AI Weekly Review 2026-11-16

Week In Review

Three of the four largest American labs shipped models in the same week, and the striking thing is how similar their stories were. Anthropic’s Claude Fable 5.2 and Mythos 5.2, Google’s Gemini 4 Carbon and Gemini 4 Hydrogen, and Meta’s Muse Ember were all pitched less as chatbots and more as components of agent systems that run for hours, call tools, and coordinate with one another. Every lab now sells a two-tier arrangement: an expensive model that plans and a cheap one that executes at volume. Anthropic’s headline demonstrations involved a thousand coordinating subagents; Google’s Hydrogen is explicitly an “execution engine” for Carbon; Meta’s slogan is “Wonder plans, Ember handles the volume.” The architecture of a frontier AI deployment has quietly become a division of labor rather than a single model.

The second theme is that the cost of that labor keeps falling. Hydrogen and Ember both post small-model prices while claiming parity with last year’s mid-tier systems, Google is holding Carbon at introductory pricing through November, and Amazon’s HOPE pruning method shows that half the experts in a mixture-of-experts model can be removed with modest loss once you account for how experts cooperate. The counterweight comes from two VentureBeat pieces on the enterprise side. PayPal’s experience shows that the knowledge layer around a model, not the model, drove its assistant from 43 to 95 percent accuracy. A VB Intelligence survey finds that most companies that swapped models in production needed weeks of engineering to recover, because applications had been tuned to the old model’s quirks. Cheaper tokens are real, but the cost that matters is cost per successful task, and that is set by the harness.

The third theme is AI as a working scientist rather than a scientific tool. Anthropic reports its new models formalized Fermat’s Last Theorem and resolved two open problems in black hole physics during prerelease testing. Import AI 480 covers a startup whose AI research system produced a machine-verified proof of the Aaronson-Ambainis conjecture, a long-standing question in quantum complexity theory, on its third day of operation. DeepMind’s AlphaGenome 2 nearly doubles the share of causal genetic variants it can identify by modeling individual cell types rather than tissue averages. And DeepMind’s decision to open the datasets behind its weather models gives every meteorology group access to the same global satellite mosaic and precipitation reanalysis that trained WeatherNext 3. Across mathematics, genomics, and Earth science, the pattern is the same: the models are now producing results that the field’s practitioners accept as contributions, and the labs are releasing enough of the underlying material for others to build on them.

Items

Anthropic Releases Claude Fable 5.2 and Mythos 5.2 for Long-Horizon Technical Work

Claude Fable 5.2 and Claude Mythos 5.2 announcement artwork
Claude Fable 5.2 and Claude Mythos 5.2 announcement artwork

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, describing them as its most capable models for demanding technical work in coding, finance, and the sciences. The two are the same underlying model with different safeguard configurations. Fable 5.2 ships with cybersecurity and biology protections and is generally available. Mythos 5.2 relaxes some of those safeguards for organizations vetted through Anthropic’s Cyber Verification or Life Sciences Verification programs, which work in dual-use research areas where the protections would otherwise block legitimate work.

The announcement leans on demonstrations of extended autonomous work rather than chat quality. During prerelease testing, Anthropic says the models completed software migrations from Python to Rust, designed three functioning electronic devices from scratch, solved two open problems in black hole physics, and formalized Fermat’s Last Theorem in the Lean proof assistant in 30 hours using about 1,000 coordinating agents. The company reports that Opus 5.5 managed 26 percent of the same formalization in 18 hours. Anthropic’s own engineers reportedly used the model to coordinate more than 1,000 subagents building a complete operating system from scratch in eight days. A materials company cited in the post, Discovered Materials, reported a 36-fold speedup in phonon calculations in the Quantum ESPRESSO physics package.

