Next Week in The Sequence:

We continue our series about model distillation.
The AI of the Week, dives into Google’s new TabFM model for tabular and time series data.
The opinion section, discusses that domains are a good fit for rapid progress of AI models vs which ones are challenging.

Subscribe and don’t miss out:

TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Editorial: Fable 5's Comeback, ZCode's Debut, Claude Science, and the $3.5B Deployment Land Grab

This week’s developments in AI provide a clear direction where the space is going: more capable models and the imperative of capable delivery capabilities.

Start at the model layer, where we just watched the first frontier model get hot-patched by a government. Claude Fable 5 came back online July 1 after 19 days behind an export-control firewall — pulled three days post-launch when Amazon researchers found a jailbreak that coaxed it into surfacing software vulnerabilities. The fix is beautiful in a systems-engineering way: a classifier that catches the technique in over 99% of attempts and, instead of throwing an error, gracefully degrades the request to Opus 4.8. Not a 403. A fallback route. Frontier deployment now looks like a load balancer with a compliance layer, and Anthropic, Amazon, Microsoft, and Google are literally drafting a CVSS for jailbreaks — severity scoring for prompts. Take a second with that.

While the US was busy negotiating classifiers, Z.ai shipped an existence proof. ZCode is a free “Agentic Development Environment” wrapped around GLM-5.2: a 744B-parameter MoE with 40B active, a genuine one-million-token context window, 28.5 trillion training tokens, MIT-licensed weights — and trained on Huawei silicon, no American chips required. The pitch practically writes itself: self-host the weights and there is no kill switch. The Fable suspension turned “regulatory tail risk” from a slide in someone’s deck into a lived 19-day outage, and Z.ai is selling the mitigation. Same story, opposite sides of the Pacific.

One layer up, Anthropic launched what I’d callclaude --science. Claude Science isn’t a new model — it’s a workbench: a coordinating agent with 60+ curated skills and connectors across genomics, proteomics, structural biology, and cheminformatics, rendering protein structures natively, running on your own cluster over SSH, with every figure shipping alongside the exact code and environment that produced it. Reproducibility as a first-class primitive. A reviewer agent even runs alongside the pipeline, flagging citations that don’t resolve — CI/CD for manuscripts. The bet is explicit: do for wet labs what Claude Code did for codebases. If the analogy holds even at 50%, biology just got its terminal moment.

And at the top of the stack, the money moved to the least sexy layer: deployment. Microsoft committed $2.5B and 6,000 engineers to Microsoft Frontier Co., two days after AWS put $1B behind its own forward-deployed engineering push. OpenAI and Anthropic stood up theirs in May. Four of the best-capitalized companies on Earth independently converged on the same gradient: models are no longer the bottleneck. Integration is. The Palantir FDE playbook — engineers embedded in the customer’s messy production environment — is now the industry’s default architecture.

Zoom out and the inversion is complete. Capability is the commodity. Classifiers, weights licenses, workbenches, and boots-on-the-ground are the moat. The frontier isn’t the model anymore. It’s the runtime around it.

AI Research

SKILLOPT: Executive Strategy for Self-Evolving Agent Skills

AI Lab: Microsoft

Summary: This paper introduces SKILLOPT, a controllable text-space optimizer for training agent skills as the external state of a frozen agent, using trajectory feedback to make bounded edits on a skill document. Across multiple benchmarks and harnesses, SKILLOPT significantly lifts average no-skill accuracy and outperforms competitors without requiring additional inference-time model calls.

Nemotron-Labs-TwoTower: Diffusion Language Modeling with Pretrained Autoregressive Context

AI Lab: NVIDIA

Summary: The authors propose TwoTower, an architecture that separates diffusion language modeling into a frozen autoregressive context tower for clean tokens and a trainable diffusion denoiser tower for noisy blocks. Instantiated on a 30B parameter model, Nemotron-Labs-TwoTower maintains near-baseline autoregressive quality while achieving 2.42x higher wall-clock generation throughput.

TUA-Bench: A Benchmark for General-Purpose Terminal-Use Agents

AI Lab: Meta AI

Summary: TUA-Bench is a new benchmark designed to evaluate general-purpose terminal-use agents across 120 diverse, real-world tasks spanning everyday digital activities and specialized scientific workflows. Evaluations reveal that even the strongest frontier agents currently achieve only a 65.8% success rate, highlighting significant remaining challenges in long-horizon planning, error recovery, and tool use within terminal environments.

Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs

AI Lab: Yale University & Google Research

Summary: This research introduces Reinforcement Learning with Metacognitive Feedback (RLMF), a paradigm that prioritizes and rewards large language models based on the accuracy of their self-assessed performance judgments. By applying this framework alongside targeted rewriting, models learn to faithfully calibrate and express their internal uncertainty numerically and linguistically across diverse tasks while maintaining task accuracy.

