Google
Nov 11, 2026
Gemini Decisions API adds bounded multimodal scoring
Google introduced Gemini Decisions API with Gemini Decision 1 for closed-set choices and rubric scores across text, images, audio, video and PDFs. The announcement supports up to 1,000 questions against a multimodal input of up to 256,000 tokens. It targets routing, content grading and policy checks. Calibration, speed and cost benefits are Google’s claims.
- Google announced a dedicated multimodal decision endpoint and its supported input and question limits.
Why it mattersBounded decisions can change the economics of routing and evaluation pipelines. Financial applications still require measured error rates, threshold selection and explicit escalation behavior on representative data.
Google DeepMind Research
Nov 11, 2026
AlphaGenome 2 adds cell-conditioned DNA-effect prediction
Google DeepMind announced AlphaGenome 2, which conditions DNA-sequence predictions on cell type or state and can incorporate new single-cell measurements. The publisher reports training on over 100 million cell profiles across 68 tissues and improved variant-effect prediction. Access is available for non-commercial research, with commercial research in private preview. The system’s stated use is research, and its capabilities remain publisher-reported.
- The official November 11 announcement introduces cell-conditioned variant prediction and research access.
Why it mattersCell-specific prediction may improve how research teams prioritize experiments. Independent replication, experimental validation and access terms determine the practical value for scientific workflows.
GOV.UK / Fractile
Nov 11, 2026
UK public investors join Fractile’s round of over £500 million
The UK government announced that Sovereign AI and the British Business Bank are backing AI-hardware developer Fractile in a round exceeding £500 million. The statement describes support for domestic chip-design capability and the company’s full-stack approach. The confirmed financing changes the resources available for development, while the announcement’s projected hardware and economic benefits remain prospective.
- The government’s November 11 release confirms public-investor participation and a round above £500 million.
Why it mattersPublic capital is participating in a major AI-hardware financing. Delivery milestones, measured efficiency and manufacturing execution will determine whether the investment produces competitive systems.
Microsoft Research
Nov 10, 2026
Microsoft studies dormant-neuron backdoors in tool-using agents
Microsoft Research describes Lazarus, a study in which small weight modifications create hidden triggers that redirect an agent’s actions while ordinary tests can still pass. The attack assumes write access to a model checkpoint. Microsoft describes provenance controls already used in Foundry and additional scanning and signing work under development. The report establishes a demonstrated research threat model with a specific access prerequisite.
- Microsoft describes a weight-tampering technique and distinguishes deployed provenance checks from defenses under development.
Why it mattersFor financial and enterprise agents with tool access, checkpoint integrity and permission boundaries become part of operational evaluation. The study supports targeted controls and testing under the stated threat model.
Google DeepMind Research
Nov 10, 2026
Google expands access to Gemini 4 Carbon
Google made Gemini 4 Carbon available to developers, enterprises, scientists and consumers following a limited cybersecurity-defender release. The company positions Carbon for extended coding, knowledge-work and scientific workflows that spend more tokens on exploration and verification. Its performance and cost comparisons remain provider-reported. The announcement supplies a new access milestone for an existing model family.
- The canonical announcement expands Gemini 4 Carbon availability on November 10.
Why it mattersBroader availability creates an opportunity to compare long-horizon task quality, latency and total execution cost on representative work. Domain-specific verification remains part of assessing its operational value.
Riskfuel / BBVA
Nov 10, 2026
BBVA uses a neural pricing surrogate in ePricer structuring
Riskfuel announced that BBVA uses its neural surrogate for autocallable pre-trade structuring in ePricer. Trained on BBVA’s own pricing model, it accepts more than 4,000 inputs. The issuer reports validation on over 100,000 trades and about 2.27 million valuations per second in Riskfuel–Intel benchmarks. A quoted BBVA executive confirms the integration. Accuracy and throughput measurements remain issuer-reported.
- Riskfuel’s dated release reports BBVA deployment and supplies a named bank executive’s account.
Why it mattersThe deployment connects model approximation to a live financial workflow. Assess approximation error, model-governance controls and behavior outside the validation domain alongside reported speed.