Tech • AI • Robotics • Game

VIDEO
ENFR
TodayPlay for freeBloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

OpenAI Dots, GPT-6.1 Soul and Anthropic Sonnet 5.5 surge

AIWednesday, September 30, 2026· 23 videos

Briefing

Audio player
0:00 / 0:00

OpenAI turns ChatGPT into Dots

OpenAI used Dev Day 2026 to launch Dots, always-on agents that can monitor work, act proactively and communicate across SMS, phone, email and Slack. The product pushes ChatGPT from prompt-response chat toward persistent digital coworkers with cloud computers, app permissions and specialist roles. Space was introduced as a collaborative workspace to share context, files and agent workflows across teams. The strategy is clear: keep more enterprise work, memory and execution inside OpenAI's own cloud stack.

GPT-6.1 Soul resets pricing

GPT-6.1 Soul arrived as OpenAI's new value-performance model at $2 per million input tokens and $10 per million output tokens, far below GPT-6 Astra at $10 and $50. OpenAI presented it as near-Astra quality on coding, computer use and professional tasks at roughly one-fifth to one-sixth the price. Reported benchmark figures included 75% on Deep Sway, 31.7% on Automation Bench, and about 57% on Terminal Bench Science at sharply lower cost per task than top rivals. The company also expanded its pricing ladder with GPT-6 Luna and a new $500-per-month plan with much higher limits.

Anthropic answers with Sonnet 5.5

Anthropic re-entered the center of the model race with Claude Sonnet 5.5, pitched as more than 30% faster and up to 30% cheaper per task than Sonnet 5. The model carries a 1 million-token context window, up to 128,000 output tokens, and pricing of $2 input and $10 output per million tokens. On one leaderboard it ranked third overall, and on Artificial Analysis's Intelligence Index it scored 56, only two points behind Opus 5.5 Max. The release signals a sharper Anthropic focus on coding, agentic workflows and cost-efficient frontier performance.

Sonnet 5.5 wins coding showdown

In a side-by-side game-generation test, Claude Sonnet 5.5 produced the strongest overall result against Sonnet 5, Opus 5.5 and Fable 5.1. All four models were given the same demanding prompt to build a Minecraft-style 3D world with architecture-first planning, stable movement and environmental effects. Sonnet 5.5 was judged best on visual quality, coherence and runtime stability while still undercutting higher-end models on speed and cost. The result strengthens the case that mid-tier models are now good enough to displace premium systems for many practical coding tasks.

Google tests Gemini 4 Argon

Google DeepMind quietly moved Gemini 4 Argon into limited release through its Fairwind program for selected cybersecurity partners and testers. Introductory pricing matches the new aggressive market norm at $2 per million input tokens and $10 per million output tokens, before rising to $4 and $20. The standout specification is a jump in maximum output from 64,000 to 1 million tokens, making the model notable for long-running agentic and software workflows. Google also cited a 77.9% result on SWE-bench 1.1, positioning Argon as a serious everyday multimodal contender.

AI security shifts to exploitability

Standard Chartered and Sophos separately outlined how frontier AI is reshaping cyber defense from raw flaw discovery toward exploitability and response speed. Standard Chartered said the key problem is now identifying attack paths across applications, APIs, identities and trust boundaries, not simply finding isolated vulnerabilities, and pointed to early work with Daybreak. Sophos reported that AI-assisted operations have cut average investigation times from 38 minutes to 89 seconds in many cases through its Fusion platform and more than 500 integrations. Both accounts underline the same pressure: the defender window is shrinking as AI accelerates software understanding for attackers and defenders alike.

US-China harden AI duopoly

A Trump-Xi summit in Washington put artificial intelligence alongside trade, Taiwan and Iran at the top of bilateral strategy. Publicly, the meeting produced little beyond a pledge to continue dialogue on AI risks and benefits, possibly including a direct channel between Washington and Beijing. That fell well short of binding rules and reinforced the sense that the two powers prefer managed rivalry to formal global governance. The optics were significant because the talks overshadowed simultaneous UN General Assembly and UN Security Council discussions on AI in New York.

Tesla app hints at Optimus Gen 3

Assets embedded in Tesla's Android app version 4.61 appear to show a possible Optimus Gen 3 robot design, though the files offer no proof of final hardware or public functionality. The package contains nine PNG renders in a mock asset folder, including filenames such as gold and gen 3, with one labeled robot home fullbody gen 3. File comparisons indicate the images are authentic shipped assets and first appeared in an end-of-August update, with higher-resolution versions added in early September. For now, the discovery is best read as a product clue rather than a confirmed roadmap reveal.

Videos covered

Previous briefings · AI