🚨 Fable 5.1 and Mythos 5.1 differ only in the safeguard layer — and Anthropic published what that layer costs

Almost nothing this week is a story about a model. A forecaster whose whole point is the series it was never allowed to read. A benchmark whose unit is model plus quantization plus runtime plus device. A query set that rebuilds itself hourly so nothing can memorize it. An environment that reshapes itself around the policy training on it. A research agent walled off from its own evaluator on purpose. And at the end, two models from Anthropic that are the same model — where the published gap between them is the cost of the layer wrapped around it. The model was always the easy part to name.

52.6%

Fable 5 scored 24.7%

78.4%

vs 33.8%, same size

575 ms

was ~1.6 s before

01 · Google Research · Forecasting

Google AI Releases TimesFM-3: A 330M Parameter Zero-Shot Foundation Model For Multivariate Time Series Forecasting

Every TimesFM checkpoint through 2.5 forecast one series from its own history. Real problems are not shaped that way. Google’s own example is ice cream sales, where foot traffic, weather, promotions and the neighboring product all move the number — and the model was not allowed to look at any of them.

TimesFM-3 is 330M parameters pretrained on more than 1 trillion time points, and takes multiple targets, past covariates and past-future covariates zero-shot. Alternating causal temporal and full variate attention read across series; Contiguous Patch Masking fills the entire horizon in one forward pass rather than decoding patch by patch, with 9 quantiles at every step. It takes the top average rank among pretrained foundation models on GIFT-Eval, fev-bench and TIME, on both point and probabilistic metrics. The catch is the license: the repo is Apache-2.0, the 3.0 weights ship under timesfm-non-commercial-license-v1.0. Benchmark it today. TimesFM 2.5 is still the one you ship.

Read on Marktechpost →  · Technical details  · Hugging Face  · GitHub

02 · Liquid AI · On-device

Liquid AI Open-Sources Pipette: A Reproducible Benchmarking Suite That Measures On-Device Models, Quantization, Runtime and Hardware Together

A model card reports quality under server-class, full-precision conditions. Nobody deploys that. Liquid AI’s answer is to stop benchmarking models at all. In Pipette the unit of measurement is a whole deployment configuration: model plus quantization plus runtime plus device.

The launch dataset covers five on-device metrics across more than 1,000 model × quantization × runtime × device × context configurations, 30+ models, and context lengths from 256 to 8,192 tokens, verified on a MacBook Pro M5 Max, an iPhone 17 Pro and a Galaxy S26 Ultra. The result that justifies the exercise: at Q4_K_M on the Galaxy, Granite-4.0-H-350M holds 78.4% of its decode throughput out to 4,096 tokens while Granite-4.0-350M holds 33.8%. Same parameter count, same phone, opposite conclusion. Apache 2.0 with Artificial Analysis as independent methodology validator — and one caveat stated in the docs: quality scores come from H100 reference runs matched to on-device performance, not measured on the phone.

Read on Marktechpost →  · Technical details  · Leaderboard  · GitHub

03 · Keenable AI · Search benchmarks

Keenable AI Open-Sources NEEDLE: A Live Search Benchmark That Rebuilds Its Query Set Every Hour

A search agent has a fetch tool. If the gold labels sit in a public dataset, the agent can download the answer key mid-evaluation and skip retrieval entirely. NEEDLE’s fix is to never hold still: news queries regenerate hourly from RSS feeds and Google Trends, and finance, scholar, legal and rare-entity queries regenerate daily from SEC XBRL, arXiv, Europe PMC, CourtListener and public agent logs.

Fifteen search APIs run the same query text under one protocol: one request at a time, a 2,000-character evidence cap, no page fetching, no re-ranking, and a judge that never sees the engine name. Every score is read against ultimate, a pooled oracle built from what all fifteen engines together managed to retrieve — which separates a ranking problem from a retrieval problem the whole market shares. Seven-day means ending 2026-08-28: finance is close to solved, Exa at 0.910 against a 0.965 ceiling; deep-tail, the vertical closest to real agent traffic, tops out at 0.557 of its ceiling. MIT licensed, runs execute in public GitHub Actions.

Read on Marktechpost →  · Live dashboard  · GitHub  · Artifacts

04 · Google Research · Agent training

Google AI Introduces EnvHarness: A Programmable Layer That Turns Static Agent Environments Into Adaptive Training Worlds

An agent environment behaves identically no matter who is acting on it or how much they have improved. Once a policy solves it, it has nothing left to teach. The usual fix is generating new environments, which buys you domain-specific pipelines and LLM-written verifiers nobody fully trusts. EnvHarness inverts the move: wrap the frozen environment instead of authoring a new one.

