Google released Gemini Omni 1.2 Flash as generally available on the Gemini API and Gemini for business. It generates native video clips of up to 15 seconds, up from 10 seconds in Omni 1.1. Duration can be set from 3 to 15 seconds in one-second steps. Price is $0.07 per second for 720p, a 30% cut from Omni 1.1, and outputs carry a SynthID watermark.
How we checked
Release, GA status, duration limits and pricing read from Google's own announcement page (two mirrors). Quality and consistency improvements are vendor claims and are not asserted here.
Why it matters
Longer clips, lower per-second pricing and GA status make API-driven video generation more practical for production ad and explainer workflows.
Try this
On the Gemini API, generate a short Gemini Omni 1.2 Flash clip up to 15 seconds and check the listed price against the announcement.
NVIDIA released four open-weight Cosmos 3 checkpoints: Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer and Cosmos3-Nano-Sim-Bimanual. Each is post-trained for interactive world simulation under the OpenMDW-1.1 license, per Hugging Face.
How we checked · Why it matters
How we checked
Read on the NVIDIA-authored Hugging Face release post. Model cards not separately fetched.
Why it matters
Open checkpoints for interactive world simulation give builders a starting point to test robot and auto simulation workflows.
Anthropic analyzed 207,049 consumer-plan Claude and Claude Code research sessions from Aug 2 to Sep 30, 2026, using privacy-preserving classifiers. Data analysis made up 33.7% of Claude sessions and 48.1% of Claude Code sessions, versus 27.6% of estimated researcher time. Computer science accounted for 32.7% of Claude Code sessions, while its researchers are 4.3% of the global total. Ideation was only 4.2% of Claude and 2.1% of Claude Code sessions.
How we checked · Why it matters
How we checked
Figures read from Anthropic's primary report page. The underlying usage data is Anthropic-internal and cannot be independently checked.
Why it matters
Data analysis dominates what researchers hand to Claude, while ideation is rare in these sessions (4.2% Claude / 2.1% Claude Code).
In the same study, among biology and biomedical sessions, Claude Science was used for more specialized tasks than general Claude. Ideation and hypothesis generation made up 8% of Claude Science sessions. Proposing mechanistic protein–protein or protein–ligand interaction models was 5.6x more likely there than on Claude. Over 70% of Claude Science biology/biomedical sessions ran with high AI autonomy, compared with 23.7% on Claude.
How we checked · Why it matters
How we checked
Figures read from Anthropic's primary report page. The usage data is vendor-internal.
Why it matters
Claude Science biology/biomedical sessions show more autonomous work than general Claude (70%+ vs 23.7%); user populations differ.
Hugging Face's Open ASR Leaderboard added two held-out private test splits from Liva AI. Egyptian-accented English has 1.34 hours and Nigerian-accented English has 1.55 hours, both conversational with punctuated, cased transcripts that include disfluencies. The leaderboard also now reports US vs non-US accents and Kachru Inner Circle vs Outer and Expanding Circle accents.
How we checked · Why it matters · Try this
How we checked
Read on the Hugging Face blog, which is primary for its own leaderboard update.
Why it matters
Private, held-out accent data reduces benchmark overfitting and exposes ASR gaps on underrepresented English accents.
Try this
Check the Open ASR Leaderboard's Private tab to compare your candidate ASR models on Inner Circle vs Outer and Expanding Circle accents before choosing one.
OpenAI Codex CLI 0.167.0 was released on GitHub with these changes: an agent command center opened via `codex agents --no-daemon`, with unread-result markers; cross-backend session switching in the terminal picker that keeps drafts; a gcp_setup_wizard for Google Cloud Vertex AI sign-in; and opt-in system proxy support under /experimental. Ctrl+L now clears the display while keeping history available via Ctrl+T.
How we checked · Why it matters · Try this
How we checked
Release notes read on the official openai/codex GitHub release page. Not tested; no sandbox.
Why it matters
Codex users get multi-agent session management and a supported Vertex AI onboarding path in the stable CLI.
Try this
After upgrading to Codex CLI 0.167.0, run `codex agents --no-daemon` to open the agent command center and review unread agent results.
GitHub added a copilot_code_review object to every row of the enterprise and organization repos-1-day Copilot usage metrics reports. Fields include total_comments_evaluated and total_comments_addressed, which together give an addressed rate. Both counts are broken down by severity, review tier, comment type and customization. A total_merged_no_human_change_request_after_review count is also included. Comments are counted on the day the pull request merges.
How we checked · Why it matters · Try this
How we checked
Read on GitHub's official changelog. API not called.
Why it matters
Teams can now measure per repository whether Copilot code review feedback is acted on, and whether custom instructions, skills or MCP servers change that.
Try this
With View Copilot Metrics permission, pull a repos-1-day report and compute total_comments_addressed / total_comments_evaluated per repository, split by comment type.
The Copilot usage metrics API now includes per-user fields in enterprise and organization user reports: client_telemetry_present, and usage_metrics_requirements, which marks each IDE and extension as met, unmet or unknown against minimum versions. Daily counts missing_client_telemetry_user_count and below_usage_metrics_minimum_user_count were also added. Null values mean the data is unavailable, and days before the release have no values.
How we checked · Why it matters · Try this
How we checked
Read on GitHub's official changelog. API not called.
Why it matters
Admins can find users whose Copilot metrics lack detail because of outdated editors, disabled telemetry or proxies.
Try this
In a 1-day organization user report, list users where client_telemetry_present is false or any usage_metrics_requirements entry is unmet, then follow up on their editor versions.