Google made Gemini Omni 1.2 Flash generally available in the Gemini API and business tools, adding 15-second video, conversational editing and 720p pricing at $0.07 per second.
Thinking Machines launched Inkling 2 Small with 594 billion total parameters and 17 billion active per token; Tinker access is live, but downloadable weights await safety checks.
Axiom released a Lean formalization of the roughly 7,500-page Stacks Project, making major algebraic-geometry definitions and proofs machine-checkable and reusable by research agents.
OpenAI and Ligo introduced BioBasicsBench for messy laboratory evidence; GPT-6 Astra scored 66.7% mean reward at xhigh effort versus 43.8% for GPT-5.6 Sol.
LeRobot v0.7.0 brings text-steered robot sessions, direct streaming from large robotics datasets and GPU-hosted remote policies, alongside breaking changes for existing installations.
AI NEWS ANALYSIS
The AI frontier is shifting from impressive outputs to dependable tools and verifiable evidence. Google and Thinking Machines expand what creators and developers can do with multimodal models, though availability and independent tests still matter. Axiom's Lean library makes a major body of mathematics machine-checkable, while OpenAI's BioBasicsBench exposes the remaining uncertainty in interpreting messy laboratory data. Hugging Face's LeRobot release connects language-based planning with real robots, raising the importance of latency, safety and reproducible testing. Across the five stories, operational reliability is the common challenge.
Read previous source-backed editorials and dated briefings.
Send an AI model release, research finding or industry announcement for independent editorial review.
Pitch a Story →IN-DEPTH AI NEWS / NOVEMBER 7, 2026
Original reporting on Gemini Omni 1.2 Flash, Inkling 2 Small, the Stacks Project in Lean, BioBasicsBench and LeRobot 0.7.
Google made Gemini Omni 1.2 Flash generally available in the Gemini API and business tools, adding 15-second video, conversational editing and 720p pricing at $0.07 per second.
Google launched Gemini Omni 1.2 Flash on November 5, expanding its generative-video model to 15-second clips with native editing across multiple conversational turns. The release is generally available through the Gemini API and Google business tools, rather than being a limited research demonstration. Google says it lowered the 720p generation price by 30% from Omni 1.1, to $0.07 per second. For developers building video workflows, the real shift is the ability to revise a shot without restarting the entire generation process.
The longer native clip reduces the need to stitch together separate short generations. Multi-turn editing potentially changes the cost of experimentation: a designer can ask to replace a background, preserve a character or alter the composition without rewriting all instructions. Reference inputs may help maintain brand elements and scene continuity across shots. These capabilities matter for product explainers, catalog clips and accessible educational materials, especially when a team needs several regional versions. The model still requires editorial review for accuracy, rights and misleading depictions.
Set up a controlled comparison using identical prompts and reference assets across Omni versions. Measure how often text, faces, products and actions remain consistent, plus end-to-end turnaround time after revisions. Include failed generations and upscaling fees when estimating the total cost of one accepted clip. Test audio synchronization and accessibility captions separately. For an API product, verify the exact model identifier, supported resolutions, usage limits and regional account access before presenting the service as available to every customer.
Thinking Machines launched Inkling 2 Small with 594 billion total parameters and 17 billion active per token; Tinker access is live, but downloadable weights await safety checks.
Thinking Machines Lab introduced Inkling 2 Small on November 5 as a multimodal mixture-of-experts model designed for coding, tool use, visual reasoning and research automation. Its architecture has 594 billion parameters in total, with 17 billion active for a token. That scale does not make it a small model in memory or operational terms; the product name describes its position in the company's lineup. Developers can currently sample and fine-tune the model through Tinker, while the announced Apache 2.0 weight release remains a future step.
Mixture-of-experts routing activates a subset of model parameters for each token, seeking higher capability without paying the full computation cost of a dense model of equal total size. Thinking Machines describes a sparse architecture with specialized experts and hybrid attention. Native image support lets applications process screenshots or visual documents alongside text. The actual latency, memory needs and price still depend on serving hardware, the long-context workload and provider configuration, so active-parameter counts alone cannot predict application cost.
Build tasks that resemble the intended deployment, such as coding fixes, tool selection, document interpretation and error recovery. Capture tool traces, latency, cost, safety-policy violations and outcomes that require human correction. Test how the model handles large tool catalogs, because the company acknowledges that it does not always choose the best tool. Treat the reported autoresearch exercise as a case study with a limited internal benchmark rather than evidence of reliable autonomous scientific work across every field.
