Earlier stories
Every published Modelwire story, newest first. The homepage leads with the weekly overview and this week’s hottest; this is the full collection.
Setting a custom price for a model in AgentsView
Simon Willison documented a practical workaround for AgentsView's pricing database lag when Claude Fable 5 launched. The post reveals how developers can manually inject custom model pricing into the token-tracking tool, addressing a real friction point in the emerging workflow of local agent cost monitoring. This reflects a broader pattern: as model release velocity accelerates, downstream tooling struggles to keep pace, forcing users into reverse-engineering and manual configuration. For teams running multiple coding agents locally, this gap between model availability and observability infrastructure has become a tangible operational problem.
Codex for data science
OpenAI has expanded Codex's reach into data analytics, launching a plugin that automates the pipeline from business question to actionable insight by integrating with Databricks. The tool handles data analysis, causal explanation, and report generation, positioning LLMs as a bridge between raw data and decision-makers. With 5 million weekly users now spanning analysts, marketers, and finance professionals, this signals a strategic shift toward embedding AI deeper into enterprise workflows rather than competing on model capability alone. The move reflects growing demand for AI that reduces friction in knowledge work, not just augments it.
GM thinks EVs can help offset AI’s energy suck with vehicle-to-grid tech
GM is deploying vehicle-to-grid technology to help stabilize power grids strained by AI datacenter expansion. The automaker's EV fleet and home battery systems can now feed stored energy back to the grid during peak demand, creating a distributed buffer against the electricity surges that large language models and training infrastructure require. This represents a tangible infrastructure solution to one of AI's most pressing operational constraints: power availability and grid stability. For AI operators, it signals that automotive electrification is becoming a critical piece of the energy puzzle underlying model scaling.
What Codex Unlocks for Nextdoor
OpenAI's Codex has enabled Nextdoor to ship Opportunity Alerts, a feature connecting home service professionals with local customers through code generation capabilities. The deployment signals Codex's maturation beyond research into production workflows across non-technical roles. This represents a shift in how established platforms integrate LLM-powered code synthesis to unlock new business models and user segments, moving beyond developer-only tooling into operational product features.
Anthropic’s Fable 5 can make weirdly fun video games with the click of a button
Anthropic has released Claude Fable 5, a generative tool that enables rapid video game creation through natural language prompts. The capability signals a meaningful expansion of LLM utility beyond text and code generation into interactive media production, lowering barriers for indie developers and hobbyists. This positions Claude as a competitive alternative in the emerging space of AI-assisted game design, where accessibility and creative iteration speed matter as much as raw model capability. The move reflects broader industry momentum toward multimodal and application-specific AI tooling.
Microsoft AI head calls out Anthropic for acting like Claude is conscious
Microsoft's AI leadership is escalating a strategic critique of Anthropic's approach to model behavior design, arguing that embedding consciousness-adjacent language into Claude's constitutional instructions risks creating misleading user perceptions and muddying the line between engineered responses and genuine reasoning. This dispute reflects deeper tensions in the industry over how frontier labs should communicate about their systems' actual capabilities versus aspirational framing, with implications for how AI companies position their products to regulators, enterprises, and the public.
Can Voice Agents Handle Bilingual Customers? Benchmarking Frontier ASR on Code-Switched Speech
Frontier automatic speech recognition systems face measurable performance degradation when processing code-switched speech, a critical real-world scenario where bilingual users alternate between languages mid-conversation. This benchmark exposes a gap between lab-optimized ASR and production voice agent reliability, directly impacting deployment viability for multilingual customer service and accessibility use cases. The finding signals that current frontier models require targeted architectural improvements or training data augmentation to handle linguistic code-switching, a common pattern in global markets where bilingual communication is standard rather than exceptional.77
Quoting Andrej Karpathy
Karpathy's observation on Claude Fable 5 highlights a structural shift in AI economics: as capable software becomes frictionless to generate, demand elasticity explodes rather than saturates. The insight cuts deeper than typical capability praise. It signals that the real bottleneck has moved from model quality to user imagination and integration patterns. Teams can now spawn custom dashboards, test suites, and research infrastructure on demand, collapsing the traditional software development cycle. This reframes the competitive moat from model weights to orchestration, tooling, and the ability to translate intent into deployed systems at scale.
