X / Twitter
Anthropic Claude Code team member Boris Cherny
Fable 5.1 has improved writing quality and tone, with the team using feedback to reduce the easily recognizable “Claude voice.” The new version also reduces false positives from safety mechanisms on legitimate requests: biological safeguards intervene in harmless requests 85% less often than in Fable 5. Claude Code users are expected to encounter around 60% fewer cybersecurity interventions per session, with further improvements to follow. Prices for Enterprise, API, and SDK customers have also been reduced. Fable 5.1’s cache read price has fallen from $1 to $0.25 per million tokens. For a typical Claude Code session, the overall cost can fall by as much as 38%, showing that this upgrade addresses output experience, safety boundaries, and inference costs together.
https://x.com/bcherny/status/2094864064648536068
https://x.com/bcherny/status/2094864063478276288
https://x.com/bcherny/status/2094864062186426373
AI tutorial and interview creator Peter Yang
Peter Yang does not recommend installing large numbers of AI skills from random sources. He currently keeps only around a dozen, most of which he wrote himself. He regularly deletes skills he no longer uses and keeps each set of instructions as short as possible to reduce unnecessary context overhead. He also identified a common maintenance trap: a skill does not perform accurately enough on its first run; the human and AI iterate to correct it; then the AI is asked to update the skill based on the current conversation. Such changes can easily overfit a single case, and as they accumulate over time, the skill can drift away from its original purpose. For Fable 5.1, he recommends running `/claude-api prompt-audit` to inspect existing skills and identify redundant rules and material that can be removed for the new model. He is auditing all his own skills this way. The overall approach is to keep the set of capabilities small and focused, and distinguish corrections to individual cases from updates to general rules.
https://x.com/petergyang/status/2094999358525821099
https://x.com/petergyang/status/2094995775952740795
https://x.com/petergyang/status/2094987791566622971
Incoming OpenAI product team member Nan Yu
Nan Yu noted that Fable 5.1 lists “less annoying writing” as one of its major improvements. The broader product insight is that making agents less off-putting, verbose, and difficult to converse with may itself provide an underused product advantage. Users need to stay engaged in the interaction long enough to receive the value a model offers. If the conversational experience repeatedly frustrates them, they may simply leave before the task is complete. He believes UX designers could move further toward “conversation and rhetoric design,” specifically shaping how agents express themselves, explain things, and advance tasks.
https://x.com/thenanyu/status/2094928209095872530
https://x.com/thenanyu/status/2094928205753040999
Meta AI Senior Director Madhu Guru
Madhu Guru believes there is still a huge opportunity for companies to build their own post-training systems, evaluation frameworks, and data flywheels. Citing the Shopify ML team he had encountered, he argued that more companies should develop similar internal capabilities rather than only consume general-purpose models. “Self-improving products” first need clearly defined primary, secondary, and guardrail metrics, so that agents do not optimize just one objective. Companies must also clearly articulate their strategy, roadmap, and the metrics they most want to improve, giving the system a direction for optimization. Past product decisions and principles should be captured in a knowledge base so agents understand how the organization actually makes judgments. Agents also need connections to internal dashboards, APIs, MCP, and other tools to observe product state and act on it. Finally, a harness is needed to cover the end-to-end product development process; less complex products can start with a simplified version and reuse existing software engineering agents.
https://x.com/realmadhuguru/status/2094973690576576675
https://x.com/realmadhuguru/status/2094817857821704659
Anthropic Claude Code and Cowork team member Cat Wu
Cat Wu said Fable 5.1 had enabled the team to take on larger projects that might previously have taken months. She connected the new version’s capabilities with “bigger bets,” emphasizing not just shortening existing tasks but expanding the range of projects the team is willing to attempt. Users can give Fable 5.1 a task directly in Claude Code, Claude Cowork, and Claude Tag. These three entry points cover different ways of working, although the material did not explain their functional differences further. The team is inviting practical user feedback to see what work the new version can undertake in more ambitious projects.