SkillHone: A Harness for Continual Agent Skill Evolution Through Persistent Decision History

AI Lab: WeChat, Tencent Inc.

Summary: SkillHone is a framework for the continual evolution of agent skills that maintains a persistent decision history, recording diagnoses, revisions, redacted evaluation evidence, and outcomes. This persistent context enables role-separated subagents to refine skills across sessions without re-deriving past rationale, leading to state-of-the-art performance on deep-research benchmarks like GAIA and WebWalkerQA-EN.

Introducing TabFM: A zero-shot foundation model for tabular data

AI Lab: Google Research

Summary: TabFM is a zero-shot foundation model that simplifies tabular classification and regression workflows by utilizing in-context learning to bypass the need for traditional model training, hyperparameter tuning, and manual feature engineering. Trained entirely on hundreds of millions of synthetic datasets, its hybrid architecture employs alternating row and column attention to natively capture complex feature interactions, allowing users to generate highly accurate predictions on unseen tables in a single forward pass.

烙 AI Tech Releases

Fable5

Anthropicredeployed Fable5after the US export controls were lifted.

ZCode

Z.aireleased ZCode, a new development environment based on GLM 5.2.

Claude Science

Anthropicannounced Claude Science, a workbench for scientists.

10 AI News You Need to Know About

Anthropic × Samsung custom chip— Anthropic is in talks with Samsung to explore collaborating on a custom AI chip, though it hasn’t yet decided the chip’s purpose, server role, or power level, while stressing that its diversified Google/Amazon/Nvidia hardware stack remains central to its compute strategy. Original source (The Information broke it): https://www.theinformation.com/articles/anthropic-talks-samsung-manufacture-custom-ai-chip
Meta launches Pocket— Meta quietly launched Pocket, an experimental app born from its Gizmo acquisition that lets users generate and share small interactive apps and games (”gizmos”) from AI text prompts, complete with a scrollable discovery feed. No official Meta announcement exists yet — the closest primary source is the app’s Google Play listing: https://play.google.com/store/apps/details?id=com.facebook.gizmo
Microsoft Frontier Company— Microsoft launched Microsoft Frontier Company, a new operating business backed by a $2.5 billion investment and 6,000 industry and engineering experts to deliver enterprise AI deployments, which Judson Althoff says “goes beyond” the Forward-Deployed Engineering label.
Venice AI unicorn round— Venice, the privacy-first platform offering access to 200+ AI models with client-side encryption and no server-side data storage, raised a $65M Series A led by Dragonfly at a $1B valuation — its first outside capital — while already profitable on $70M+ annualized revenue.
Etched comes out of stealth— Etched emerged from stealth with working first-pass A0 silicon on TSMC’s N4P process, $800M raised (most recently $500M at a $5B post-money valuation), and over $1 billion in signed customer contracts for its rack-scale “frontier inference clusters” shipping this summer.
Amazon’s $1B FDE org— AWS created a dedicated Forward Deployed Engineering organization backed by a $1 billion investment that embeds engineers and purpose-built agents directly inside customer teams, pitching an agentic-first model that compresses deployment timelines from months to days and leaves customers self-sufficient.
Arena hits $100M— Arena, the UC Berkeley-born crowdsourced AI leaderboard, reached $100 million in annualized run-rate revenue just eight months after launching its commercial AI Evaluations service — up from $30M at its January Series A.
Crusoe raising ~$3B— Crusoe, which supplies AI data center capacity to the likes of Meta and Oracle, is in talks to raise about $3 billion in a round investors expect to land around a $30 billion valuation — roughly triple its $10 billion mark from October. Bloomberg is the original source (scoop, no primary release): keep your link.
ElevenLabs $22B tender— ElevenLabs has held early talks with investors for a secondary tender offer letting employees sell shares at a roughly $22 billion valuation — double its February round — with the tender expected by September. Bloomberg is the original source (scoop): keep your link.
Meta cloud business— Meta is developing plans for a cloud infrastructure business selling access to AI compute and models — including a Bedrock-style offering hosting its Muse Spark models — putting it in direct competition with AWS, Azure, and Google Cloud; the news sent Meta shares up ~9% while hitting neoclouds CoreWeave and Nebius.