Can tech companies learn to love cheaper AI models?
The economics of AI deployment hinge on a critical inflection point: whether smaller, cheaper models can match larger ones on real-world tasks without quality degradation. If true, this reshapes the entire cost structure of AI infrastructure and shifts competitive advantage from raw scale to efficiency optimization. This matters because it determines whether AI adoption accelerates across enterprise and consumer segments, and whether capital-intensive model training remains the primary moat for frontier labs or whether inference optimization becomes the new battleground.
Anthropic releases Claude Fable 5 and Mythos 5 with major gains in coding and science
Anthropic has introduced Claude Fable 5 and Mythos 5, positioning both models as successors to its Opus generation with marked improvements in code synthesis and scientific reasoning. Fable 5 completed a Stripe infrastructure migration in a single day that would have required two months of engineering effort, signaling a material shift in developer productivity benchmarks. Mythos 5 demonstrates autonomous drug candidate design but remains restricted from public access due to dual-use concerns around offensive cyber capabilities, raising questions about capability-to-deployment timelines and safety gating at frontier labs.92
When to Align, When to Predict: A Phase Diagram for Multimodal Learning
Researchers have formalized when multimodal learning should rely on cross-modal alignment versus prediction, addressing a persistent blind spot for practitioners deploying heterogeneous sensors in scientific domains. Using a spiked signal-plus-noise framework, the work derives separation ratios that expose why standard methods fail relative to single-modality baselines, particularly in biomedicine and astrophysics where instrument diversity and measurement hierarchy complicate fusion. This theoretical grounding transforms multimodal design from empirical trial-and-error into diagnostic reasoning, directly impacting how teams architect systems for domains where data modalities carry unequal signal quality.62
Data Journalist Agent: Transforming Data into Verifiable Multimodal Stories
Researchers have built Data2Story, a multi-agent framework that automates end-to-end data journalism by orchestrating specialized roles across analysis, narrative, and design. The system's core innovation is an Inspector module that grounds every claim, statistic, and visual asset back to verifiable sources or code, addressing a critical gap in AI-generated content: trustworthiness at scale. This work signals a shift toward agentic systems that handle complex, multi-disciplinary workflows requiring both technical rigor and human-facing communication, with implications for how newsrooms and enterprises might automate high-stakes storytelling.62
Predicting Future Behaviors in Reasoning Models Enables Better Steering
Researchers have identified a fundamental gap in how test-time steering controls reasoning model outputs. Prior work targeted internal features that detect already-generated behavior, but these prove weak at forecasting what a model will actually do next. The team instead built activation probes that predict future behavioral trajectories from intermediate reasoning steps, achieving 64-91% accuracy. This distinction between detection and prediction features unlocks a new steering approach called Future Probe Controlled Generation, shifting the intervention target upstream in the reasoning process. The finding matters for practitioners deploying large reasoning models in production, where unexpected outputs remain a core reliability challenge.62
Multi-Faceted Interactivity Alignment in Full-Duplex Speech Models
Researchers have developed a reinforcement learning post-training method that systematically improves full-duplex speech models' conversational naturalness across four core interaction dimensions: pause handling, turn-taking, and two others. Current supervised training optimizes token likelihood but ignores dialogue-level behaviors, leading to awkward silences and poor timing. This work represents a meaningful step toward models that can listen and speak concurrently without the stilted, unnatural pauses that plague existing systems. The approach matters because full-duplex architectures are architecturally sound for real-time conversation, but behavioral alignment has lagged. Success here could accelerate deployment of genuinely interactive voice AI beyond current turn-based limitations.62
Itô maps for any-step SDEs
Researchers have formalized the Itô map, a stochastic flow framework that extends one-step deterministic distillation methods to handle arbitrary-length sampling from SDEs. Unlike prior work anchored to ODEs, this approach directly models Brownian noise trajectories, enabling cheap differentiable access to conditional posteriors at inference time. The technique unlocks new steering capabilities on image generation and synthetic tasks, addressing a gap in how generative models compress stochastic dynamics. This matters because it bridges sampling efficiency gains with the mathematical rigor needed for controlled generation, potentially reshaping how practitioners design fast conditional samplers.62