Three components — Stage, Contract and Chain — operate strictly through reset() and step(), so the simulator and its human-built verifier are never touched. An LLM designer called EnvRigger observes five baseline rollouts, diagnoses a systemic flaw, writes the wrappers as real Python, and validates on five fresh ones. Across five benchmarks in four domains, mined skills gain up to 9.0 points on ALFWorld’s out-of-distribution split, and SWE-bench Verified moves 49.88 to 52.58 while average steps fall 55.01 to 49.61. Apache-2.0 — the hard prerequisite is a resettable environment, which rules out live user accounts and physical robots.

Read on Marktechpost →  · Paper  · GitHub  · Project page

05 · Princeton, Ant Group, Stanford · Quant research

Researchers from Princeton, Ant Group and Stanford Introduce AQuA: A Two-Part Agentic Framework for Autonomous Factor Discovery and Model Development in Quantitative Finance

A research agent that writes its own experiments can poison the evidence it later learns from. A leaky feature that scores well gets filed as a successful precedent, and the next iteration builds on it. Prompt-level instructions and reviewer agents do not close this, because author and reviewer share the same blind spots. AQuA’s answer is to make the leaking action unavailable rather than discouraged.

Splits, feature and label definitions and the evaluator are all fixed before iteration one, and the agent emits only a constrained factor expression or a single config diff. The authors call it asymmetric freedom: the agent explores inside its DSL, the evaluator sits outside the adaptive surface, and what improves across runs is the research process. Two independent loops share no agents, memory or state. Part I reaches roughly 0.190 combined validation IC on a crypto five-minute universe against 0.171 for an adapted AlphaMemo; Part II reaches +0.0843 per-stock IC against +0.0613 for a GRU on identical data and the same evaluator. The equity book holds a +2.50 Sharpe at a 2 bps two-leg cost and stays positive every year from 2021 to 2025. The paper states plainly that the two parts use different IC conventions and should not be compared to each other.

Disclosure: the source article notes that Ant Research supported this content for promotion. The figures above are the paper’s.

Read on Marktechpost →  · Paper

06 · OpenClaw · Agent runtime

OpenClaw Releases OpenClaw 2.0: Guided Model Setup, 575 ms Control UI Startup, and One Trust Boundary Per Gateway

The OpenClaw team shipped 106 releases in 230 days, then shipped nothing for nearly seven weeks. What came back is not a model story. It is installation, storage, a browser app, and an honest sentence about where the trust boundary sits.

v2026.8.1 lands 16,000+ pull requests from 933 contributors, about half the project’s total merge history. Guided setup now reuses a verified Codex, ChatGPT or Claude CLI sign-in, an API key, or an installed Ollama or LM Studio model — and proves the exact choice can answer before saving it. In a simulated default-chat test with a mocked Gateway at 50 ms HTTP/1.1 latency, the rebuilt Control UI cut startup from roughly 1.6 s to 575 ms and JavaScript requests from 140 to 45. Shared cloud sessions add real multiplayer, and the docs state plainly that they are not tenant isolation and not a security boundary. Sessions moved to SQLite; take a verified backup first, because downgrades are not free.

Read on Marktechpost →  · Announcement  · Release notes  · Docs

One more thing…

Anthropic · Model releases

Anthropic Releases Claude Fable 5.1 and Claude Mythos 5.1: 52.6% on Terminal-Bench-Science and 75% Cheaper Cache Reads

Every story above moved the work off the model and into the layer around it. This one puts a number on that layer. Anthropic shipped two models that are the same model: Fable 5.1 and Mythos 5.1 differ only in the safeguards wrapped around them. Fable 5.1 is generally available as claude-fable-5-1; Mythos stays restricted to vetted organizations inside Project Glasswing.

On Terminal-Bench 4.0, Fable 5.1 reaches 55.8% and Mythos 5.1 reaches 60.9%. Five points, one model, and the difference is the cost of safeguard interventions — an unusually direct thing to publish. On Terminal-Bench-Science 0.1, Fable 5.1 takes 52.6% against 29.0% for Opus 5, 24.7% for Fable 5 and 22.4% for GPT-5.6 Sol, with a standard error of 3.5 to 4.5 points per model, so read the margin rather than the ranking. Elsewhere: CursorBench 3.2.0 at 73.4%, Humanity’s Last Exam at 60.9% without tools and 65.0% with, OSWorld 2.0 at 41.7% strict.

The commercial number is cache reads falling 75%, from $1.00 to $0.25 per million — 0.025 times base input, against 0.1 on every other Claude model. Base input and output are unchanged at $10 and $50. Anthropic measures roughly 25% lower cost on typical workloads and up to about 45% on context-heavy agentic ones. Read the breaking changes before you upgrade: tool_choice set to any or tool now returns a 400; thinking blocks are model-bound, so router fallbacks lose reasoning on the way down; and editing earlier turns invalidates thinking blocks, enforced for accounts created on or after August 31, 2026. Anthropic documents regressions too — parallel tool calling is more variable, and the model prefers whole-file rewrites over targeted edits.

Read on Marktechpost → · Announcement · Project Glasswing

Sponsor AI Dev Signals

Put your product in front of the people who build with it.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.