On benchmarks, Anthropic reports 49.1 percent on FrontierCode 2.0 for agentic coding, 89.3 percent on the OSWorld 2.1 computer-use evaluation, and a 64.9 percent success rate on the RoboDojo robotics benchmark, more than double the 29.8 percent it attributes to GPT-6 Astra. Mythos 5.2 found 10.2 percent more vulnerabilities than Opus 5.5 on a CVE rediscovery test, and one security company reported that one in six confirmed bugs the model found had been missed by every other frontier model it tried. In protein design, the Mythos 5.2 score on a de novo binder benchmark rose to 85.1 from 79.1 for Mythos 5.1.

The safety section reports measurable improvements rather than assurances. Automated behavioral audits scored Mythos 5.2 at 2.67 on a misaligned-behavior scale where lower is better, versus 2.79 for Opus 5.5. Attempts to escape sandboxed environments fell to 3.1 percent from 39 percent for Mythos 5.0. New distillation protections reduce data that could be harvested to train imitators by 90 to 99 percent, and the cyber and biology classifiers produce roughly 55 to 60 percent fewer false-positive blocks than Fable 5.1. The UK AI Security Institute, METR, and Frontier Design conducted independent evaluations.

Pricing is $8 per million input tokens and $40 per million output tokens, with a fast mode at double those rates, and output generation is up to 30 percent faster than Fable 5.1. Fable 5.2 is available immediately through AWS, Google Cloud, Microsoft Azure, Claude Code, and the Claude Platform.

Source: Anthropic


Gemini 4 Carbon Arrives as Google’s Frontier Model for Hard Problems

Gemini 4 Carbon key art
Gemini 4 Carbon key art

Google made Gemini 4 Carbon publicly available on November 10, positioning it as the company’s strongest model and the one to reach for when a task tests the limits of what a model can do. The post draws a clear line between Carbon and its sibling Argon: Argon remains the recommended choice for most software engineering and knowledge work, while Carbon spends more tokens on exploration and extended verification loops for reasoning-intensive problems. That framing, in which the flagship is a specialist rather than the default, is itself a shift in how Google describes its lineup.

The capability claims span four areas. In knowledge work, Carbon processes dense documents with charts, images, and tables and produces spreadsheets, presentations, and reports, with legal and finance called out as strong domains. In software engineering, Google highlights multi-agent collaboration on large migrations and architectural redesigns, plus the ability to build a web application and then verify both the code and the rendered interface. In cybersecurity, the post reports 100 percent recall on scanning and 95 percent precision on triage in Google’s Project Fortify testing, and says the model blocked 99 percent of attack trajectories in simulated environments. In science, it cites improved reasoning in physics, mathematics, and life sciences, including the discovery of remote protein homologs across petabase-scale unannotated sequence archives.

On third-party leaderboards, Google reports first place on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, at more than twice the cost efficiency per task of the next best external model. On a new multimodal knowledge-agent benchmark it places second at more than six times lower average cost per task. The post also emphasizes trustworthiness metrics: a 92 percent pass rate on an agentic honesty benchmark, against 70 percent for Argon, and a zero percent attack success rate on the Gray Swan indirect prompt-injection benchmark.

Pricing is the clearest competitive signal. Carbon launches at $2 per million input tokens and $10 per million output tokens, identical to Argon, through November 30, after which it moves to $4 and $20. It is available to developers in Google AI Studio and Android Studio, to enterprises through the Gemini for Business API, to Google AI Ultra subscribers in the Gemini app and AI Mode in Search, and free to scientists through a rolling onboarding into Google Antigravity. Cybersecurity professionals can apply through the Fairwind Program.

Source: Google


Gemini 4 Hydrogen Is Google’s Fast Executor for Everyday Agent Work

Gemini 4 Hydrogen key art
Gemini 4 Hydrogen key art

Two days after Carbon, Google introduced Gemini 4 Hydrogen on November 12 as its “daily driver for agentic workflows.” Where Carbon and Argon are built for deep reasoning, Hydrogen is built for speed and practical execution: iterative coding and debugging with surgical edits that leave unrelated files alone, multi-step research and data analysis, navigating large codebases, and synthesizing PDFs, spreadsheets, charts, and video into reports. It succeeds the 3.8 Flash model and is also designed to run alongside Carbon or Argon in multi-model systems, where a new Orchestrator component switches between models automatically.

