shared by input and output; a single response can be up to 1M tokens
• Pricing: $2 input / $10 output per million tokens on the paid Gemini API
• Access: developers use the API through AI Studio and the Gemini platform. AI Ultra subscribers also get it in Google Antigravity, and Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search
• Cyber defenders: can apply separately to the Fairwind Program for Argon access without cyber safeguards

The table shows selected Argon scores compiled by MarkTechPost. Argon leads the Vals Index and ties GPT-6 Astra on CWE-bench v1. On FrontierSWE v2 it ranks last of the 4 models in Google's comparison table.

Limitations: Under the new tier structure, free users and Plus subscribers only get the weaker Flash models. MarkTechPost rates its agentic coding results as mixed. The Decoder says Argon is aggressively priced but trails Claude Opus 5.5 and a Meta model in some comparisons. The Decoder also reports that Google is testing an internal successor codenamed "Carbon" that is said to be stronger at coding; Google has not confirmed this.

Source: Google blog · MarkTechPost · The Decoder · Fairwind Program

LightSeek Foundation says its open-source TokenSpeed inference engine produced 685 output tokens per second per user running Alibaba's Qwen3.8-Flash-Next. Qwen posted congratulations on October 14.

• Model: Qwen3.8-Flash-Next
• Hardware: four NVIDIA GB300 GPUs
• Precision: NVFP4
• Result: 685 output tokens per second per user
• Engine: TokenSpeed, code published on GitHub

Limitations: LightSeek has not disclosed prompt length, concurrency or other benchmark conditions. RuntimeWire notes that without them, operators can't infer much about real deployments from the figure. This is a self-reported result and has not been reproduced by a third party.

Image source: GitHub; mirrored on Jiufeng R2.

Source: TokenSpeed on GitHub · LightSeek post · RuntimeWire

Anthropic opened its fifth Asia-Pacific office in Singapore and signed an MOU with the Ministry of Digital Development and Information covering AI safety, evaluation and responsible adoption.

Anthropic calls Singapore one of its most active markets worldwide. It says Claude already powers services from GovTech Singapore, Grab and DBS that millions of people use every day. Permanent Secretary Kai Fong Chng and Anthropic co-founder Ben Mann signed the MOU. Dale Finlay, Anthropic's General Manager of ASEAN, leads the local team.

Limitations: This is Anthropic's own announcement. The excerpt lists only the areas the MOU covers, with no specific projects, funding or timeline, and no independent coverage was found.

Source: Anthropic announcement

On October 13, Anthropic asked the full U.S. Court of Appeals for the D.C. Circuit to rehear its appeal over the Pentagon's blacklist designation.

The petition argues that the panel overlooked evidence that the Pentagon was punishing Anthropic for its public position on AI safety. If the court grants a rehearing, the appeal would be reopened.

Limitations: This is only a petition; the court has not decided whether to rehear the case. The excerpt gives no details of the panel's earlier ruling, and this item has a single reporting source.

Source: RuntimeWire

Global AI news

SpaceX Super Intelligence has combined X Premium, Grok and Cursor into a single plan called "The X Subscription," priced from $8 to $200 a month.

TechCrunch frames the bundle as a bid for more subscribers in a paid AI market where rivals include OpenAI and Anthropic. X's official subscriptions account announced the bundle.

Limitations: The excerpt does not list what each tier includes.

Source: X Subscriptions post · TechCrunch

NVIDIA has committed $250 million in compute over five years to strengthen U.K. research on agentic AI safety and security. It is also joining the Earthshot Prize as a Global Alliance Founding Partner.

• Fuse Energy: moving its customer AI onto in-house NVIDIA B300 Blackwell Ultra infrastructure, serving more than 450,000 U.K. households
• Locai Labs: released Nemotron-Personas-UK, an open dataset of one million synthetic personas reflecting the U.K.'s demographic, regional and cultural diversity
• Earthshot Prize: NVIDIA will provide compute and technical expertise to climate and sustainability innovators

Limitations: All of these details come from NVIDIA's own blog. The excerpt does not say which institutions will receive the $250M in compute or how it will be allocated.

Source: NVIDIA blog

Bengaluru-based Scimplify raised $25 million in a strategic round. Its annualized revenue run rate has tripled to $200 million, and it plans to expand into the U.S.

• Lead investor: Hitachi Ventures. Customers Henkel and Teknor Apex joined as new investors
• Existing backers: Accel India, Bertelsmann India Investments, Omnivore and 3one4 Capital
• Total equity raised: $80 million
• Growth markets: semiconductors, defense and energy

Limitations: The company declined to disclose its valuation. The $250M post-money figure comes from a person familiar with the deal. After its $40M Series B in March 2025, the company was valued at $150M.

Source: TechCrunch

Regional and early signals

Dobot launched VEGA 7, a seven-axis force-controlled arm aimed at wheeled dual-arm robots and embodied AI.

• Joints: a torque sensor and brake in every joint
• Force control accuracy: 0.15N
• Payload and reach: 7kg payload, 725mm reach
• Reliability: 50,000-hour mean time between failures (MTBF)

Limitations: Pandaily is the only source so far, and the excerpt gives no price or delivery date. The MTBF figure is a vendor claim and has not been independently tested.

Source: Pandaily

StepFun's device brand STEPX launched an early-access edition of STEPX Neo, its first agent phone, in Shanghai. Pre-orders are open at 3,999 yuan, with first sales on October 16.

• What the price covers: the phone, Step AOS 1.0, the Step Studio developer tool and Step model services. Buyers get a Token Plan of 10 billion tokens a month for 12 months
• Device-cloud split: an on-device model handles simple commands such as settings and alarms, and some work offline. Cloud models plan and break down complex multi-app tasks, which the device then carries out
• Hardware: a six-microphone array, a low-power audio chip and the Whisper 1.0 voice-pickup algorithm, plus a rear MiniLED dot-matrix screen that shows the agent's status

Limitations: Chinese-language source. This is an early-access edition, and the details come from the launch event and on-stage demos. No model parameters, on-device model specs or third-party tests have been published.

Source: QbitAI · InfoQ China (Chinese-language source)

On October 14, Qwen Office launched an e-commerce mode for online merchants. Access is currently by invitation only.

Features include a cross-platform operations dashboard, trend and selling-point analysis, batch generation of marketing assets, and cross-platform ROI optimization. It already connects to merchant tools including Shengyi Canmou, Youzan and Qianniu, and plans to add more merchant tools and industry software.

Limitations: Chinese-language source, and Leiphone is the only outlet reporting it. The edition is invite-only, and the source gives no pricing and does not say which Qwen model version it uses.

Source: Leiphone (Chinese-language source)