AI NewsOctober 12, 2026

OrcaCyber Zero 1.5 is built for authorized security work such as reproducing vulnerabilities and testing defenses. It reads up to 1 million tokens at once, and access is limited to vetted researchers.
Security teams drown in reports of possible bugs. OrcaCyber Zero 1.5 aims to prove which are real, like a mechanic test-driving the car instead of guessing from the noise. It reads up to 1 million tokens (a very large codebase) at once.
OrcaRouter reports a perfect score on 39 Cybench hacking challenges, priced at $3 in and $7.50 out per million tokens. Access is limited to approved security researchers and red teams.
Key Capabilities:

• Proves, not just flags: Aims to confirm real flaws, such as running a suspected injection in a real browser.
• Big memory: Takes 1 million tokens of input and returns up to 128K tokens of output.
• Restricted access: Available only to the Security Research tier through an OpenAI-compatible API.

Falcon-ASR is a 1.6B speech-to-text model for Arabic and Emirati speech that also handles English, French, Spanish and Portuguese. It runs only as a hosted demo for now.
Most speech tools stumble on regional dialects because little transcribed data exists. TII trained Falcon-ASR on Emirati and Gulf Arabic, like hiring a local who knows the accent. One set of weights covers all five languages, with no language setting needed.
In TII’s own tests it averages a 20.92% word error rate on six Arabic test sets, beating the best published 23.17%. Lower is better. It also returns word-level timestamps for subtitles.
Key Capabilities:

• Dialect focus: Trained on Emirati, Gulf and standard Arabic, plus English, French, Spanish and Portuguese.
• Small but strong: At 1.6B parameters, it beat larger Arabic models, including a 7B one, in TII’s comparison.
• Demo only: No weights or license yet, with an API and native apps planned

Sakana AI built an AI reviewer and tested whether it can find mistakes deliberately planted in research papers, not how closely it copies human reviews. It runs on off-the-shelf Claude models for about $0.47 per review.
Most AI paper reviewers are judged on how well they imitate humans. Sakana AI tested something harder: planting 1,164 contradictions in 257 papers and checking whether the AI finds them, like hiding typos for a proofreader.
Its Multi-Layered Review system reads the paper and appendix before judging. With four reviews it caught 73.43% of errors in a paper’s main claims, versus 14.81% for the best baseline, though it still fell for hidden prompt injection.
Key Capabilities:

• Planted-error test: 1,164 inserted contradictions across 257 papers from five venues.
• Cheap to run: Uses off-the-shelf Claude Sonnet 4 and Haiku 3.5, about $0.47 per review.
• Real-world gap: On real retracted arXiv papers, only 16.11% were exact matches.

Clef-omni is a decision model, meaning it returns probabilities for fixed options instead of writing text. It is a 30B model that runs only 3B per request, free to download under Apache-2.0.
Most AI pipelines transcribe audio and caption video before deciding anything. Cloudflare’s Clef-omni skips those steps, taking text, images, audio and video in one call and returning a probability for each allowed option, like a judge watching the raw footage.
It’s open-weight under Apache-2.0 and live on Workers AI at $0.15 per million input tokens (chunks of text). Cloudflare also cut Clef-flash pricing and made hosted Clef about 2x faster.
Key Capabilities:

• One call, four inputs: Scores text, images, audio and short video together in a single request.
• Small footprint: Only 3B of its 30B parameters run per request, built on Qwen3-Omni.
• Strong on intent: Leads on sorting customer intent but trails the text-only Clef on harder reasoning tests.