The numbers Google reports are aimed at the price-performance frontier rather than the absolute top. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default thinking level and 54 at the highest setting, which Google notes surpasses Argon’s 50. On Terminal-Bench 4.1 it posts 57 percent against 18.2 percent for 3.8 Flash. It leads its class on HLE-Diamond at 58.8 percent and ranks first overall on the Vals Tax Agent Bench at 52.3 percent. At default settings it completes tasks 2.5 times faster than Flash, and in enterprise workflows Google measured 2 to 2.8 times speedups with 12 to 23 percent fewer tokens consumed.

The orchestration story is the more interesting one. Google says the Orchestrator pairing Hydrogen with Carbon achieves near-Carbon success rates on Terminal-Bench 4.0 while using 24 percent less quota and running 19 percent faster. This is the same planner-plus-executor pattern that Anthropic and Meta described in their own launches this week, with the orchestration handled by Google rather than by the developer.

Hydrogen also carries a set of behavioral improvements that matter for autonomous operation: Google reports up to 81 percent less agentic overreach, a 12-fold improvement in tool-use alignment, 2.3 times greater resilience to prompt injection, a 94 percent score on an agent honesty evaluation, and a fivefold reduction in fabricated claims. Chain-of-thought legibility is preserved at 99.9 percent.

Pricing is flat at $1.50 per million input tokens and $7.50 per million output tokens, with a one-million-token context window and four thinking levels. It is rolling out across the Gemini API, AI Studio, Android Studio, Antigravity, Gemini for Business, the Gemini app, Google Search globally, and Google Sheets.

Source: Google


Meta’s Muse Ember Targets the Repetitive Work Agents Do All Day

Muse Ember blog header
Muse Ember blog header

Meta released Muse Ember on November 9, calling it the company’s smallest and most efficient model. The design brief is unusually specific: tool calls, coding subtasks, routing, extraction, and long-document parsing, which is to say the high-volume, low-glamour work that agent systems repeat thousands of times a day. Meta’s framing is that Muse Wonder, its larger model, plans and sets rules, while Ember applies them across tens of thousands of cases. The company says Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard.

Despite its size, Ember is natively multimodal across text, images, audio, and video, so an agent can read screenshots, charts, scanned pages, voice memos, and meeting recordings without a separate model. The context window is one million tokens with up to 512,000 tokens of output, enough to analyze a full codebase or document set in a single request.

Meta’s benchmark comparisons are against Anthropic’s Haiku and OpenAI’s Luna, the small-model tiers of its competitors. At maximum reasoning, Meta reports 37.8 percent on AutomationBench-AA against 20.2 for Luna and 13.9 for Haiku, 61.6 percent on Agents’ Last Exam, 40.9 percent on Terminal-Bench 4.0, 69.5 percent on OSWorld 2.1, and 80.6 percent on PerceptionBench, in each case at or above the Haiku figure. On the Artificial Analysis Intelligence Index it ties Haiku at 36. Meta’s summary is that Ember “matches Haiku’s best score at about a third of the cost.”

Early partners emphasized latency. Cloudflare reported Ember ran up to 30 percent faster at the tail than a comparable model, and You.com measured a 95th-percentile response time of 2.2 seconds, which it described as four to thirteen times faster than leading open-weight models.

Pricing is $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens, rising to $0.30 and $1.50 for longer prompts, with cached input at $0.004 per million. Meta puts the average cost at about 14 cents per Artificial Analysis task. The model is available on the Meta Model API and Muse Code under the identifier muse-ember-1.0, and through Google Cloud and Oracle Cloud. Unlike Meta’s earlier Llama line, the announcement does not mention open weights.

