The Manifest
Claude Sonnet 5 arrives as a cheaper agentic model and Anthropic's Fable 5 returns from a two-week government ban; container spot rates hit multi-year highs as port congestion reaches a four-year peak; CBP's new forced labor enforcement rules put full sub-tier documentation requirements on every importer.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Container Shipping: Why Rates are Skyrocketing (It's NOT Demand)FreightWaves
- 02Anthropic launches Claude Sonnet 5 as a cheaper way to run agentsTechCrunch AI
- 03CBP's new 'forced labor' rules burden every link in the chainThe Loadstar
- 04Anthropic's Fable 5 is back worldwide after a two-week government ban over a jailbreakThe Decoder
- 05CMA CGM set to buy large chunk of FedEx Logistics to shore up CevaThe Loadstar
Models & Releases
5 storiesAnthropic launches Claude Sonnet 5 as a cheaper way to run agents
Sonnet 5 matches or exceeds Opus 4.8 on key benchmarks at Sonnet pricing. Any agentic workflow currently running on Opus should be tested against Sonnet 5 before the next billing cycle.
Anthropic's Fable 5 is back worldwide after a two-week government ban over a jailbreak
A new safety classifier blocks the jailbreak technique in over 99% of cases, with a modest increase in false positives on harmless requests. Organizations that paused Fable 5 deployments during the ban should review the classifier tradeoffs before re-enabling.
OpenAI paper reveals three GPT-5.6 Pro models, breaking with single top-tier strategy
Three Pro tiers instead of one signals OpenAI is moving toward task-complexity pricing. Teams budgeting for agentic applications should expect the cost model to shift again when 5.6 ships.
Google launches Nano Banana 2 Lite for fast AI images and Gemini Omni Flash for video via API
Four-second image generation at $0.034 per image and text-to-video via API opens generative media pipelines for product catalog and operations training content at meaningful scale. Chaining both is now a single API workflow.
Meituan's LongCat-2.0 shows China can train massive AI models without Nvidia
A 1.6 trillion parameter model trained entirely on domestic Chinese chips weakens the assumption that export controls will materially slow Chinese AI capability. Procurement teams evaluating AI tools from Chinese vendors should factor faster-than-expected model quality convergence into their assessments.
Supply Chain & Ops
5 storiesContainer Shipping: Why Rates are Skyrocketing (It's NOT Demand)
A 300% spot rate increase from March to June on flat demand reflects carriers using capacity management, not trade growth. Shippers with open H2 bookings on spot are exposed; those with locked contract rates should verify those rates are actually being honored.
Rate rises loom as port congestion hits four-year high
3.7 million TEU, or 11% of the global fleet, is waiting at berths, with North Asia accounting for 38% of that congestion. Planners should add transit buffer to any East Asia origin shipment scheduled for Q3 or Q4 arrival.
CBP's new 'forced labor' rules burden every link in the chain
CBP now requires importers to document every supplier in a production chain or face denied entry. Procurement teams that have not mapped beyond tier 2 are operating a live compliance risk, and this accelerates the business case for supply chain visibility investments.
Truck driver shortage now a 'structural threat' to businesses
Nearly 3 million vacancies globally with an aging workforce means carrier capacity constraints are structural, not cyclical. Network planners and procurement teams should model driver availability as a long-term constraint when setting service level commitments.
New BAFs another headache for shippers, but a 'digital twin' may ease the pain
July 1 bunker adjustment factors reflect Hormuz closure fuel costs from Q2 arriving all at once, on top of already-elevated spot rates. Shippers without fuel surcharge visibility in their TMS may not see the full landed cost impact until invoices arrive.
Deals & Market
5 storiesCMA CGM set to buy large chunk of FedEx Logistics to shore up Ceva
Acquiring North American 3PL assets from FedEx would give Ceva the scale to compete with DHL and Kuehne+Nagel in the US. Shippers relying on Ceva should track how the ownership change affects service commitments and contract pricing.