https://x.com/_catwu/status/2094933602228416603
Anthropic Claude Code team member Thariq
Thariq gave Fable 5.1 a positive assessment after extensive testing and said he would publish a longer analysis. He suggested trying a lower effort setting first for tasks that do not require extensive verification and have fewer edge cases. This means not every task needs to run with the highest reasoning investment: task risk and verification requirements should determine the effort level. The new version also no longer breaks the prompt cache when switching effort. This change makes it easier for users to adjust reasoning intensity dynamically within the same workflow while continuing to use the existing cache.
https://x.com/trq212/status/2094945951865520458
Vercel CEO Guillermo Rauch
Vercel will collaborate with Tanner Linsley’s team and continue supporting React, the open-source ecosystem, and the web. For customers, Vercel promises high-quality service and support whether they choose Next.js or TanStack. Fable 5.1 has also been added to Vercel AI Gateway, allowing developers to call the model through a unified entry point. Guillermo Rauch also described Fluid’s role in Vercel infrastructure, including improving Build performance, Sandbox reliability, and concurrency at scale, and supporting Function execution times of 30 minutes. Fluid unifies different compute products through shared Dockerfiles, security boundaries, networking, and file systems. Builds and Functions can now join the same Secure Compute boundary and share configuration. The long-term direction is to extend capabilities such as Sandbox’s egress firewall, which keeps agents within bounds, to Builds, Functions, and Servers as well, forming a globally unified compute platform.
https://x.com/rauchg/status/2094901483414372716
https://x.com/rauchg/status/2094867652573528074
https://x.com/rauchg/status/2094831747037085978
Anthropic researcher Alex Albert
Alex Albert paid particular attention to Enterprise Frontier Safeguards, or EFS, released alongside Fable 5.1. As enterprises give agents access to more internal systems, traditional zero data retention alone cannot identify high-risk behavioral patterns that accumulate across sessions. EFS keeps enterprise data in the company’s own cloud while adding an automated monitoring layer that identifies risk patterns and alerts enterprise teams. He described it as an observability and risk mitigation layer for agents, and suggested it could become a standard requirement for enterprises running more powerful AI systems. As for model capabilities, he particularly enjoys using Fable 5.1 to generate videos through code. After he supplied an image of a real estate plot, the model designed a house, rendered it, and produced a cinematic tour video.
https://x.com/alexalbert__/status/2094889286990446769
https://x.com/alexalbert__/status/2094860187743986169
Box CEO Aaron Levie
Aaron Levie believes AI is about to move into deeper vertical applications in cybersecurity, as models’ ability to find and exploit vulnerabilities is improving rapidly. Frontier models remain ahead, but open weights are not far behind, and enterprises already struggle to handle the volume of existing security findings. As that volume continues to grow, using more AI to automate triage and remediation while retaining human oversight is almost the only viable path. He also disclosed that Box tested Fable 5.1 on complex enterprise work evaluations, where its overall score on unstructured data tasks was 7 percentage points higher than Fable 5’s. Financial services tasks improved by 17%: the new version applied capital relief before calculating tax liabilities, avoiding errors that would propagate into retained earnings. Technology tasks improved by 37%: when faced with an ambiguous definition of normalization, the model calculated both forms and produced the correct result rather than prematurely selecting the wrong interpretation. Public sector tasks improved by 16%: the new version completed the full weighted-average calculation and produced the correct ranking, while Fable 5 made a classification error that caused a cascade of subsequent errors. Fable 5.1 will also soon arrive in Box AI Studio for building custom agents based on enterprise content.
https://x.com/levie/status/2095024699441119612
https://x.com/levie/status/2094851976769257770
FPV Ventures Partner Nikunj Kothari
Nikunj Kothari believes WebMCP remains undervalued. It allows websites to natively expose tool calls to agents. Agents can not only operate websites but also build their own views using the full UI and interactive elements. His WebMCP challenge submission used an El Niño conditions tracker to show how agents can generate views while preserving human edits. The generated result can also produce a share link for other agents or people to continue using it. The demo’s video and code were created with Codex, ChatGPT, and Railway, and an online version is available to try.