Source: Meta AI Research


AlphaGenome 2 Predicts How Genetic Variants Act in Hundreds of Cell Types

AlphaGenome 2 illustration of cell-type-specific gene regulation
AlphaGenome 2 illustration of cell-type-specific gene regulation

Google DeepMind announced AlphaGenome 2 on November 11, a model that predicts how DNA sequence and genetic variation affect gene regulation in specific human cell types and states. The first AlphaGenome predicted regulatory activity at the level of tissues, averaging over the many cell types a tissue contains. The new version bridges single-cell biology and human genetics, which matters because a variant that drives disease often does so by acting in one cell population that is a small minority of the tissue around it.

The key architectural change is from parallel prediction across thousands of experiments to a conditional model that handles one biological condition at a time. DeepMind says this lets the system capture subtle differences between closely related cell types instead of spreading its capacity thinly across all tissues at once. The model was trained on more than 107 million individual cell profiles spanning 68 tissues, using both bulk and single-cell data from human and mouse samples.

The reported gains are substantial. On immune cell datasets, AlphaGenome 2 identified causal variants affecting gene expression in 30 percent of cases against 16 percent for its predecessor, nearly double. Accuracy in predicting the direction of a variant’s effect on a gene rose to 43 percent from 23 percent. The model also tracks continuous changes in gene expression as cells differentiate with much greater specificity than before.

DeepMind illustrates the model with three kinds of use. It reproduced findings from landmark studies, including identifying variants that drive inflammatory bowel disease by selectively activating in gut immune cells. It can impute unmeasured properties such as chromatin accessibility from measured gene activity, which helps laboratories decide which experiments to run next. And it generalized to cellular states it had never seen, analyzing novel conditions from COVID-19 vaccine studies without retraining.

Access for non-commercial research is through the AlphaGenome API. Commercial research access is in private preview through Google Cloud, and DeepMind plans to launch AlphaGenome 2 skills inside its Antigravity platform.

Source: Google DeepMind


DeepMind Opens the Satellite and Rainfall Datasets Behind Its Weather Models

Global satellite mosaic from the SatWeave dataset
Global satellite mosaic from the SatWeave dataset

On November 13, Google DeepMind released two foundational Earth-observation datasets that it used to train WeatherNext 3, its global weather model, making them openly available on Google Cloud Storage. The move addresses a bottleneck that has little to do with model architecture: much of the raw material for AI weather forecasting has been scattered across incompatible agency archives, and assembling it is a project in its own right.

SatWeave is a unified global mosaic of imagery from five geostationary satellites, GOES-East, GOES-West, Meteosat at zero degrees, Meteosat Indian Ocean, and Himawari, updated hourly across 17 visible, water-vapor, and infrared channels. DeepMind notes that producing coherent global composites had been difficult because the satellites use different projections, formats, and scan schedules. The historical archive spans more than a decade.

PARDIG, which stands for Precipitation AI Reanalysis: Densely Inferred GPM, is a global precipitation dataset at 15-minute temporal resolution and roughly 10-kilometer spatial resolution. It combines sparse spaceborne radar from NASA’s Global Precipitation Measurement Core Observatory with SatWeave imagery and atmospheric data from the European Centre for Medium-Range Weather Forecasts to infer rainfall everywhere, including the large parts of the world without ground radar. The catalog runs from 1999 to 2026 and is also accessible through Google Earth Engine.

DeepMind reports that PARDIG improves hourly precipitation detection skill over IMERG-Final, the standard satellite precipitation product, by more than 30 percent on light-to-moderate rainfall, and that it performs especially well on extreme precipitation events. The evaluation compared against radar networks in the United States, Europe, and Japan and against more than 36,000 weather stations worldwide.

The applications DeepMind lists go beyond forecasting: solar power output prediction, wildfire monitoring, flood prediction, drought assessment, and landslide detection, with the largest benefit in regions lacking dense radar coverage. Collaborators on the release include ECMWF, NASA, JAXA, NOAA, EUMETSAT, and the Korean, Japanese, and Dutch meteorological agencies.

