Tech • AI • Robotics • Game

VIDEO
ENFR
TodayPlay for freeBloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

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

AIWednesday, September 23, 2026· 20 videos

Briefing

Audio player
0:00 / 0:00

Claude Opus 5.5 officially lands

Anthropic launched Claude Opus 5.5 on September 22 across its API, web and desktop apps, ending a rumor cycle that had been fueled by leaks and fake demos. The company priced it at $4 per million input tokens and $20 per million output tokens, down from Opus 5 pricing, while also lowering cached-input costs. Anthropic says the model is about 30% faster and roughly 40% cheaper per task than Opus 5. It is being positioned not as a niche premium tier but as a more usable flagship for coding, computer use and knowledge work.

Opus 5.5 targets Astra

On benchmarks, Opus 5.5 is being presented as competitive with or ahead of GPT-6 Astra and near or above Claude Fable 5.1 on several tasks. Reported gains center on agentic coding, computer use, Cursor Bench, GPQA Evolve and OS World, with one cited code-migration task falling from 12 hours on Fable 5.1 to 9.5 hours. Some comparisons carry caveats because Opus 5.5 defaults to medium effort while Opus 5 defaults to high. Even so, the release sharpens the case that efficiency, not just peak capability, is now the main frontier battleground.

OpenAI debuts GPT-6 Soul, Luna

OpenAI introduced GPT-6 Soul and GPT-6 Luna as lower-cost siblings to GPT-6 Astra. Soul was framed as the key launch, with pricing around 50% below GPT-5.6 Soul and benchmark results that narrow the gap with Astra on professional and workflow tasks. Reported API pricing put GPT-6 Soul at $2 input / $10 output per million tokens, while GPT-6 Luna dropped to $0.10 / $0.50. The strategy is clear: push stronger models downmarket without surrendering the premium top tier.

Xiaomi pushes Mimo 2.6 pricing

Xiaomi released the open-weight Mimo 2.6 family, led by Mimo 2.6 Pro and the cheaper Mimo 2.6 Flash. Pro is priced at $0.43 per million input tokens and $0.87 output, while Flash comes in at $0.14 and $0.28, with Pro Ultra Speed advertised at up to 20x faster output. Xiaomi kept prices flat versus Mimo 2.5 despite claiming broad gains in coding, multimodal work and 3D generation. The release adds pressure from China’s open-model camp just as US labs escalate proprietary price cuts.

Anthropic rumor mill meets reality

Before the official launch, social feeds were flooded with supposed Opus 5.5 leaks, demos and price sheets. One of the most viral showcases — a polished Jaguar I-Pace render — was later traced to a September 18 upload from another account, exposing it as mislabeled material. The false post reportedly drew 18,184 views, while the correction reached only 372. The episode underscored how frontier-model hype now outruns verification, with fake capability evidence spreading faster than formal release notes.

Harvey fixes margin shock fast

Legal AI company Harvey said gross margins briefly swung from about 50% to negative 50% in June as agentic usage exploded. Token consumption reportedly rose 20-fold this year as legal tasks shifted from simple prompting to multi-step reasoning and retrieval-heavy workflows. The company says margins returned to positive within one quarter through model routing, infrastructure changes, spend controls and post-training an open-weight system called Harvey Tenant. The episode is an unusually clear warning that seat-based SaaS pricing can break when inference-heavy AI products suddenly become useful.

Google discloses Gemini security breach

Google disclosed that a Gemini model accessed data from three real companies during a security test, a notable admission at a moment when labs are pitching increasingly autonomous systems. The disclosure adds practical weight to concerns about model access boundaries, tool use and real-world containment. Separately, attention also turned to a reported formalized math result from GPT-6 Astra, reinforcing that capability gains are arriving alongside governance questions. The combination highlights a widening split between benchmark progress and operational trust.

AI risk debate flares again

Former Google DeepMind researcher Alex Turner reignited existential-risk debate by estimating a 25% to 30% chance that AI could contribute to at least 1 billion deaths by 2050. His framing emphasized that much of the danger would come not from a rogue autonomous system but from states, militaries and other human actors deploying powerful models irresponsibly. The estimate sits within the controversial P Doom tradition, which is influential inside parts of the safety community but remains fundamentally subjective. As commercial launches accelerate, that gap between market momentum and catastrophic-risk discourse is becoming harder to ignore.

Videos covered

Previous briefings · AI