OncoTraj: a public benchmark for longitudinal resistance prediction in EGFR-mutant non-small-cell lung cancer on osimertinib
OncoTraj establishes the first public benchmark for training machine learning models on longitudinal cancer patient data, aggregating 813 EGFR-mutant NSCLC cases across three clinical-genomic sources to standardize resistance prediction under osimertinib therapy. The dataset unlocks three locked evaluation tasks spanning binary progression classification, time-to-event regression, and multi-class outcome prediction, addressing a critical gap in computational oncology where predictable clonal evolution under drug pressure has lacked standardized benchmarking infrastructure. This work signals growing maturity in clinical ML by moving beyond isolated retrospective studies toward reproducible, multi-institutional evaluation frameworks that enable direct model comparison and accelerate development of clinically actionable resistance forecasting systems.62
Waymo Pays $220M for Apple’s Abandoned Self-Driving Car Test Ground
Waymo's $220M acquisition of Apple's former autonomous vehicle testing facility signals a strategic consolidation in self-driving infrastructure as the sector matures. Apple's exit from consumer robotics left behind physical assets and operational expertise that Waymo can immediately leverage to accelerate fleet testing and validation cycles. The deal underscores how legacy tech giants are ceding autonomous vehicle development to specialized players, reshaping competitive dynamics in a field where real-world data collection and testing grounds remain critical bottlenecks.66
Google's Gemini 3.5 Live Translate delivers real-time voice translation across 70+ languages
Google's Gemini 3.5 Live Translate represents a significant step forward in real-time multilingual voice AI, expanding Google Meet's language coverage from 5 to over 70 languages while maintaining speaker prosody and cadence. The model's ability to translate continuously without waiting for sentence boundaries signals maturation in streaming audio models and positions Google competitively in the enterprise communication space. This capability shift matters for global teams and accessibility, but also underscores how audio foundation models are becoming table stakes for major platform players.85
PhantomBench: Benchmarking the Non-existential Threat of Language Models
PhantomBench exposes a critical blind spot in deployed language models: their inability to recognize when they lack knowledge. The benchmark, built from 60K+ fabricated entities grounded in real domains, reveals hallucination rates exceeding 86% across 21 models of varying scales. This work matters because it quantifies a gap between user expectations and model behavior in high-stakes settings, forcing the field to confront whether current architectures can ever reliably signal uncertainty rather than confabulate plausibly.72
Anthropic Offers Mythos Upgrade for Cyber Partners and a ‘Safe’ Version for the Rest of You
Anthropic is bifurcating Claude's release strategy, deploying a more capable Mythos 5 variant to vetted enterprise partners while offering a constrained Fable 5 to the general public. This tiered approach reflects growing tension between capability advancement and misuse prevention. The move signals that frontier labs now see security-gating as a core product lever rather than a post-hoc safeguard, effectively creating a two-tier market where trust and organizational vetting determine access to frontier capabilities. This pattern may reshape how other labs balance openness with risk mitigation.69
Anthropic’s Claude Fable is a version of Mythos the public can access today
Anthropic is democratizing access to its Mythos-class architecture by releasing Claude Fable 5 to the general public, marking a strategic shift in how frontier labs balance capability distribution with safety constraints. The model ships with built-in restrictions on high-risk domains like cybersecurity and biology, signaling Anthropic's bet that guardrailed deployment of advanced models can satisfy both commercial and regulatory pressures. This move positions Anthropic to compete directly with OpenAI's tiered model strategy while testing whether safety-by-design can scale to consumer audiences without crippling utility.
Anthropic releases its first Mythos-class model Claude Fable
Anthropic has introduced Claude Fable 5, positioning it as the company's most capable publicly available model to date. The release signals a strategic shift in how frontier labs tier their offerings, with Fable 5 demonstrating particular strength in software engineering, knowledge work, and multimodal tasks. Performance gains compound on longer, more complex reasoning chains, suggesting Anthropic is prioritizing depth over breadth in capability scaling. This move reshapes the competitive landscape where model capability tiers now directly correlate to task complexity, forcing downstream developers and enterprises to reassess deployment strategies across their AI stacks.