Source: Google DeepMind


Import AI 480: An AI System Proves a Quantum Complexity Conjecture

Jack Clark’s November 9 issue of Import AI leads with a result that he frames as a milestone: Physical Superintelligence, an AI science startup, reports that its research system, running on an Anthropic model, produced a complete proof of the Aaronson-Ambainis conjecture. The conjecture, a long-open question about the structure of bounded low-degree polynomials with implications for how much quantum algorithms can outperform classical ones, had resisted human attempts for years. According to the newsletter, the system generated the proof on its third day of operation, the proof was verified in the Lean proof assistant, and the system then refined its result twice before final acceptance.

Clark’s commentary is characteristically direct. He describes the episode as “mostly automated synthetic researchers extending the frontier of knowledge” and suggests it is the kind of event that marks an entry into a period where technological progress outpaces human comprehension. The point is not that an AI assisted a mathematician, which has become routine, but that the system set out, found, and machine-checked the proof with humans mostly watching.

The issue’s second major item is a distributed training result from Core Automation, a startup founded by OpenAI and Google alumni. The company trained a 32-billion-parameter mixture-of-experts model with 32-fold sparsity across two clusters 9,000 miles apart, using 160 B300 GPUs in the United States and 160 GB300 GPUs in Asia, synchronizing every 50 training steps with a variant of the Distributed Shampoo optimizer. Clark reads this as a sign that compute scarcity is pushing newer labs toward training across geographically dispersed “compute islands,” and notes the startup’s stated ambition to eventually train trillion-parameter models on orbital hardware.

Rounding out the issue, Clark covers Meta’s Tactus tactile sensing glove for collecting manipulation data and revisits a 2016 NASA Kennedy Space Center paper on bootstrapping a self-sustaining lunar industrial economy. His observation on the latter is pointed: the 2016 authors expected “mouse-level” AI by 2030 and “human-level” by 2050, and Clark judges that the 2020s have already delivered systems well past the first marker, which makes the paper’s robotic-industry scenario more plausible than its authors assumed.

Source: Import AI


Amazon’s HOPE Prunes Half the Experts From a Mixture-of-Experts Model by Modeling How They Cooperate

HOPE expert pruning results chart
HOPE expert pruning results chart

Mixture-of-experts models have become the dominant architecture for large open-weight systems because they activate only a fraction of their parameters per token, keeping compute costs down. The catch is memory: Amazon researchers Alex Tseng, Luca Zancato, and colleagues note in a November 9 Hugging Face post that in a model like Qwen3.5-122B, 95 percent of parameters belong to expert layers that sit idle for any given token yet must still be loaded. Pruning rarely used experts is the obvious answer, and it has been tried, but existing methods evaluate each expert in isolation.

HOPE, for Higher-Order Pruning of Experts, starts from the observation that experts form cooperative groups that are consistently activated together, so removing one member can damage the others’ usefulness in ways a per-expert importance score never sees. The method builds what the authors call an F-matrix, whose diagonal entries capture each expert’s individual importance and whose off-diagonal entries capture pairwise cooperative contributions. Pruning then becomes a quadratic program that minimizes a tractable upper bound on the error introduced. In practice it takes three steps: calibrate on representative data, solve the program, and write out the pruned checkpoint.

The authors tested HOPE on Qwen3.5-122B, Qwen3.5-35B, and the 106-billion-parameter GLM-4.5-Air against four competing methods. Across the benchmark suite it achieved an average rank of 2.07 out of five, finishing first 39 percent of the time. Its advantage widens at aggressive pruning rates: at 50 percent pruning its average rank improves to 1.58 against 2.42 for REAP, the strongest prior method. On SWE-Bench Pro with Qwen3.5-35B pruned by half, HOPE scored 42.3 against REAP’s 36.2, and the authors report gains of up to 6.1 points on agentic coding tasks.

The practical consequence is that a model which needed two accelerators to serve might fit on one, with modest capability loss, and the open-source implementation on GitHub under awslabs/HOPE lets any operator of an open-weight MoE model try it on their own hardware.