",{"_256":202,"_257":286,"_258":287},"46bb33d0-1875-419c-8656-491c5e24f541",{"_21":286,"_33":288,"_204":289,"_35":290,"_226":291,"_263":291,"_23":233,"_264":292,"_266":293,"_219":42,"_70":-5,"_72":42,"_182":294,"_256":202,"_283":307},"Fable is back","Organizing a taxi using an AI agent involved providing it with necessary context from calendar and emails, then allowing it to use available tools like a web browser to complete the booking. This demonstrates how agents can perform tasks independently when context and tools are sufficiently available. The article also reviews recent AI model updates, including the return of Fable 5, the introduction of Claude Sonnet 5, and new Gemini media models like Nano Banana 2 Lite and Omni Flash.","and there's a new sonnet","https://substack-post-media.s3.amazonaws.com/public/images/3a5162c9-9497-4820-bf1d-0b3ae332bebb_3380x1940.png","2026-07-02T13:04:04","https://www.bensbites.com/p/fable-is-back",{"_21":295,"_185":296,"_187":116,"_70":297,"_213":304,"_214":305,"_216":306},"57b9d75c707dad93778e0a1b3520150659fa1b0760616be5c6f2a16002d99e42","1ab4768e8d7085fda138193f9b7fb4d9e5fd46d638071a7af985b793ce3b5dd5058ef625174b0965d74c926607d8287749c5a21429c5de28085fbe6841e4a5d3",[298,299,300,302,303],[219,288],[37,291],[275,301,42,277],"30023:0b16efb3db6165f2311a31e04c8f874f8c569a4cc6cc219a3295cdfa9110b8b8:46bb33d0-1875-419c-8656-491c5e24f541",[210,42],[212,304,42],"0b16efb3db6165f2311a31e04c8f874f8c569a4cc6cc219a3295cdfa9110b8b8","Fable is back\nand there's a new sonnet\n\n\nnostr:naddr1qvzqqqr4gupzqzcka7eakct97gc35v0qfj8cwnuv26dye3kvyxdr99wdl2g3pw9cqqjrgdnzvgenxeps95cnsde4956rzwtr95urvdfk956rjvtrx4jnydrxx56rz3e7ue6 ",1782997444,"

Hey folks,

When I was in Greece, I needed to organise a taxi from the airport home. I pinged codex and 1m37s later it was done, I got a text to say the car was booked.

\"\" — Codex also added the redacted bit to this image

It’s not the most impressive task by all means but there’s a system as to why it works (and this system is how a lot of tasks work).

So first, all my day-to-day agent conversations start in my ‘bites’ folder, which has links to files that store my memories, what I’m working on, ben’s bites, about me etc.

Codex loads my AGENTS.md and my task then gets to work.

It’s got the default Google Calendar and Gmail tools connected, so starts with looking at if my flight is in my calendar (which it was, thanks to another Codex thread I did pre-trip). Finds the trip and then looks in my emails for any prior task/transfer emails, finds one to get my home address and the company I used on the way out (these could also be noted in my folders’ files, in a ‘TRAVEL.md’ file, linked from my AGENTS.md).

So now it’s got all the details it needs, this is the ‘agenty’ bit, it’s gathered enough context to complete my task. With that context it determines the next thing to do: make the booking. It knows the tools it has available to it, like computer use, so it went on to the taxi’s website (I had my computer stay on all trip so I could connect), filled in the booking form and made the payment.

Working with agents is a lot about making sure the right context and tools are available to do the task. If both are covered, agents can get stuff done without you interfering.

We have to think about tasks in this way - what would an agent (or a person) need if they had no knowledge of you or your task? It’s your job to make sure it’s available.

Let’s get to it.

p.s. I wanted to shout out my friend’s first business he launched solo -Fixxa- voice quotes and invoices for UK tradespeople, with WhatsApp delivery and Stripe links.

Ben’s Bites is brought to you byAttio

Introducing Attio: the agentic CRM.With agents and automations that build pipeline, chase signals, and move deals forward,Attioorchestrates your revenue work around the clock. Loved by high-growth startups like Granola, Modal, and Wispr Flow.Start for free today.

Headlines

Fable 5 is backfor all paid users. Anthropic’s blog post mentioned stronger guardrails in this re-release, but I haven’t hit any of them yet. Fable 5 is only included in the subscription plans till July 7 and you can only use 50% of your limits on Fable.This benchmarkclaims Fable can do 16% of remote work projects, double the amount of Opus 4.8.
- Theo has quick tips ongetting the max out of Fablein the next 5-6 days.
Before Fable came back, we also gotClaude Sonnet 5- it’s benchmarked close to Opus 4.8 on most agent tasks, and cheaper on a per-token basis, but in practice it costs roughly the same as Opus per task. It is now the default for Free/Pro, available in Claude Code and API, and has launch pricing of $2/$10 per million tokens until Aug 31. Mine and others’ experiences with it is it’s expensive and slow, I can’t see it being something I use over other models.
Two new Gemini media models:Nano Banana 2 LiteandGemini Omni Flash. Both models are available in the Gemini app and API. In the API, Nano Banana 2 Lite gives you fast (under 4 seconds) and cheap images (~30 images at 1K resolution for a dollar). Omni Flash lets you generate and edit videos at $0.10/sec.
Bridgewater and Thinking Machinestrained a specialist model. It hit 84.7% accuracy on financial triage at 13.8x lower cost than the best frontier model tested.Factory’s Droid Shield 2.0also fine-tuned two detectors to catch exposed secrets in sessions and reduce false alarms.

