This Week in AI — Week of September 28, 2026

This Week in AI is our weekly digest of the stories that actually matter in AI — sourced first party wherever possible, with the original link attached. Here’s the week of September 28, 2026.

🔝 Top story

Trump’s super intelligence rename sends Slovenia’s .si domains surging

At the United Nations General Assembly on September 22, Donald Trump said the US government would call artificial intelligence super intelligence in its official documents, using SI as the abbreviation. SI is also Slovenia’s country code, and the rush to register .si names followed almost immediately. Register.si, the national registry, now publishes daily and monthly registration charts to track what it describes as an increase in new .si domain registrations, and its own counters showed 46,066 new domains over the previous month and 20,300 in the previous 24 hours, against 321,700 total — in a country of about 2.1 million people that had been adding roughly a thousand a month before the speech.

🔗 https://www.register.si/en/news/follow-the-growth-of-si-domains/

Most of the buyers look like domain investors, founders defending their own name and people betting .si becomes as durable as .ai, rather than companies adopting it on merit. We wrote it up in more detail. Is this a durable second AI extension, or a two week story?

⚡ Models

Gemini 4 Argon raises the output token ceiling to 1 million

Google DeepMind shipped Gemini 4 Argon on September 30 with an output limit of 1 million tokens, up from 64K — the number that makes very long generations practical instead of iterative. Introductory pricing is $2 per million input and $10 per million output, reverting to $4 and $20 after the intro period, and Google publishes a footnote saying so. The part worth reading twice is the cyber posture: Argon is trained to find, validate and patch vulnerabilities and ships without cyber guardrails to trusted defenders through the Fairwind Program, while standard Frontier Safety Framework guardrails still apply to everyone else.

🔗 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/

GPT-6.1 Sol lands near Astra at a fifth of the price

OpenAI’s September 29 release frames the gain as cost per task rather than a new capability ceiling: $2 per million input, $10 per million output, and $0.10 cached input, which is 95% under standard input pricing and matters most for agents that resend the same context repeatedly. On Terminal-Bench Science 0.1 it averages $5.47 per task against $23.80 for Astra — though Astra still holds the top score at 68.1%. Model ID is gpt-6.1-sol.

🔗 https://openai.com/index/introducing-gpt-6-1-sol/

Claude Sonnet 5.5 gets faster than Opus for less money

Anthropic’s Sonnet 5.5, released September 28, generates more than 30% faster than Sonnet 5 at unchanged pricing of $2 input and $10 output per million tokens. It scores 70.6% on Terminal-Bench 4.0 against 10.3% for Sonnet 5 and 66.4% for Opus 5.5, and it is the first Sonnet to ship with Opus grade cyber safeguards. The more interesting claim is cost per task: roughly a tenth of Sonnet 5’s at the same scores.

🔗 https://www.anthropic.com/claude-sonnet-5-5

Ideogram 4.5 is built for the edit loop, not the first draft

Ideogram’s 4.5 model targets drift — the pixel shifts, color changes and texture artifacts that pile up every time you edit an image. It generates from text at native 1K and 2K, edits using up to four reference images with an optional mask, and prices across four tiers from 0.8 cents to 22 cents per image. One caveat: Ideogram published no dated announcement for it, so the September timing comes from its own API documentation and from integrators rather than from the company.

🔗 https://ideogram.ai/models/4.5

Kling 4.0 does 30 second takes with 15 references

Kling AI’s 4.0 release is built around native 30 second generation in a single pass, with no stitching, plus Omni Reference accepting up to 15 reference assets across seven subjects and 10 keyframe images for pose, expression and transitions. Prompt complexity goes to 8,000 tokens with automatic expansion, and audio is two channel stereo with improved lip sync.

🔗 https://klingai.com/blog/kling40-kling-visual-realism-creative-control-introducing-storytelling

🚀 Tools & platforms

GitHub Copilot gets computer use on the desktop

Public preview on macOS and Windows lets Copilot read app content and visual context, then click, type, scroll, drag and navigate across applications. GitHub names the target plainly: legacy or GUI-only software with no API, no CLI and no MCP integration, the long tail where an agent otherwise has nothing to call. It asks for approval before controlling an app, and organization managed settings can disable it entirely.

🔗 https://github.blog/changelog/2026-10-01-github-copilot-can-now-interact-with-desktop-apps

Higgsfield’s AI Influencer builds a character, then puts it in video

Higgsfield shipped the character builder on October 2. You start from scratch, from a reference image or from your own photo, refine through a Character Builder with 19 categories covering type, age, style and accessories, then push the result into an existing clip with Use in Genjutsu for Motion transfer or Object swap. Five free generations per user went live the next day so it can be tried before credits are spent.

🔗 https://higgsfield.ai/changelog

NVIDIA cuts DGX Spark’s entry price to $4,999

A 64GB unified memory configuration arrives October 23 through Acer, ASUS, Dell, Gigabyte, HP and MSI. NVIDIA’s own explanation for the new tier is the more interesting detail: the 128GB DGX Spark has climbed to $6,950, roughly 75% above launch, largely on RAM price inflation. The GB10 Grace Blackwell hardware is unchanged, and a single unit still handles models up to 100 billion parameters.

🔗 https://blogs.nvidia.com/blog/local-ai-dgx-spark-64GB-sync/

Suno’s Speech beta generates voice and music together

Suno calls it the first audio model that produces spoken audio and music as one cohesive track, rather than generating speech and then backing it with a generated track. It is open beta with no pricing, duration cap, plan tier or model name published yet, and Suno is candid about the rough edges: British accents may drift to Australia and back, and dramatic pauses may be very dramatic.

🔗 https://suno.com/blog/introducing-speech-beta

💬 Community is talking about

Meta moves its loss of control rules onto training itself

Meta’s update to the Meta Superintelligence Scaling Framework extends loss of control requirements to containment during training and evaluation, not just deployment. Before any reinforcement learning run it will assess whether the model being trained could compromise its own environment, and require mitigations before the run begins if so. High risk run environments must sit in internally vetted sandboxes. There is no model attached to any of this — the interesting question is whether the other labs copy it.

🔗 https://research.meta.ai/blog/developing-capable-models-responsibly

GitHub retires four Copilot models in one sweep

Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7 were all deprecated on October 2, each pointed at a named replacement. Enterprise admins may need to enable the replacements manually before they appear in the model selectors, and GitHub published no end of life date beyond the effective date and no migration path for individual plans.

🔗 https://github.blog/changelog/2026-10-02-selected-models-in-github-copilot-deprecated

OpenAI attributes a reasoning extraction campaign to Moonshot associates

OpenAI published an account of disrupting a coordinated effort to copy encrypted reasoning out of one conversation and have a model in a different conversation decrypt and transcribe it. Volume spiked on July 24 and 25 with 16,000 requests from more than 4,000 users across a wider cluster of over 15,000, and was disrupted by July 28. OpenAI attributes a core cluster to individuals associated with Moonshot AI, developer of Kimi — while stating explicitly that attribution is uncertain, the figures are attempted rather than successful extractions, and no specific model is named.

🔗 https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/

---

That’s the week. All stories above link to the original source — no aggregators in the chain. Catch the next digest next Monday, or browse the full AI tools directory while you’re here. Got feedback? Tell us.