AI NewsOctober 2, 2026

Three new models handle both ends of a voice conversation: one turns speech into text as someone talks, the other two turn text into natural-sounding speech in 23 languages.
Microsoft AI released MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.1-Flash on October 1. The transcription model ranked #1 on an independent benchmark, with a 2.5% word error rate (a standard accuracy measure, lower is better) and results appearing just 0.13 seconds after someone stops speaking.
CEO Mustafa Suleyman said it's the most accurate real-time transcription model available today. One voice in MAI-Voice-2.1 can speak all 23 supported languages while keeping the same accent and identity, useful for things like a multilingual tutoring app or customer support line.
Key Capabilities:

• Fast and accurate: 2.5% word error rate, the lowest among models tested on this benchmark.
• Same voice, many languages: One voice can switch between 23 languages without sounding different.
• Built for speed: The Flash variant trades some quality for lower cost and latency.

FLUX 3 Image generates and edits images using bounding boxes, a grid system that tells the model exactly where each element belongs, aimed at pixel-perfect product shots and layouts.
Black Forest Labs released FLUX 3 Image on October 1. Users place elements on a 0-to-1000 grid with IDs and short descriptions, then the model renders everything inside its assigned box. Edits to one box leave the rest of the image untouched, which BFL says stays bit-identical, meaning the pixel values literally don't change.
It supports up to 10 reference images per composition and generates at up to native 4K. Partners including Higgsfield AI, Krea, and OpenRouter integrated it immediately, with open weights and a 50% API discount through October 8 still to come.
Key Capabilities:

• Bounding-box layout: Place and edit elements with exact coordinates on a 0-1000 grid.
• High resolution: Generates natively at up to 4K.
• Multi-reference: Composes a scene from up to 10 separate reference images.

Built by fine-tuning MiniMax's H3 video model, HeyGen Video starts at a penny per second through October, well below rivals like Veo 3.1 and Kling 3.0 Pro.
HeyGen released HeyGen Video on September 30, priced at $0.01 per second during an October launch discount (half its $0.02 standard rate). It's live now through HeyGen's API and partners including OpenRouter, Runware, and ComfyUI. HeyGen says it performs comparably to pricier rivals in blind side-by-side tests.
The model caps out at 768p, below MiniMax H3's own 2K option, so it targets volume use cases like ads and social clips rather than high-end production work.
Key Capabilities:

• Low price: $0.01 per second through October, among the cheapest video models available.
• Built on MiniMax H3: A fine-tuned version of MiniMax's base video model, not built from scratch.
• Resolution cap: Tops out at 768p; no 2K option like the base MiniMax H3 offers.

GLM 5.3 and its faster sibling, GLM 5.3 Flash, are now selectable inside Cursor. Open weights means anyone can download and run the model themselves, unlike closed models from OpenAI or Anthropic.
Cursor added GLM 5.3 and GLM 5.3 Flash from Chinese lab Z.ai on October 1. On Cursor's own CursorBench 4.0 chart, GLM 5.3 Max is the best-scoring open-weight model, landing near Sonnet 5.5 and Astra on the quality-versus-cost curve, though still a step behind the top closed models.
Developers have been asking Cursor to add GLM for weeks, since open models often cost far less to run than proprietary ones. GLM 5.3 ships under the permissive MIT license, with a 1 million-token context window.
Key Capabilities:

• Open weights: Free to download, run locally, or host yourself.
• Long context: Supports up to 1 million tokens, roughly 750,000 words, in one conversation.
• Three effort levels: Low, High, and Max reasoning modes, with High as the default.

[9:45 AM]

OpenAI reopened sign-ups for its $200 Pro plan, but changed the math so new subscribers get roughly half the usage value of the old plan at the same price.
OpenAI reopened $200 Pro sign-ups on September 30 after a three-week pause caused by demand for GPT-6 Astra. Product lead Tibo Sottiaux says the new usage allowance nets out to about half the old plan's value when converted to API pricing, now 10 times Plus usage instead of 20 times.
Existing subscribers keep their old allowance until October 29, after which it also drops. OpenAI isn't bringing back the old 5-hour usage window, and reactions have been sharply mixed, with some users calling it a scam and others welcoming the lower sticker shock on a new $500 Pro tier.
Key Capabilities:

• Allowance cut: New subscribers get 10x Plus usage, down from the previous 20x.
• Grace period: Existing users keep old limits only until October 29.
• New tier: A $500/month Pro plan now offers 25x Plus usage.

Since launching September 22, Opus 5.5 has led most independent benchmarks against GPT-6 Sol and GPT-6.1 Sol, at a higher price per task.
Anthropic's Claude Opus 5.5 scores 58 on Artificial Analysis's Intelligence Index at max effort, ahead of GPT-6 Sol's 48. It also leads on Terminal-Bench 4.0 (66.4% vs. 43.9%) and most knowledge-work tests, while GPT-6.1 Sol, released a week later, narrowed some of that gap.
The trade-off is cost. Opus 5.5 runs about $5.98 per task on that index, compared with roughly $1.06 for GPT-6 Sol, so OpenAI's models remain the cheaper pick for high-volume, well-defined work.
Key Capabilities:

• Independent lead: 58 vs. 48 on Artificial Analysis's Intelligence Index at max effort.
• Coding edge: 66.4% on Terminal-Bench 4.0, versus 43.9% for GPT-6 Sol.
• Price gap: Costs roughly 5-6 times more per task than GPT-6 Sol on the same index.

Tavus says its Griffin model is the first to pass a video Turing test, holding real-time video conversations where it listens, watches, and talks at once, like a person.
Tavus introduced Griffin on October 1, calling it the first Human Interaction Model. In a blind study of 54 people on one-minute video calls, 48% believed they'd spoken to a real human, compared with under 3% for earlier systems. It scored 3.83 out of 5 on NVIDIA's video benchmark, against 3.92 for an actual human.
Griffin is a research preview, available only as the smaller Griffin-Lite to select testers. Tavus says a model this convincing needs more safety review before a public release, given the risk of impersonation or scams.
Key Capabilities:

• Full-duplex design: Listens, watches, and speaks simultaneously, rather than waiting for silence to respond.
• Live reactions: Demos show it reading a held-up object or coaching a Rubik's cube solve from what it sees.
• Limited release: Only Griffin-Lite is available now, to a small group of testers.