Gemini 4 Carbon, Nemotron 3.5 Transcribe, olmOCR 2 archives

Four items today. One frontier launch with a limited rollout, one open speech model, and two science stories that show where this work lands outside the labs.

Gemini 4 Carbon

Google announced Gemini 4 Carbon today and calls it its strongest model, with a focus on cybersecurity and long-horizon work. It sits at the top of the Artificial Analysis Intelligence Index v5.0 with a score of 59. Google also reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching, plus 1768 on WebDev Arena, which ranks fourth.

Access is narrow for now. The rollout starts with a subset of cyber defenders and widens to trusted partners through the Fairwind Program. Google says it will reach developers, businesses and consumers as soon as possible but gives no date. Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later, with cached input at 95% off.

The page does not state a context window. Several claims, such as the Chrome team's patch quality comparison and the doubling of secure code success against Argon, come from Google's internal testing.

Nemotron 3.5 Transcribe and Granary v2

NVIDIA released an open-weight speech model that handles transcription, speaker attribution, translation and audio Q&A in one checkpoint. It transcribes 40 languages, translates 31 to or from English, labels up to eight speakers and takes recordings up to an hour. It is built for recorded audio, not live streaming.

NVIDIA reports 5.33% average word error rate on the English Open ASR Leaderboard, first among open-weight models. Fine-tuning on a medical benchmark from Eka Care cut held-out error from 22.44% to 12.17%. The post does not name a license, which I would check before using it commercially. Granary v2, the companion dataset, now covers 48 languages and 463,062 hours across 16 corpora.

olmOCR 2 on historical archives

Ai2 profiled two projects built on olmOCR 2. A historian built a tool that runs vision-language models locally to make thousands of archival photos searchable without uploading them anywhere. A TU Wien team fine-tuned the model on 57 handwritten pottery records from the Roman site Carnuntum, and error on 450 test records fell from 15.58% to 1.11%. The historian also notes the model can write fluent text that is wrong, which is the failure to watch in archives.

Anthropic AI for Science

Anthropic expanded its AI for Science program with 20 highlighted partners: funders such as the Simons Foundation and the US National Science Foundation, institutions such as the Wellcome Sanger Institute and EMBL, and initiatives such as FlyWire and gnomAD. It also launched a discounted Team plan for scientists. Anthropic says many collaborations are early and it will share results later.

What I'd watch next

Whether Carbon reaches general availability and what outside testers find on the security numbers. Whether NVIDIA publishes the quantized Transcribe builds and a license. And whether the pottery result holds when the team runs the roughly 70,000 records in the full archive.