The Manifest
Anthropic and OpenAI keep slashing AI prices and reshaping chat into interactive interfaces, while diesel costs and an AI-driven trade boom squeeze freight budgets, and new scrutiny lands on AI pricing bias and teen safety failures.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Claude Haiku 5.5 arrives with massive price cuts proving the AI pricing arms race is far from overThe Decoder
- 02ChatGPT with GPT-6 ditches mostly text output for interactive UI with charts, buttons, and mini appsThe Decoder
- 03DHL: Tech beats tariffs – the AI boom 'has become a powerful driver of global trade'The Loadstar
- 04Diesel's supply crunch may hit CFOs through freight contractsSupply Chain Dive
- 05Study: Claude, ChatGPT Offer Different Shopping Prices Based on Wealthbloomberg.com
Models & Releases
5 storiesClaude Haiku 5.5 arrives with massive price cuts proving the AI pricing arms race is far from over
Haiku 5.5 cuts token prices up to 90 percent and jumps from 15.7 to 72.4 percent on computer-use benchmarks, though a new tokenizer eats into some of the savings. Re-run your cost models before assuming the discount is as big as advertised.
ChatGPT with GPT-6 ditches mostly text output for interactive UI with charts, buttons, and mini apps
GPT-6's Intelligent UI turns answers into charts, buttons and mini apps and starts streaming responses while still thinking. For planners who live in dashboards, this narrows the gap between asking a question and acting on it.
Mistral says 'Le Chonk' can challenge the best AI models
Mistral is positioning an open-weight model against closed frontier systems again. If it holds up in third-party tests, procurement teams get another credible self-hosting option outside the US hyperscalers.
Liquid AI Releases Open-Weight d1-3B and d1-omni-600M: Multimodal Decision Models With Zero Output Tokens
These models skip text generation and return calibrated yes/no or category answers in one pass, built for real-time decisions rather than chat. That fits exception triage or fraud flags better than a general chatbot does.
Perplexity AI Releases pplx-embed-v2-late: A 0.6B Edge Model and a 9B Model Scoring 92.4% on MADQA
A 0.6B edge model plus a 9B indexing model, both MIT-licensed and self-hostable. Useful for teams building internal search over contracts, SOPs or supplier catalogs without sending data to a third party.
Supply Chain & Ops
5 storiesDiesel's supply crunch may hit CFOs through freight contracts
Chevron's CFO expects energy prices to stay elevated, and carriers that locked in rates before diesel climbed are already eating the difference. Anyone renewing freight contracts this quarter should build in a fuel surcharge reopener.
DHL: Tech beats tariffs – the AI boom 'has become a powerful driver of global trade'
DHL's Globalization Tracker says AI infrastructure buildout is now outweighing tariffs and geopolitics as a driver of trade growth. Expect capex tied to data center construction to keep showing up in freight and airfreight demand numbers.
Tractor Supply opens Idaho distribution center with automation, dedicated AI team
The retailer is standing up an in-house group to apply AI to operational problems rather than buying a platform off the shelf. That build-versus-buy signal is worth watching before deciding how to staff your own automation push.
Man gets 2 years for $3.5M Amazon fraud using 23 trucking accounts
Over a thousand Amazon Logistics assignments flowed through fake carrier accounts before anyone caught it. A reminder that carrier onboarding and invoice verification controls matter more than any routing algorithm.
Why Adding 3PLs Can Backfire on Shippers
More brokerage partners does not automatically mean lower rates, and managed transportation gets oversold as an AI fix when the real issue is often routing guide discipline. Worth a re-read before adding another 3PL to the mix.
Deals & Market
5 storiesNous Research confirms it hit $1.5B valuation, launches AI agents for business users
The Hermes Agent maker raised a $90 million Series B and is pushing straight into business agents. Another signal that agent tooling for enterprise ops is where the funding is chasing right now.
Zim lifts profit forecast 72% on strong demand
Rising freight rates and strong demand are flowing straight to the bottom line for the liner operator. Budget owners should expect less room to negotiate spot rates down through year end.
Radiant Logistics' latest deal expands its bi-modal broker platform
Radiant picked up Whitacre's Dallas intermodal brokerage to round out its road-and-rail network. Expect more small intermodal brokerages to get rolled up as larger platforms chase density.
Healthleap raises $38M for its AI that flags hospital patients who may need a closer look
Sequoia and First Round backed the seed, Hummingbird led the Series A. Healthcare ops keeps proving a reliable place to fund narrow, high-stakes AI tools rather than general chat products.
Zuckerberg's Biohub leads a $1.8 billion push to build AI models that predict cell behavior
Meta, Google DeepMind, Isomorphic Labs and the DOE are funding compute and lab equipment for a shared biology model effort. It is a capex story as much as a science story, and a sign of where infrastructure spend is shifting next.
Research & Frontier
5 storiesNVIDIA PivotOPD Teaches Multi-Turn AI Agents to Recover From Pivotal Mistakes
NVIDIA's training method helps agents notice and correct early mistakes instead of compounding them across a long task. Relevant for anyone piloting agents on multi-step procurement or planning workflows where one bad early call wrecks the whole run.
Meta AI Open-Sources Rebalancer: A C++ Assignment Solver That Runs About 40 Million Placement Problems a Day
Meta's internal tool for shard and server placement is now pip-installable under Apache 2.0, wrapping local search and MIP solvers like Gurobi and HiGHS. The same assignment logic applies to slotting inventory, labor or load assignments if you have the data to feed it.
Bounded Autonomy and Verifiable Safety for Agentic AI Enabled Automation
A framework for keeping agent reasoning inside expert-defined guardrails as tasks get more complex. Useful reading before anyone signs off on letting an agent make unsupervised calls in a regulated process.
FluidPD: In-Place Elasticity for SLO-Aware Prefill-Decode Disaggregated LLM Serving
A serving architecture that reallocates inference capacity on the fly between prompt processing and generation as traffic bursts. The kind of infrastructure work that quietly determines whether your AI vendor's latency and pricing hold up under load.
Just for FUNS: LLM-Guided Spatio-Temporal Graph Node Generation for Forecasting Unobserved Node States
Targets exactly the gap most supply chain forecasting hits: sensor and data coverage never matches the full network. Early-stage work, but the forecasting-unobserved-nodes problem is a real pain point for anyone planning around partial visibility.
Org & AI Architecture
4 storiesStudy: Claude, ChatGPT Offer Different Shopping Prices Based on Wealth
Researchers found AI shopping assistants quote different prices depending on signals about a user's wealth. If procurement or customer-facing teams use chatbots for quotes, this is a bias and trust problem worth auditing now, not later.
ChatGPT rated 'unacceptable risk' for teens after parental alerts failed during suicide conversations
Common Sense Media's audit found parental alerts never triggered during self-harm conversations across thousands of test prompts. For any org rolling out AI chat tools to vulnerable users, this is the kind of safety gap that only shows up under adversarial testing, not vendor demos.
Tony Fadell on why the first wave of AI gadgets failed — and what comes next
The iPod's co-creator argues first-gen AI hardware failed to solve real problems and burned user trust. Worth remembering before greenlighting another AI pilot that is a solution looking for a problem.
AI-powered hacking tools enabled a likely single attacker to breach multiple South Korean banks
CrowdStrike says an open-source AI penetration testing tool let one person breach several banks and steal more than 25,000 customer records. The barrier to serious attacks against financial and supply chain systems keeps dropping, and vendor risk reviews need to catch up.