Source: Hugging Face


PayPal Raised Its AI Assistant From 43 to 95 Percent Accuracy Without Touching the Model

Illustration of an engineer swapping AI components
Illustration of an engineer swapping AI components

Taryn Plumb’s November 10 VentureBeat report on PayPal’s internal AI assistant is a case study in where enterprise AI performance actually comes from. The assistant answers analytical questions against PayPal’s business data, and its accuracy rose from 43 percent to 95 percent while the underlying large language model stayed exactly the same. “The LLM is still the same,” said Ren Zhang, PayPal’s vice president and global head of AI product. “If the model is getting better, I’ll just plug in and play.”

What changed was a separate knowledge layer the team built around the model. It contains business-specific definitions, “golden queries” that encode the standard way to pull a given metric from PayPal’s databases, and analytical playbooks that lay out step-by-step investigation procedures for different question types. Because the layer is model-agnostic, PayPal can adopt a stronger model as soon as one appears without rebuilding its institutional knowledge. The team also tested fine-tuning smaller 8-billion-parameter models to cut latency, and found the gains real but accompanied by an ongoing maintenance burden every time a new base model arrived.

The results PayPal cites are operational rather than benchmark-driven: more than 1,000 questions answered and roughly 1,600 hours saved over ten weeks. The company has extended the pattern into multi-agent systems, including a seven-agent incident response workflow monitoring more than 33,000 services and an 18-agent security review workflow, both with human verification checkpoints. “Every decision that is proposed is vetted against the source of truth,” said Deepika Khera, senior director of AI. An episodic memory system records each session and distills lessons nightly.

The article pairs PayPal’s story with survey data on why so many enterprise deployments disappoint. Forty-two percent of surveyed organizations reported pre-deployment evaluations that later failed in production, and in 49 percent of those cases the failures reached customers. Thirty-nine percent lack automated production monitoring and learn about failures from users. The lessons PayPal draws are to separate knowledge from models, test on the real production query distribution, weigh customization against its maintenance cost, and keep humans verifying consequential decisions.

Source: VentureBeat


Swapping AI Models in Production Takes Most Companies Weeks, Survey Finds

Model swap survey lead image
Model swap survey lead image

With three frontier labs releasing new models in a single week, the question of how easily an enterprise can actually move to a better one is timely. Matt Marshall’s November 12 VentureBeat piece reports results from the VB Pulse survey, run by VentureBeat Intelligence in October 2026 among organizations with at least 100 employees. It asked whether companies had replaced the AI model in a production system over the past year while keeping the application intact, and what happened when they did.

Forty-seven percent of respondents said their companies had swapped a production model. Of those, 70 percent needed weeks or more of engineering work beyond the model change itself, and 36 percent reported swaps that ran past a month. Only 13 percent described the swap as approximating a configuration change. The outcomes split nearly evenly: 55 percent saw equal or better performance afterward, while 45 percent experienced degradation that forced rework or a reversion to the original model.

The explanation the article offers is that applications become tailored to a specific model’s behavior. Prompts, tool definitions, and tests are built around the first model’s tendencies, and fine-tuning does not transfer at all. The surrounding harness, meaning the code that wraps the model and controls its tool access, turns out to dominate results. Marshall reports that the same model can cost five times more or less depending on harness configuration, and that smaller models are especially sensitive, showing swings of 20 to 40 points on evaluations depending on how they are wrapped.

The economic point follows directly. Cost per token and cost per successful result are different quantities. A cheap model that fails 80 percent of the time costs more than an expensive one that succeeds 90 percent of the time once retries are counted. The article’s prescription is practical: build evaluation sets from real usage, test candidate models inside the actual harness rather than in isolation, separate business logic from the model so it survives upgrades, and monitor production performance after the swap. Companies that treat a model change as a configuration edit are the ones most likely to land in the 45 percent.

Source: VentureBeat