Subscribe now

My feed

Modelence Mobile Builder- build native mobile apps by chat, using the same Modelence auth and backend. (portfolio company)
Browserbase Agents- one prompt or API call for browser automation in a hosted harness.
Safari MCP- let agents use Safari to open tabs, debug pages and catch Safari-only bugs.
Option AFK- Wispr-flow alternative with local transcription (nothing leaves your device), ability to transcribe multi-hour files and a CLI for your agents.
Claude Science- Mac/Linux workbench for scientists with code-backed figures, 60+ scientific skills/connectors and more. In beta for Pro, Max, Team and Enterprise.
Ramp PorTAL- move fine-tuned tasks between models at about half the usual cost.
xAI Voice Agent Builder- no-code Grok Voice agents at $0.05/min.
plain-writing-skill- rules for agents to write plainly, with an HTML diff of what changed.
Bond- AI chief of staff for founders and execs that tracks context, blockers and follow-ups.
Interfere- watches production, investigates issues and fixes problems before users notice.
AI-native PM leverage- how PMs move from text help to prototypes, PRs and evals. (course)
/wizard- skill to build interactive CLIs for annoying setup tasks.
Understanding agent-written code- a case for still knowing what the code does when agents write it.

Afters

@dylan522p@shaunmmaguirein today's Training Data episode. Nobody is a more trusted industry insider to the biggest infrastructure build-out in history. \\n\\nThe story of how Dylan became@SemiAnalysis_is even more awe inspiring: a young motel kid, \",\"username\":\"sonyatweetybird\",\"name\":\"Sonya Huang \",\"profile_image_url\":\"https://pbs.substack.com/profile_images/1468351747097694209/ZHz8EWnS_normal.jpg\",\"date\":\"2026-06-30T18:22:19.000Z\",\"photos\":[{\"img_url\":\"https://substackcdn.com/image/fetch/$s_!n-IS!,w_1028,c_limit,f_auto,q_auto:best,fl_progressive:steep/l_play_button_usfui2,w_88,e_colorize:0/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F__ss-rehost__tw-video-preview-13_2072020723321999360.jpg\",\"link_url\":\"https://t.co/EQiFiEYoG8\"}],\"quoted_tweet\":{},\"reply_count\":9,\"retweet_count\":21,\"like_count\":246,\"impression_count\":153358,\"expanded_url\":null,\"video_url\":\"https://video.twimg.com/amplify_video/2072020723321999360/vid/avc1/1280x720/xAV17Kn-QJzaYL3y.mp4\",\"belowTheFold\":true}\" data-component-name=\"Twitter2ToDOM\">@saranormous. \",\"username\":\"stanine\",\"name\":\"Matt MacInnis\",\"profile_image_url\":\"https://pbs.substack.com/profile_images/1249720502995705856/nBKDsWgZ_normal.jpg\",\"date\":\"2026-06-30T17:34:41.000Z\",\"photos\":[{\"img_url\":\"https://substackcdn.com/image/fetch/$s_!EdpP!,w_1028,c_limit,f_auto,q_auto:best,fl_progressive:steep/l_play_button_usfui2,w_88,e_colorize:0/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F__ss-rehost__tw-video-preview-13_2072010123204333568.jpg\",\"link_url\":\"https://t.co/mnHKSBv9fX\"}],\"quoted_tweet\":{},\"reply_count\":10,\"retweet_count\":29,\"like_count\":179,\"impression_count\":45358,\"expanded_url\":null,\"video_url\":\"https://video.twimg.com/amplify_video/2072010123204333568/vid/avc1/1280x720/nO9s3Q__t75NW3nt.mp4\",\"belowTheFold\":true}\" data-component-name=\"Twitter2ToDOM\">

Share Ben's Bites

Find me onX,Linkedin, orYouTube
Readabout meand Ben’s Bites
thumbnail via@keshavatearth

* sponsors who make this newsletter possible :)
Wanna partner with us for the next quarter?
Email us atshanice@bensbites.comork@bensbites.com

Ben's Bites is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

",[],"selectedArticleImage","rootArticleImageUrl","consensusHtml","tweets",[],"userSchema","relatedArticles",[259,287],"actionData","errors"]

Anthropic's Fable 5 Model Returns After Security Incident

Copertura della storia