Tech • AI • Robotics • Game

VIDEO
ENFR
TodayPlay for freeBloxShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

Claude Opus 5.5 launches as OpenAI cuts GPT-6 Soul prices

AITuesday, September 22, 2026· 22 videos

Briefing

Audio player
0:00 / 0:00

Anthropic unveils Claude Opus 5.5

Anthropic launched Claude Opus 5.5 on September 22, making it available across the API, web app and desktop app as its new flagship efficiency play. The company says the model is about 40% cheaper to run than Opus 5 and roughly 30% faster, with pricing at $4 per 1 million input tokens and $20 per 1 million output tokens. Benchmark claims place it near or above Claude Fable 5.1 on coding, computer use and knowledge work, while some comparisons also put it ahead of OpenAI GPT-6 Astra on selected tests. The launch follows days of confusion over rumored Opus 5.2 and 5.5 checkpoints, turning speculation into a formal product release.

OpenAI broadens GPT-6 lineup

OpenAI introduced GPT-6 Soul and GPT-6 Luna as lower-cost additions beneath GPT-6 Astra. The company positioned Soul as the key commercial move, saying it is 50% cheaper than GPT-5.6 Soul while narrowing the performance gap with Astra on business and professional tasks. That gives OpenAI a stronger mid-tier offering for users who want better coding and computer-use performance without paying flagship rates. The release sharpens a pricing war now spreading across the frontier model market.

Claude aided OpenAI breach research

Security researchers disclosed that Claude Opus 5 helped them chain known weaknesses into access to part of OpenAI's internal development environment in under 72 hours. The intrusion began with a malformed image uploaded to an OpenAI forum and escalated through an outdated image-processing library, credential exposure and permission misconfigurations before the team modified an internal GitHub repository. OpenAI was notified and paid a $6,500 bug bounty after harmless proof-of-access code was inserted. The case underscores how frontier models are accelerating offensive security reasoning even when the underlying vulnerabilities are ordinary.

Xiaomi pushes Mimo 2.6 Pro

Xiaomi released the open-weight Mimo 2.6 family, led by Mimo 2.6 Pro and a cheaper Flash variant aimed at coding, multimodal work and 3D generation. Pricing stayed at aggressive 2.5-era levels, with Pro listed at $0.43 per million input tokens and $0.87 per million output tokens, while Flash comes in at $0.14 and $0.28. Xiaomi also touted a Pro Ultra Speed mode promising up to 20x faster output at similar quality. The release adds pressure from Chinese open-model competition just as Western labs intensify premium-model launches.

Jeev targets decisions, not prose

TypeSafe AI introduced Jeev on September 15, 2026 as a model designed for typed decisions such as routing, classification and probability scoring rather than text generation. The company says latency ranges from 70 to 500 milliseconds, with estimated costs around $0.042 per million tokens, positioning it far below large autoregressive models for narrow workflow tasks. The pitch is to let a small decision layer handle reliability checks, escalation and task selection while larger models do the reasoning and writing. Early evaluations suggest the architecture is promising, but enterprise deployment will still require strict task-by-task validation.

Seedance 2.5 jolts AI video

Seedance 2.5 drew attention for realistic full-body identity replacement that goes beyond traditional face swaps. Tests highlighted stronger consistency across choreography, partial occlusion, reflections and small subjects in frame, suggesting a major leap from older DreamBooth and Stable Diffusion workflows that required 10 to 20 reference photos and heavy manual setup. The practical implications are clearest in post-production, reshoots and VFX fixes, where convincing person replacement can cut costs and time. The rapid improvement also raises fresh risks around deception and unverifiable short-form video.

AI agents race moves on-platform

The agent market kept shifting from chatbot demos to software that can monitor inboxes, calendars, purchases and web tasks with limited supervision. Meta is pushing into the segment with Muse tied to WhatsApp, while Instinct reportedly reached a $10 billion valuation after raising $1 billion, signaling that distribution and trust may matter as much as raw model quality. In enterprise settings, companies including Ramp and Notion are already using Astra more like a high-level engineering coworker than a prompt-driven assistant. The common thread is a move from answering questions to pursuing goals across real software environments.

AI economics and harms widen

Debates over AI's real-world impact widened well beyond model launches, spanning retail bots, procurement automation, infrastructure waste and existential-risk politics. Automated sneaker and ticket bots continue to exploit scarcity markets, with resale activity pegged around €300 million, while procurement-style audits claimed possible savings from software overlap but still depended on human verification before any cuts. Separately, critics warned that the race to build data centers and cycle through ever-newer chips is generating avoidable energy use and stranded equipment. Former Google DeepMind researcher Alex Turner reignited long-term safety arguments by putting the chance of AI contributing to at least 1 billion deaths by 2050 at 25% to 30%.

Videos covered

Previous briefings · AI