We are not estimators. We build software, and we build it with AI models doing much of the typing. AI labs charge by the token, a chunk of text that is often part of a word, priced per million. This piece reads the fall 2026 model releases from that seat, then carries one lesson over to your quoting desk.
Most software teams keep a list of jobs nobody will ever get to: the small bug that annoys three customers, the report someone asked for in March, the old code everyone is afraid to touch. Those jobs are real work with real value. They sit on the list because starting each one costs more than it returns.
In the past year, that list started to shrink from the bottom, where the small jobs live. Jobs that were never worth starting became worth starting, because an AI could do the first pass for almost nothing and an engineer only had to check it.
We call the point where a job costs more to start than it can return the not-worth-it line. Most businesses have one. In a manufacturer's office, it runs straight through the RFQ inbox.
1. What shipped in September, in plain terms
A model is the AI itself: the part a lab such as Anthropic, OpenAI, or Google trains on huge amounts of text, code, and images. Products such as ChatGPT and Claude run on top of models. An agent is a model set up to take steps on its own, such as reading files or running tools, inside limits a person sets.
Six releases matter for this piece:
| Model | Maker | Released | What it is | Price per million tokens (in / out) |
|---|---|---|---|---|
| GPT-6 Astra | OpenAI | Sep 3 | OpenAI's "most capable model we have ever broadly deployed" | $10 / $50 |
| Jev | TypeSafe AI | Sep 15 (early access) | Returns a decision and a confidence score instead of text | $0.042 / free |
| Claude Opus 5.5 | Anthropic | Sep 22 | Performs "at the level of Claude Fable 5.1 on most work," Anthropic says (Fable is its top model), for 40% less than Opus 5 | $4 / $20 |
| Claude Sonnet 5.5 | Anthropic | Sep 28 | Within 2 points of Opus 5.5 on a test of office work, at half the price | $2 / $10 |
| GPT-6.1 Sol | OpenAI | Sep 29 | Close to Astra on coding and computer tasks, per TechCrunch, at about a fifth of the price | $2 / $10 |
| Gemini 4 Argon | Sep 30 (limited) | Google's newest top model, open only to selected security teams so far | $2 / $10 announced |
Sources: OpenAI's GPT-6 Astra system card, OpenAI pricing, Anthropic on Opus 5.5, Anthropic on Sonnet 5.5, TypeSafe on Jev, and Google on Gemini 4 Argon.
Read the table as two stories. The first is that top-tier work got cheap fast. Opus 5.5 matches Anthropic's top model on most work at a mid-tier price, and six days later Sonnet 5.5 came within 2 points of Opus 5.5 at half that price.
Source: Anthropic pricing.
The second story is Jev, a new kind of model. It does not write sentences. You give it a question with fixed answers ("quote it, ask a question, or decline") and it returns one answer with a probability, in 70 to 500 milliseconds by TypeSafe's count. TypeSafe and Vercel describe it for software that has to make thousands of small decisions: route this, score that, send the unsure ones to a person. Vercel, which runs a gateway developers use to reach many models, reported that nearly 13% of its paid teams were using Jev within 24 hours.
2. In software, cheap AI changed which jobs got done
How much faster AI makes engineers is still unsettled. METR, a research group that tests AI models, measured experienced developers in early 2025 and found their tasks took 19% longer with AI. When METR ran the study again in late 2025, it had to change its design because 30% to 50% of developers would not submit some tasks, since "they did not want to do them without AI."
That last finding is the useful one. Some developers no longer wanted to do certain jobs by hand at all.
Anthropic, which makes Claude, studied its own engineers and found that 27% of their AI-assisted work "wouldn't have been done otherwise." Separately, 8.6% of their tasks in Claude Code, Anthropic's coding agent, were small fixes that made daily work less annoying, the kind of job that sits at the bottom of every list for years.
The big version of the same pattern showed up in Google's Gemini 4 Argon announcement. Google says Argon agents are moving old code at Google into a newer, safer programming language, up to 800,000+ lines for one piece of its operating system. Google says the rewrites go through automated and manual review.
