AI NewsSeptember 25, 2026

AMD partnered with Perplexity to bring Portable Computer to Windows PCs with Ryzen AI Max processors, letting users run AI agents locally without sending data to the cloud.
AMD partnered with Perplexity to bring Portable Computer to Windows PCs built around Ryzen AI Max Series processors, announced September 24. It's a new hardware path for existing software: the same agent harness, orchestrator, and 27B model now run on AMD chips.
Ryzen AI Max pairs CPU and graphics processing with shared memory, letting a local model run alongside other apps. Inference completed on the PC skips Computer credits, and cloud escalation to 15+ frontier models remains available but requires explicit permission.
Key Capabilities:

• New hardware option: Runs on AMD Ryzen AI Max Series processors, not just NVIDIA RTX GPUs.
• No credits used: Inference completed on-device skips Perplexity's Computer credits entirely.
• Permission-gated cloud: Escalating to 15+ frontier models requires explicit user approval.

Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise, adding real-time video avatars with lip-syncing and tool calling for customer service and interactive kiosks.
Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise on September 24, adding near real-time generated video to its native voice models. Enterprises get video avatars with synchronized lip-syncing, fluid turn-taking, and asynchronous tool calling.
Customers deploy from a library of curated pre-built avatars, while custom avatars are gated behind a strict enterprise verification process to prevent misuse. Every generated audio and video stream carries an invisible SynthID watermark, and the service ships with US and EU endpoints.
Key Capabilities:

• Talking avatars: Real-time video with synchronized lip-syncing added to voice conversations.
• Misuse prevention: Custom avatars require strict enterprise verification; SynthID watermarks every output.
• Enterprise-ready: US and EU endpoints with provisioned throughput and compliance controls.

fal launched H3 Max Extend Video, letting users upload up to 15 seconds of existing footage and extend it to 30 seconds while preserving the original's look and feel.
fal launched H3 Max Extend Video on September 23, letting users upload up to 15 seconds of existing footage and extend it to 30 seconds while preserving the original's look and feel. It's powered by H3 Max, fal's post-trained version of MiniMax's open-weight H3.
fal's post-trained H3 Max already ranks #1 on independent benchmarks from Artificial Analysis and Design Arena for overall quality, prompt understanding, and aesthetics. The extend feature turns short generated clips into longer usable footage.
Key Capabilities:

• Double the length: Extends up to 15 seconds of footage to 30 seconds automatically.
• Consistent style: Preserves the look, motion, and feel of the original clip.
• Top-ranked base: Built on H3 Max, ranked #1 for quality on independent video benchmarks.

Higgsfield AI says it's on track to top $1 billion in annualized revenue within a year while already cash-flow positive, and it's giving away $20 million in API cashback to celebrate.
Higgsfield AI said on September 24 it's on pace to cross $1 billion in annualized revenue within a year, up from about $50 million a year ago, and already cash-flow positive since May. Revenue under contract from business customers has grown tenfold since June.
To celebrate, Higgsfield launched a $20 million cashback program offering 100% back on API spend across models like Seedance 2.5 and Kling 3.0, up to $100,000 per business, through September 30. About 70% of the company's activity now comes from commercial advertising.
Key Capabilities:

• Fast growth: On pace for $1 billion in annualized revenue, up from $50 million a year ago.
• Already profitable: Cash-flow positive since May, unusual among AI video startups.
• Cashback promo: 100% back on API spend, up to $100,000 per business, through September 30.

Sarvam AI released Vision 2.1, a document AI model that extracts structured data from forms and tables and reads handwritten text across English and 22 Indian languages, topping key OCR benchmarks.
Sarvam AI released Vision 2.1 on September 23, adding structured extraction from multi-page tables and forms plus handwritten Indic text recognition across English and 22 Indian languages. It scores 87.3 on olmOCR-Bench, ahead of Gemini 3.6 Flash and Opus 5.
It places second on OmniDocBench, behind PaddleOCR-VL, but Sarvam says the two scores together make it Pareto-optimal. The model uses a layout parser and reading-order network feeding a vision-language model, available now via Sarvam's Digitize and Extract APIs.
Key Capabilities:

• Handwriting support: Reads handwritten text across English and 22 Indian languages.
• Top benchmark scores: 87.3 on olmOCR-Bench, ahead of Gemini 3.6 Flash and Claude Opus 5.
• Structured output: Extracts key-value pairs and table structure, not just raw text.

Cursor reduced token costs by 7% with no drop in agent quality, through tighter prompts, selective tool loading, better caching, and reading fewer line numbers when scanning files.
Cursor announced on September 23 that changes across its agent harness cut token costs by 7% with no drop in agent quality. The system prompt and tool definitions, sent every turn, were among the biggest controllable costs, so tightening them cut cold cache misses by 20%.
Another fix targeted file reads: instead of numbering every line, the agent now numbers only every tenth, still enough to cite code accurately, cutting cache-read tokens by 1.6%. Longer agent runs also delegate more work to subagents that start with fresh context windows.
Key Capabilities:

• Trimmed context: Tighter system prompts and selective tool loading cut cold cache misses by 20%.
• Smarter file reads: Numbering only every tenth line cut cache-read tokens by 1.6%.
• No quality trade-off: All changes were validated to preserve agent output quality.