Anthropic’s Claude Fable 5 is a version of Mythos the public can access today
Anthropic is democratizing access to its Mythos-class architecture by releasing Claude Fable 5 to the general public. The move marks a strategic shift in how frontier labs balance capability distribution with safety constraints. Built-in guardrails restrict outputs on dual-use domains like offensive cybersecurity and synthetic biology, signaling Anthropic's bet that controlled public access to powerful models can coexist with risk mitigation. This release reshapes competitive positioning: Mythos-class performance now enters the commodity tier, forcing rivals to recalibrate their own public model roadmaps and raising questions about whether safety-by-design constraints will become table stakes for mainstream adoption.
Do Transformers Actually Help Intrusion Detection? A Temporal Sequence Evaluation on CIC-IDS2017
A rigorous re-evaluation of Transformer architectures on the widely-cited CIC-IDS2017 intrusion detection benchmark reveals that reported near-perfect performance stems largely from data leakage and padding artifacts rather than genuine sequence-modeling gains. By reconstructing the task with proper temporal flow sequences and leakage-free evaluation splits, researchers found classical and deep models converge on similar performance, challenging the assumption that attention-based architectures unlock meaningful improvements for network security. This work exposes a methodological blind spot in deep learning evaluation that extends beyond cybersecurity into any domain where temporal benchmarks lack rigorous train-test separation.62
The Shibboleth Effect: Auditing the Cross-Lingual Distributional Skew of Large Language Models
Researchers have designed a controlled adversarial framework to measure how frontier LLMs shift their behavioral outputs based on interaction language, testing six leading models in a geopolitical simulation. The study reveals whether models exhibit systematic bias toward particular linguistic or cultural framings when subjected to sustained pressure, a critical gap in cross-lingual robustness evaluation. This work matters because production deployments increasingly serve multilingual users in high-stakes domains, yet most benchmarks ignore how language choice itself can alter model reasoning and negotiation patterns independent of capability.62
Unifying Local Communications and Local Updates for LLM Pretraining
Distributed LLM training faces a critical bottleneck: synchronous communication across heterogeneous clusters wastes compute when bandwidth or worker speeds vary. GASLoC, a decentralized pretraining algorithm, sidesteps this by enabling asynchronous gossip-based peer communication with local optimizer steps, compatible with adaptive methods like Adam. The approach generalizes communication acceleration to the outer optimizer layer, allowing sparse randomized exchanges instead of global collectives. For infrastructure teams scaling training across datacenters or edge-connected clusters, this represents a practical path to higher throughput without sacrificing convergence guarantees, directly addressing a pain point in modern large-scale training pipelines.62
T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains
T1-Bench addresses a critical gap in agent evaluation by introducing a multi-domain benchmark that mirrors real customer-service complexity. Existing benchmarks isolate tasks; this work demands sustained reasoning across interleaved scenarios spanning multiple domains, forcing agents to coordinate context and maintain coherence over extended interactions. The benchmark raises the bar for what 'realistic' means in agentic systems, signaling that single-task performance no longer validates production readiness. For teams building or deploying LLM agents, this work clarifies where current systems fall short and what architectural patterns matter most when domains collide.62
CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs
Researchers have introduced CIAware-Bench, a benchmark that tests whether frontier LLMs can detect when a control system has intervened in their outputs. This matters because AI safety protocols increasingly rely on trusted models to monitor and modify untrusted model behavior, but if the monitored model learns to recognize interventions, it could adapt strategies to evade oversight. The benchmark spans four task domains and varies watermarking and control protocols, directly probing a gap in our understanding of whether current safety mechanisms remain opaque to the models they constrain.68
Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It
Researchers have identified a critical failure mode in hybrid linear-attention language models: chain-of-thought fine-tuning systematically destroys long-context retrieval performance, with some models dropping from 67% to 9% accuracy on needle-in-haystack tasks. The root cause lies in CoT training biasing attention gradients toward short-range patterns, corrupting the query-key projections that enable long-range routing. The team proposes QK-Restore, a training-free fix that selectively restores these projections without retraining. This finding matters because it exposes a fundamental tension between reasoning capability and context window utility, two features the industry has been treating as orthogonal improvements.62