- NATURAL 20
- Posts
- Gemini 4 Argon Arrives as OpenAI Targets a $1.4 Trillion Valuation
Gemini 4 Argon Arrives as OpenAI Targets a $1.4 Trillion Valuation
PLUS: Gemini replaces Gems with Skills, and Ideogram 4.5 targets multi-turn editing drift.
Wes Roth
October 02, 2026
In partnership with
The best in influencer marketing. And you’re invited.
Get ready for Return on Influence Festival '26. An entire day dedicated to influencer marketing with speakers running some of the best programs in the world.
Already booked:
Maya Shaff, ŌURA
Leah Walker, Adobe
Georgia Humphries, Stanley 1913
Tyler Vaught, Edelman
Josh Rangel, Ogilvy
and many more to be announced soon…
It’s free, it’s online, and you’ll take away something you can apply to your job the very next day. Scout’s honor!
Today:
Gemini 4 Argon brings a 1M-token output limit to long-horizon professional work
OpenAI reportedly seeks at least $30 billion at a $1.4 trillion valuation
DeepSeek open-sources Huawei Ascend tooling built around TileLang
Ling-3.1-flash brings a 560B MoE design and up to 1M context
Ideogram 4.5 targets image drift across repeated edits
Gemini 4 Argon Pushes Long-Horizon AI Into a New Scale
Google’s new Gemini 4 Argon is built for work that stretches across large codebases, long documents and multi-step professional tasks. At the same time, OpenAI is reportedly preparing another huge private funding round, while DeepSeek and Huawei are trying to close one of the biggest infrastructure gaps in China’s AI stack: the software layer around domestic accelerators.
Together, the three stories show how the frontier is widening. Model capability still matters, but so do output length, operating cost, financing, chip software and the ability to keep difficult workflows running for much longer.
Gemini 4 Argon is Google’s new frontier model for long-horizon software engineering, enterprise knowledge work, multimodal reasoning and cybersecurity defense. Google is beginning with a phased rollout through its Fairwind Program for trusted cyber defenders while it gathers feedback and tests safeguards before broader developer, enterprise and consumer access.
The headline specification is a one-million-token output limit, up from 64,000 tokens on Google’s previous generation. That gives the model room to sustain much longer reasoning and generation trajectories for tasks such as large code migrations, deep financial or legal research and multi-stage technical work.
Google reports 77.9% on DeepSWE v1.1, 51.3% on AutomationBench, 65.4% on Vals Finance Agent v2 and 68.0% on CWE-bench v1. The model does not lead every benchmark: Google’s own table shows it behind GPT-6 Astra on FrontierSWE v2 and Terminal-Bench Science and behind Claude Opus 5.5 on Terminal-Bench 4.0.
Argon will launch at an introductory API price of $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95%. Google has not announced a broad public release date, so the initial launch is more limited than the pricing and benchmark announcement might suggest.
OpenAI is reportedly in early discussions to raise at least $30 billion in a new private funding round at a valuation of roughly $1.4 trillion before the new capital is added. The talks come after the company pushed its planned public offering beyond 2026.
The proposed financing would act as a bridge to a later IPO and would follow the company’s March round, which included $122 billion in committed capital at an $852 billion valuation. Investor demand is reportedly driving the new discussions, but the size, valuation and terms could still change.
The timing reflects the amount of capital required to keep expanding frontier-model training, inference capacity and consumer and enterprise products. OpenAI’s annualized revenue run rate has also been reported near $70 billion, giving the company a much larger revenue base than it had during earlier fundraising rounds.
No new round has been finalized. Until financing closes, the $30 billion target and $1.4 trillion valuation should be treated as reported negotiating targets rather than completed transaction terms.
DeepSeek has released open-source programming infrastructure developed with Huawei for the Ascend family of AI accelerators. The release includes TileLang, a high-level language for writing performance-critical AI kernels, along with compute and communication libraries designed to make Ascend hardware easier to program.
The goal is to reduce dependence on Nvidia’s CUDA ecosystem, which remains the dominant software platform around AI accelerators. DeepSeek says TileLang uses a simpler programming model while still aiming to expose the performance of the underlying hardware, and the new work includes optimizations for Huawei’s Ascend 950 chips.
The companies are also working on larger Ascend deployments, including a 128-chip supernode design, while DeepSeek has reportedly planned a data center using at least 160,000 Huawei accelerators. The software release matters because replacing Nvidia requires more than competitive chips; developers also need mature languages, kernels, communication libraries and tooling.
The release does not mean CUDA has already been displaced. Nvidia’s ecosystem remains far more established, but DeepSeek and Huawei are now making more of the software needed for a domestic alternative freely available to developers.
🧠RESEARCH
RIDE improves on-policy distillation by extrapolating reinforcement-learning changes directly in hidden representation space instead of noisy output probabilities. Across four base-and-teacher model pairs, the method approaches or exceeds the RL-trained teacher on every pair and consistently outperforms output-space extrapolation, especially when teacher and base models are already similar overall evaluations.
ReaLVR trains latent visual reasoning models to preserve the image evidence that actually determines an answer, addressing weak supervision in hidden reasoning tokens. Across three model families, it reaches a 63.7% five-task average on Qwen2.5-VL-7B and continues improving at scales up to 235 billion parameters in reported experiments overall evaluations.
UniEvo-VL lets one multimodal model improve image generation by learning from its own critiques instead of relying on a larger external teacher. Built on Qwen-image-2512, the method raises GenEval from 0.747 to 0.808 and GenEval2 Soft-TIFA from 32.97 to 35.53, though gains were uneven on text-rendering tasks in reported experiments.
📲SOCIAL MEDIA
🗞️MORE NEWS
Ling-3.1-flash is a new mixture-of-experts model with roughly 560 billion total parameters and about 25 billion activated per token, aimed at agents, search, office work and software development. The model is designed for up to a one-million-token context window, although its initial two-week free trial is capped at 256,000 tokens. Ant Ling says the larger context will become available after the trial and plans to open-source the model at the same time.
Ideogram 4.5 is built around precise multi-turn editing, where repeated changes can otherwise introduce pixel shifts, color drift and accumulating artifacts. The company says the new model better preserves details as users edit the same image repeatedly instead of forcing a full regeneration after each change. It is live in Ideogram, the API and launch-partner platforms, with open weights planned for a later release.
Gemini Skills let users save reusable instructions and trigger them with a slash command instead of rewriting the same prompt each time. Skills can include reference files, combine with other skills and be created from existing conversations, with rollout beginning globally in Gemini chat and later expanding across Workspace plans. Google plans to retire Gems in stages and automatically migrate personal and Workspace Gems into Skills as those older features are phased out.
OpenAI says it disrupted a coordinated adversarial-distillation campaign that tried to recover protected reasoning from its models without breaking encryption, databases or stored user conversations. The company says activity began in early July, spiked to about 16,000 extraction-pattern requests from more than 4,000 users over July 24–25, and was linked in part to individuals associated with Moonshot AI; these are OpenAI’s findings and attribution. OpenAI says it closed replay pathways, strengthened checks that prevent protected reasoning from appearing in streamed output and shared indicators with industry and government partners.
The Federal Trade Commission has opened an investigation involving major AI developers and evaluators, including OpenAI, Anthropic and METR, focused on potential consumer risks from increasingly autonomous agents. The agency can use legal information demands and executive testimony as it examines whether existing consumer-protection law applies to cybersecurity incidents and other harms involving agentic systems. The probe is an investigation rather than a finding of wrongdoing, and the companies had not publicly been found liable for the conduct under review.
What'd you think of today's edition? |
Reply