People did not leave the loop. Across about 400,000 sessions with Anthropic's coding agent, people made about 70% of the planning decisions and about 20% of the execution decisions. At Google, Sundar Pichai said 75% of new code is AI-generated and approved by engineers. The AI does the typing. A person decides what gets built and signs off on what ships.
3. Your RFQ inbox has the same line
Quoting has the same cost structure as that software backlog. Every RFQ costs estimator time before you know whether it will pay.
Fabricators take about 3 to 4 days to return a quote and win about 30% of them, according to Fabricators & Manufacturers Association (FMA) data reported by The Fabricator. FMA's benchmarking survey is small (28 US companies in its 2023 report), so treat these as rough. Even so, about 7 of every 10 quotes are estimator hours that never turn into an order. Machine shops that rank well in Modern Machine Shop's Top Shops survey quote in about one day and convert more of their quotes into orders.
The average hides a wide spread. One machining shop told Production Machining its hit rates by type of request:
Source: Production Machining, "Sometimes No-Quoting RFQs is Essential for Efficiency" (2020). One shop's numbers. The shop gave "in the 80% range" and "below 10%," shown here as 80 and 10.
At that shop, a new customer's RFQ wins less than 1 time in 10, while a repeat part wins almost every time. In our view the new customer's RFQ also takes more estimator time, because there is no past job to start from. In FMA's data, new customers brought fabricators just 3% of revenue in 2021. Estimators are also scarce: almost 1 in 3 US cost estimators is 55 or older, per the Bureau of Labor Statistics, and all of the roughly 17,600 openings a year are expected to come from people leaving the job.
So shops ration their estimators' time, and some RFQs fall below the line: the five-piece prototype, the one-off repair, the unfamiliar material, the buyer nobody has heard of, the request that arrives during the week the senior estimator is on vacation. Some get a polite no. Others sit in the inbox until the due date passes.
We could not find a published figure for how many RFQs manufacturers decline or let expire. Our guess is that most companies do not track it, because a quote nobody sent leaves no record behind.
The shaded area is the opportunity. When the cost of a first draft drops, the line moves right, and jobs that were never worth quoting become worth answering. Some RFQs stay below the line, and they should. The aim is to decide on purpose which ones those are.
4. How the new models split the work
The September releases fit the quoting problem in two places, and they are different kinds of AI.
The first is triage: deciding, for every RFQ that arrives, whether to quote it, who should own it, and how sure you are. A decision model like Jev is built for this. Here is the rough math, which is our own: an RFQ email plus your notes on that customer runs about 3,000 tokens. At $0.042 per million, deciding on 10,000 RFQs a year costs about $1.26 in model fees. The cost that matters is the setup and the review, which you should plan for. Jev also has stated limits that matter here: InfoQ reports it is unreliable at counting, arithmetic, and comparing dates. It should sort requests, and it should never set a price.
The second is the first pass on the quote itself. A large model such as Opus 5.5 or GPT-6.1 Sol reads the drawing and the email, finds your past jobs on similar parts, and drafts a price from your own costs, rates, and margins. Opus 5.5 can take in about a million tokens at once, enough for a full drawing package, the spec, and the email thread in a single request. It leaves what it is unsure about flagged for the estimator.
The estimator stays where the coding-agent studies put the engineer: on the decisions. They check the material and the routing, adjust the price for what they know about the customer, and press send. The AI has no permission to send anything to a buyer.
5. Who you can serve that you could not before
The customers below the line are usually small, new, or odd. In software terms, they are the small fixes and the old code nobody could justify touching. A few examples, varied across the industries we serve:
- A prototype buyer wants five machined brackets by next week and is shopping three suppliers.
- A rig operator needs a one-off repair quote on a part with no drawing, only photos and a worn sample.
- A sign customer sends a sketch and wants a ballpark for twenty locations.
- A circuit board startup asks for a small run with a material your estimator has not priced in a year.