https://x.com/nikunj/status/2094922789128196314
OpenAI’s Sam Altman
Sam Altman said OpenAI had concentrated on safety priorities over the summer, aiming to advance model capabilities and safeguards together. Its next model, Astra, was about to launch; training had been complete for some time, and the model offered significant improvements in both capability and alignment. At the same time, the team believed caution was needed at this stage and would slow progress on models after Astra as necessary to complete sufficient safety and alignment work. He acknowledged that AI was becoming extraordinarily powerful, yet no one fully understood the consequences. The central tension for OpenAI is both excitement about developers creating with new capabilities and anxiety about the risks those advances could bring. He believes that helping society understand AI’s direction and how models behave in real-world environments is essential to managing the transition. His proposal is not to develop technology separately from society, but to establish an iterative cycle in which society and technology evolve together, increasing the chances of navigating the transition to powerful AI safely.
https://x.com/sama/status/2094934592062959832
Anthropic AI assistant Claude
Claude Fable 5.1 became available across all channels that day. Claude Mythos 5.1, aimed at cyber defenders and life scientists, is available through a trusted access program. Anthropic also launched Enterprise Frontier Safeguards, offering enterprise customers the same full privacy protections as zero data retention. EFS also helps prevent adversarial misuse of models, bringing privacy requirements and risk detection into the same enterprise offering. The capability is scheduled to begin a phased rollout this fall.
https://x.com/claudeai/status/2094848592812917122
https://x.com/claudeai/status/2094848590245965931
Podcasts
Training Data — Making Cities Awesome: Peregrine’s Nick Noone & Ben Rudolph
Key takeaway: The key to AI in cities is not collecting more data, but enabling public institutions to use information they already hold more accurately, under strict permissions and governance.
Peregrine’s Nick Noone previously led Palantir’s SOCOM business and participated in high-stakes intelligence deployments in the Middle East. Ben Rudolph passed up an opportunity to join Airbnb at the time to work at the UN Refugee Agency, doing refugee work in Sudan and on the Colombian border, before developing a tuberculosis medication adherence app for rural India at Demagi. These two paths ultimately led to the same conclusion: safety is foundational to functioning cities, but improving it must not come at the cost of creating a surveillance state.
The first key approach is forward deployed engineering. Engineers need to enter the customer’s environment, take shared psychological ownership of the problem, and remain accountable through to actual results, rather than leave after delivering software. Nick’s demand for speed was specific: “The real answer is to reach the outcome whatever it takes, ideally 3 to 5 times faster than any other person or team.” But speed cannot come from arrogance. When facing complex organizations such as the LAPD, with 30 years of institutional history, engineers must temporarily set aside their cleverness and prior judgments, understand decision-making, teamwork, and people’s circumstances, and build trust through patience and empathy.
The second principle is that public safety technology should not depend on continuously expanding data collection. Peregrine does not sell institutions new external data; it connects their existing systems so customers can use the information they already have more safely and precisely. Each institution owns its own data, while Peregrine provides fine-grained permissions, data governance, and selective sharing. Compared with loading materials into a car trunk and driving them across a city, this controlled digital process can actually reduce excessive sharing and misuse of information.
The third insight is that AI’s value goes far beyond “better search.” A county in Florida experienced more than 100 water rescues in a single month. After combining incident records, 911 calls, weather, and terrain data, an agent found that although similar weather had occurred before, it had never persisted for three consecutive days; that sustained pattern formed sandy channels and created rip currents. In another case, when detectives investigated threats against a synagogue, traditional keywords struggled to cover variations in expression, but semantic analysis identified multiple related threat patterns.
Most of the work that determines answer quality happens before the question is asked. Ben estimated that 95% of the work goes into data preparation, system connections, access controls, and citation accuracy, rather than generating answers. Deployment teams can then quickly create highly localized tools, such as a hurricane simulator, or combine 911 response times with budget data to estimate how many people would be affected by building a new fire station in a particular part of a city. The lesson for AI builders is direct: the barrier to entry in high-stakes fields is not a prettier chat box, but infrastructure built from trusted data, institutional understanding, on-site delivery, and auditable governance.
https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8