GPT-6 Astra can now operate your inbox and ERP. What that means for ship chandlers.

On 3 September 2026, OpenAI released GPT-6 Astra and called it a “generational leap” toward what it terms the AGI era. The claim worth scrutinising is not the marketing. It is one concrete capability: Astra is built for agentic work. Give it a goal, and it breaks the goal into steps and executes them across real software — operating a browser, filling forms, reading email, and working inside enterprise systems.

Ship chandling is one of the clearest places to test that. The core job — pricing a large RFQ that arrives as an email with a PDF or a messy Excel attachment — is exactly the kind of multi-step, tool-heavy task Astra was designed for.

Why the “agentic” part is the story

Previous models were strong at the language step: read a request, suggest a reply. The gap was everything around it — opening the attachment, checking the catalogue, applying margin rules, updating the order board. Astra is designed to carry a task through those steps in one session. On OSWorld 2.0, a benchmark that measures a model actually operating a computer, it scored 72.6%, taking roughly 40 minutes on tasks a person would also need real time to complete.

Two other numbers matter for procurement specifically. Astra handles a context window of about 1.05 million tokens and scored 96.3% on a long-context retrieval benchmark at the 512K to 1M range. In plain terms, it can hold an entire 800-line RFQ, your relevant catalogue slice, and a supplier price list in view at once without losing the thread. It is also built for long, multi-hour sessions and keeps notes across context windows, so a job that spans a morning does not reset halfway through.

Concrete applications in a chandling operation

• Reading the RFQ. Astra can open an email, parse a PDF or spreadsheet in any layout, and extract quantities, specs and codes, then hold all of it while it works down the list.
• Matching against IMPA and ISSA codes. Messy or non-standard descriptions (“2 x deck brush, stiff”) are where rule-based matching stalls. A stronger reasoning model raises first-pass match quality and flags genuinely ambiguous lines instead of guessing. See IMPA code matching.
• Drafting a margin-protected quote. With your per-category markup rules supplied as context, it can assemble a priced quote in your format and surface anything that would breach a margin floor.
• Chasing suppliers. For lines you do not stock, it can draft and track outbound price requests, then reconcile the replies back into the quote.
• Reconciling POs and invoices. Checking a multi-page supplier invoice against the PO, line by line, unit by unit, is tedious and error-prone. This is well suited to an agent that never skims.

The limits worth stating plainly

First, cost at scale. API pricing is 10 dollars per million input tokens and 50 dollars per million output tokens. A single long agentic session on a big RFQ, with a large catalogue in context, adds up. For repetitive, high-volume quoting, a purpose-built pipeline that reserves the expensive model for the hard lines is more economical than pointing a general agent at every email.

Second, oversight. Astra’s reasoning uses a new “recurrent depth” technique that is less transparent than earlier approaches. When it prices a thin-spec item or picks a supplier, you want a reviewable trail, not a black box. Keep a human sign-off on anything that leaves your building.

Third, it augments expert judgment rather than owning it. Knowing that a particular customer accepts a substitute brand, or that a port call is tight enough to justify a rush premium, is chandler knowledge. The model does not have your relationships.

Where Tidal fits

This is the workflow Tidal already runs for maritime distributors: reading email RFQs, matching every line to IMPA and ISSA codes, applying your margins, and returning a quote in under 30 minutes, with order tracking and invoice reconciliation behind it. The value of a model like Astra is not that you should wire a raw agent into your inbox tomorrow. It is that the underlying capability — reading messy procurement email and turning it into accurate, margin-safe pricing — keeps getting stronger, inside software built for the maritime edge cases a general model does not know. See how the RFQ-to-quote process works end to end, or the wider maritime procurement platform.