Each of these RFQs can lead to repeat work, and each takes estimator time out of proportion to its size. With a first pass, the estimator reviews a draft instead of starting from a blank sheet, and in our view the shop can answer all four within a day. Our view is that some of those buyers become the large customers of 2028, because they were the ones nobody else answered.
Small buyers are worth more than they look. Protolabs, which serves prototype and low-volume buyers, earned about $11,000 per customer contact in 2025, and its quotes come from software, not an estimator's afternoon.
The buyer side is changing too. Xometry's marketplace revenue rose 30% in 2025, and its annual report names its main competition for buyers: "local manufacturers who may not be digitally enabled and do not provide online instant quote capabilities and lead times." Gartner predicts that by 2028, 90% of business buying will go through AI agents. Treat that as a forecast. We expect buyers to send more RFQs to more suppliers, and the shops that answer each one with a checked price to win more of the work.
6. The worries, answered
The first worry we hear is that the AI will price a job wrong. It will, sometimes, like a new estimator does. It drafts from your own cost data, flags what it is unsure about, and every change the estimator makes is logged so the drafts improve.
The second worry is security. OpenAI rated GPT-6 Astra "Critical" for cybersecurity, meaning it can find unknown security flaws and work out how to exploit them. That applies to attackers' tools as much as yours. The answer is the same rule we use for every AI role: its own login, read access only to what the job needs, and no access to anything else. Your pricing data stays yours: Anthropic and OpenAI both say they do not train their models on business API data by default (Anthropic, OpenAI).
The third worry is timing. If a better model ships next month, should you wait? Anthropic alone shipped two Opus releases in two months. New models will keep shipping, so waiting has no end point. Build around your own costs, past jobs, and review rules, and choose a setup where the model can be swapped. The model is the part that improves without you.
What to do this week
- Pull every RFQ from the last 90 days, including the ones you declined or let expire.
- Mark each declined or expired one with the reason: too small, too unfamiliar, no time, or not a fit.
- Count how many fell into "too small" and "no time." Those are the quotes you never sent, and they are the ones a first pass helps with.
- For one of those groups, list what an estimator needs to price it: which cost tables, which past jobs, which rules of thumb. That list is the start of a first-pass system.
- Decide who on your team would review and sign AI-drafted quotes, and write down what they would check every time.
For the common questions about AI quoting, from data to accuracy, read AI Quoting Questions, Answered. For how to set up any AI role with clear limits, read The New AI Org Chart.
Questions people ask
Which AI model is best for quoting?
The ranking changes every few weeks. A quoting system can use a fast decision model to sort RFQs and a larger model such as Claude Opus 5.5 or GPT-6.1 Sol to draft the price.
How much does AI cost per quote?
Model fees are small. At list prices, sorting an RFQ with a decision model like Jev costs a tiny fraction of a cent. By our math, an Opus 5.5 draft that reads 100,000 tokens of drawings and email and writes 5,000 tokens costs about 50 cents. Most of the real cost is building the system around your data and the estimator's review time.
Will AI send quotes without a person checking them?
Not in the systems we build. The AI drafts and flags. A named estimator reviews, adjusts, and sends every quote, and the AI has no permission to email a buyer.
Terms in this piece
- Model
- the AI itself, trained by a lab such as Anthropic, OpenAI, or Google. Products like ChatGPT and Claude run on top of models.
- Agent
- an AI model set up to take steps on its own, such as reading files or running tools, inside limits a person sets. Claude Code is a coding agent.
- Token
- a chunk of text, often part of a word. Labs price their models per million tokens.
- Decision model
- a model that returns a choice from a fixed list, with a probability, instead of writing text. Jev is one example.
- First pass
- the AI's draft of a piece of work, such as a quote, that a person reviews before anything leaves the building.
- Triage
- sorting incoming requests by what to do with them and who should handle them.
- Not-worth-it line
- the point where a job costs more to start than it is likely to return. Cheaper first drafts move the line.
