Google AI Updates: November 4, 2026

1. Gemini API Deprecates Gemini 2.5 Pro and 2.5 Flash, Auto-Routing Both to Newer Flash Models on December 1

Google deprecated gemini-2.5-pro and gemini-2.5-flash in the Gemini API on November 2, and neither can be selected in Google AI Studio any longer. From December 1, 2026, every API request to gemini-2.5-pro will be routed automatically to gemini-3.8-flash and billed at Gemini 3.8 Flash rates, and requests to gemini-2.5-flash will be routed to gemini-3.5-flash-lite and billed at Flash-Lite rates. Google lists breaking API changes for both migrations: changes to function calling and thought_signature validation, thinking_level replacing thinking_budget, and the removal of sampling parameters and prefilled model turns, so applications that still set a thinking budget, pass sampling parameters or prefill an assistant turn need code changes, not just a new model string, before the cutover. Source

2. Google Releases Physical-Engineering’s Last Exam, an 85-Task Agent Benchmark Where No Model Reaches 50%

Google’s AI & Infrastructure team, working with Mercor, announced Physical-Engineering’s Last Exam (PLE), a benchmark and open competition for agents doing mechanical (30 tasks), electrical (27), civil (18) and structural (10) engineering work. Each of the 85 long-horizon tasks runs in a containerized project folder built on Mercor’s Archipelago environment with an average of 61 files (5,170 in total) such as PDFs, spreadsheets, emails, 2D CAD drawings, 3D BIM models and document revisions, and the agent must find the governing codes and physical limits, run calculations and produce deliverables a licensed reviewer would sign off; 44 tasks were written by practicing engineers and 41 were assembled by an agent pipeline from engineer-designed specs, then validated and made harder by engineers. Points are awarded only if a design passes every hard safety check: the top configuration met 91% of individual constraints on co-authored tasks, but only 41.3% of its deliverables passed every hard safety and feasibility check, and model choice mattered far more than harness (up to a 50-point spread across models against a median 3-point spread across harnesses for the same model). On the leaderboard at ple-bench.com, which scores each model on its native harness at maximum reasoning over the 56 hidden tasks, Claude Opus 5.5 in Claude Code leads at 0.483, followed by GPT-6 Astra in Codex (0.458), Claude Sonnet 5.5 (0.435) and Google’s Gemini 4 Argon in Antigravity (0.406), with Grok 4.7 at 0.301; Gemini 4 Carbon is not listed. Google released 29 public tasks, grading scripts and the evaluation container on GitHub, a paper on arXiv, and a submission process for the private set, with 13 Google-sponsored models running free on its servers and prize details still to come. Source

3. Google Makes Gemini and Search Opt-In for K-12 Student Accounts, and Schools That Refuse AI Lose Search Entirely

Google announced at its AI in the Learning Era event on November 3 that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be entirely off for K-12 student accounts until school administrators explicitly turn them on, with specifics to be shared with schools “in the coming months” before the change rolls out (TechCrunch reports it is planned for next year). Because AI features are built into Search, education VP Chris Phillips told The New York Times that schools restricting AI altogether “will not have access to search,” and students will see a message from their school instead. Alongside the controls, Guided Learning, the Gemini study mode that coaches with step-by-step prompts instead of final answers, is rolling out globally in Google Classroom as teacher-assigned activities anchored on uploaded course material, with insights showing teachers how students engaged, and is coming to Search for student accounts through a pilot in select U.S. schools later this month. Google also published a randomized controlled trial with more than 300 University of Tokyo students in which half a day of statistics study with Guided Learning raised test scores by 8.1 percentage points against 5.3 points with standard web resources alone, plus a four-week pilot with Ark across three UK secondary schools, and announced Google.org funding for five organizations building AI tools for teachers. Source Research TechCrunch

4. Google DeepMind Adds Cheminformatics Skills to Antigravity’s Science Plugin

Google DeepMind released Cheminformatics Skills for Google Antigravity, available now by installing the Science plugin in the Antigravity app, and for enterprise users who sign in with Google Cloud credentials. The skills split the work so the language model handles orchestration, database navigation and scientific reasoning while deterministic tools grounded in the open-source RDKit toolkit run the molecular calculations, with connectors to ChEMBL and PubChem. In a walkthrough built on the lung cancer drugs crizotinib and its cyclized successor lorlatinib, the agent pulled thousands of compounds from the literature, filtered out buffers and solvents, ran matched molecular pair analysis, trained and validated potency models (using chronological splits where dates exist, otherwise scaffold splits) across four experimental readouts, and proposed new rigidified macrocycles with synthesis routes and electron-pushing diagrams in about an hour, a workflow DeepMind says typically takes multiple days; results export to standard chemistry software formats and can be prepared for AlphaFold or molecular dynamics. DeepMind cautions that sparse data, inter-lab assay noise and narrow dynamic ranges limit predictive precision and that every proposed design still has to be synthesized and tested. Source