Axiom released a Lean formalization of the roughly 7,500-page Stacks Project, making major algebraic-geometry definitions and proofs machine-checkable and reusable by research agents.
Axiom announced on November 4 that it had formalized the Stacks Project, a long-running open mathematical reference for modern algebraic geometry, in Lean. The company describes a unified library of roughly 16 million lines built from a reference of about 7,500 pages. Unlike a language model answering a theorem question, a formal proof must pass a proof assistant's logical checks. This is a significant infrastructure milestone for AI-supported mathematics because researchers and agents can reuse verified definitions and theorems instead of relying entirely on plausible prose.
Algebraic geometry contains layers of definitions, lemmas and theorems that other proofs depend upon. When those dependencies live in a proof assistant, an automated agent can search for reusable statements and submit derivations for checking against a formal kernel. This reduces the risk of silently assuming a false intermediate step. The scale of the Stacks library also makes it a useful environment for testing retrieval, dependency navigation and long-context planning rather than only competition-style mathematical puzzles.
Researchers can import selected declarations, inspect the stated assumptions and compare the machine-readable form with the cited Stacks tag. Reproducible builds, version pinning and clear documentation are central: a library of this size is only useful when other teams can compile and extend it. AI researchers should evaluate whether their agents can find the correct lemma, construct a valid proof and report failure honestly. They should also distinguish Lean checking from the independent mathematical judgment needed to formulate an interesting new theorem.
OpenAI and Ligo introduced BioBasicsBench for messy laboratory evidence; GPT-6 Astra scored 66.7% mean reward at xhigh effort versus 43.8% for GPT-5.6 Sol.
OpenAI introduced BioBasicsBench on November 6 in collaboration with biotech company Ligo to evaluate agents on routine wet-lab evidence. Its 100 tasks include imperfect gel images, colony plates, microscopy and instrument outputs, not only curated figures from published studies. The benchmark is notable because laboratory decisions often depend on recognizing failed controls, ambiguous patterns and the absence of enough evidence. OpenAI reports an improvement for GPT-6 Astra over earlier models, while the remaining error rate makes human scientific oversight essential.
A wet-lab agent may need to decide whether a gel supports the presence of a protein, whether a plate has usable colonies or whether the experimental controls invalidate an apparent result. This is different from recalling textbook facts. The benchmark rewards interpretations grounded in raw observations and appropriate uncertainty. Because laboratory errors can cascade into unnecessary experiments, correctly saying that data are inconclusive can be more valuable than producing an impressive but unsupported answer.
Developers should reproduce the same kind of evidence checks on their own assay formats, equipment and protocols, documenting where the models fail. Evaluate whether the agent requests missing controls, detects mislabeled samples and distinguishes measurements from guesses. Track the cost of higher reasoning settings and the amount of scientist review required before accepting a conclusion. Where a model is integrated with laboratory software, avoid granting authority to discard data, order materials or change experiments without clear human approval.
LeRobot v0.7.0 brings text-steered robot sessions, direct streaming from large robotics datasets and GPU-hosted remote policies, alongside breaking changes for existing installations.
Hugging Face released LeRobot v0.7.0 on November 5, reshaping its open robotics stack from dataset collection through training and live control. The release lets operators give new text instructions while a robot policy is running, use remote GPU hardware for inference and train directly against hosted datasets without copying all the source videos locally. The Hub now hosts almost 300 terabytes of open physical-AI data in LeRobot format. That scale matters for teams whose bottleneck is data movement and policy deployment, not just the sophistication of the robot model.
Running a robot requires a loop between observations, policy decisions and safe actuator commands. Previously, many setups loaded a model and followed a preselected instruction until completion. Interactive rollouts allow operators to ask what a robot sees, redirect a task or delegate subtask selection without restarting the whole session. Remote policy execution decouples expensive GPU inference from the low-power computer mounted on a machine, while buffered action chunks help smooth network delays. Both capabilities increase flexibility but introduce additional timing and safety risks.
Create a controlled, low-force evaluation environment with a reliable emergency stop. Verify that new language commands respect workspace limits and that remote network interruptions do not cause unsafe motion. Measure decision latency, packet loss, data freshness and success over repeated episodes. Test streaming throughput for the actual video resolutions and storage system. Review the migration notes carefully: the v0.7.0 stack changes dependencies and replaces its earlier asynchronous inference module. A version-pinned simulator is preferable to debugging installation and motion behavior simultaneously.