MSC's terminal arm TiL takes 49% stake in Vizhinjam port (total deal value = $2.85bn)
MSC deepening ownership of deep-water port infrastructure on India's west coast positions it to capture transshipment volumes as India trade grows. Buyers sourcing from South Asia should track how this shapes carrier routing options over the next two to three years.
Prologis presses Segro after $16.6B offer rejected
A Prologis-Segro combination would consolidate logistics real estate leadership across North America and Europe. Tenants in markets where both operate should assess renewal terms before a combined entity gains further pricing leverage.
Nvidia competitor Etched hits $5B valuation, $1B in sales for AI chip
$1 billion in contracted inference chip sales is the strongest evidence yet that enterprise demand for non-Nvidia inference hardware is real. Supply chain teams evaluating AI infrastructure should add delivery risk assumptions alongside the usual performance comparisons.
Wayve launches $85M employee tender offer at $8.5B valuation
Wayve's valuation holds at $8.5 billion despite a tighter funding environment for autonomous vehicle software. The employee tender format doubles as a retention tool in a market where AI engineers have many competing offers.
Research & Frontier
5 storiesGoogle AI Introduces TabFM: A Hybrid-Attention Tabular Foundation Model for Zero-Shot Classification and Regression
Zero-shot tabular classification with no feature engineering or per-dataset training has direct relevance for demand sensing, inventory risk scoring, and supplier classification. Worth comparing against existing AutoML and gradient boosting baselines on internal supply chain datasets.
NVIDIA Releases Nemotron-Labs-TwoTower: an Open-Weight Diffusion Language Model Built on a Frozen Autoregressive Backbone
A diffusion LM built on a pretrained autoregressive backbone addresses the serial token-by-token throughput bottleneck in text generation. Higher throughput matters for high-volume operations workflows where latency or token budgets constrain agent performance.
What Drives Interactive Improvement from Feedback?
Multi-turn accuracy gains in AI agents can come from resampling or format correction rather than the feedback itself. Teams building human-in-the-loop correction into agent workflows should measure whether feedback is actually driving improvement or just adding compute.
When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models
Learned early exits in reasoning models reduce unnecessary compute at fixed budget checkpoints. For operations teams paying per token on high-volume classification or triage tasks, this research tracks directly to inference cost management.
Meta AI Releases Brain2Qwerty v2: A Non-Invasive MEG Brain-to-Text Pipeline Decoding Typed Sentences at 61% Word Accuracy
61% word accuracy on a non-invasive brain-computer interface is a meaningful research milestone, not a near-term operations tool. The open training code release accelerates the research timeline, and warehouse accessibility applications are worth tracking over a five-year horizon.
Org & AI Architecture
5 storiesAI & Logistics: Will a 'Job Apocalypse' Hit the Supply Chain?
Trimble's chief platform officer argues AI's near-term impact in logistics is unlocking value in legacy data and systems more than eliminating roles. The more useful frame for managers is identifying which decision-support roles shift toward AI oversight and which manual workflows get automated first.
Amazon launches new $1 billion FDE org, following OpenAI and Anthropic
Amazon embedding engineers inside enterprises to deploy purpose-built agents is a direct play for platform stickiness. Supply chain teams evaluating AWS AI tools should assess what the FDE engagement model requires in terms of data access and long-term platform commitment.
New attack provides one more reason why AI browsers are a bad idea
Asserting false context, such as that a basic math fact is wrong, is enough to override AI browser safety guardrails. Any agentic procurement or supplier research workflow that browses external pages should treat prompt injection from adversarial sites as a live operational risk.
San Francisco's AI boom is pricing out six-figure tech workers who can't find rent under $5,000
Median SF rent at $3,827 with homes averaging $1.7 million creates compounding retention and recruiting costs for AI teams headquartered there. Supply chain technology organizations outside San Francisco have a structural cost advantage when competing for technical talent.
Exclusive: No More Side Hustles: Why AI Startup Omnea Will Give Employees $250K To Openly Plan Their Next Startup
A London-based AI procurement spend management company funding employees to launch their own startups after five years of service is an unusual retention and talent development model. Worth watching as a signal of how AI-adjacent procurement companies compete for engineering talent.