← Back to Insights

Insight

Two Days, Then Forever

Ariel Agor
Updated
The cover illustration with subtle ambient motion, for "Two Days, Then Forever".

Key facts

FactValueSourceAs of
Budget Replit's Missions agents used to build the DataCat monitoring platform$3,000replit.com
Time DataCat took to alert on a planted database bug38 secondsreplit.com
MAI Code 2 average cost per task across Microsoft's benchmarks (promotional pricing)$0.309unite.ai
Claude Opus 5.5 average cost per task in the same Microsoft comparison$6.306unite.ai
MAI Code 2 promotional input price in GitHub Copilot$0.50 per million input tokenseesel.ai

Listen · Read by Ariel

0:00 / n/a

On November 17, 2026, two companies published build decisions within hours of each other. Replit launched Missions, its agent swarm product. The main exhibit in the launch post was a monitoring platform called DataCat, which the agents built for Replit's own use in two days. The same day at Ignite, Microsoft released MAI Code 2 and MAI Code 2 Flash. These are coding models it trained itself, and it put them into GitHub Copilot next to the frontier models it already offers there. In the first case a software company replaced a vendor tool. In the second, the largest software company in the world decided it would rather own a model than pay OpenAI and Anthropic by the call for the same work. If you are deciding when to build vs buy AI this quarter, the two announcements make the same point from opposite ends. Building has become cheap and owning has not, and every sensible answer to the build or buy question now starts from that gap.

A monitoring platform for the price of a used car

Here is what Replit says happened. Its team gave Missions a goal it describes as "bringing application monitoring in-house to reduce costs and gain more control." The post gives the result in one sentence: "In two days, within a $3,000 budget, a swarm of agents built dashboards, alerts, and application tracing."

Replit then tested the thing hard. It planted a database bug on purpose, and "DataCat raised an alert in 38 seconds, beating our 60-second target." It ran three hours of sustained load with 26 million activity records and 4.4 million log entries, and reported no data loss. DataCat passed 51 of 51 data-accuracy and alert-reliability checks. On Postgres it took in 49,000 traces per second. Replit also had Claude Code build its own version of the same platform for comparison, and that version managed 9,800.

The vendor being replaced is never named. That is a polite omission, but anyone who has paid an observability bill can guess the category. Monitoring is one of those line items that grows every year in ways nobody planned, because it is priced on data volume and data volume only goes up.

Replit was careful in one place, and the care tells you something. It says its controlled tests "do not establish performance against the original platform's hosted service." So Replit claims to have built something that works. It does not claim to have built something better than what it was paying for, and it says nothing about what DataCat will cost to run in its second year.

That is the honest version of a story that is about to be told many times, badly. Over the next few quarters a lot of executives will see a $3,000 figure and conclude that their SaaS contracts are optional. Some of them will be right. Most will find out what the $3,000 left out.

The build estimate stopped deciding anything

For thirty years the build or buy decision was really a decision about the build estimate. Someone in engineering said a thing would take two quarters and four people. Someone in finance compared that with the vendor's three-year quote. The build usually lost, because engineers estimate badly, opportunity cost is real, and the vendor could show you the product working that afternoon.

The method had a flaw nobody minded much, because the flaw was dwarfed by the build cost. It ignored most of the life of the software. A system that took six months to build would then be run, patched, secured, upgraded, migrated and explained to new hires for six years or more. Everyone knew this, and nobody priced it carefully, because by the time the build estimate was settled the decision was usually made.

Replit's post shows the build estimate shrinking toward nothing. Two days and $3,000 would not cover one week of a single senior engineer's loaded cost in most American cities. Once the build costs that little, it stops being the number that decides anything, and the weight moves to the part the old method ignored: the forever.

Forever has several parts, and none of them shrank on November 17. Someone has to be on call when DataCat itself goes down at 3 a.m., which is a strange kind of outage because the thing that would have told you something was wrong is the thing that broke. Someone has to keep up with security disclosures for every library the agents chose. Someone has to decide what happens when a product team wants a feature that the hosted vendor shipped last spring. Someone has to know, a year from now, why the agents chose ClickHouse for one storage engine and Postgres for the other, because that person will be asked when the bill for one of them jumps.

A vendor spreads all of that across thousands of customers. That was always what you paid for. The software was a byproduct of the vendor's real product, which was its willingness to own forever on your behalf.

What agents change about forever

The optimistic answer is that the same agents that built DataCat can maintain it, and the answer is partly true. Missions works to an agreement with a spending cap and a list of actions that need human approval, and an independent AI judge checks the work against your success criteria. If maintenance can be written as missions, with budgets and acceptance tests, then forever gets cheaper too.

It gets cheaper, though, without getting simpler. Someone still has to write the success criteria, notice when they no longer describe what the business needs, and sign off on the approvals. The agents take over the labour of ownership. They do not take over the responsibility, and responsibility is the part that runs for years.

What Microsoft actually built

The Microsoft announcement is the more useful one for a board, because Microsoft had every reason to keep buying.

Microsoft's own figures show MAI Code 2 reaching "98% of GPT-6.1 Sol's performance at 40% of its cost, and 96% of Claude Opus 5.5's performance at 5% of its cost," according to Unite.ai's write-up. Averaged across four coding benchmarks, MAI Code 2 passed 63.2% of tasks at $0.309 per task. GPT-6.1 Sol passed 64.1% at $0.761. Claude Opus 5.5 passed 65.3% at $6.306.

Those three numbers explain the whole decision. The pass rates differ by about two percentage points. The costs differ by a factor of twenty. For a company running coding tasks through Copilot at Microsoft's volume, two points of accuracy are not worth twenty times the inference bill. Microsoft built because it had the volume, and volume turns a small per-task saving into a line item large enough to justify a model team.

Two caveats belong next to those figures, and Microsoft supplied both. The cost numbers use a 50% promotional price. Eesel's breakdown of the pricing puts MAI Code 2 at $0.50 per million input tokens and $2.50 per million output tokens until May 17, 2027, after which list prices double to $1.00 and $5.00. That moves the cost per task to about $0.62, which is still well under a tenth of the Opus figure. Microsoft also says both models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work."

Read that limitation as a design choice rather than a weakness. Microsoft did not try to build a frontier model. It built a model for the large middle of everyday coding work, where the tasks are frequent, well defined and gradable, and it left the hard tail to the vendors.

The narrow job is where building wins

The clearest evidence is in the Microsoft 365 numbers. Microsoft tuned Flash further into a variant called MAI Work Flash and tested it on generating DAX queries, the query language behind Power BI's Chat with your Data feature. On Microsoft's offline scoring, MAI Work Flash scored 0.77. GPT-6 Astra scored 0.54 and Opus 5.5 scored 0.30.

A small in-house model beat a frontier model by a wide margin on one narrow task that Microsoft runs millions of times and controls completely. Microsoft owns the query language, the product and the telemetry, so it knows what a right answer looks like. A general model has to be good at everything and was never trained to be excellent at DAX.

That gives you most of a decision rule. You should build when the task is narrow, frequent and gradable and when you hold the data that defines a correct answer. Everything else you should buy, and switch vendors whenever someone ships a better model.

AT&T found a third door

Eesel's summary of the model page includes a detail that most coverage skipped. AT&T has a custom version of Flash that beat GPT-6.1 Sol on standard 5G network analysis at a much lower cost. AT&T did not train a model and did not buy the general one off the shelf. It tuned weights it licensed through a Microsoft program called MAI Licensed Weights.

This matters more than the headline benchmarks. The old choice had two options: write it yourself, or pay someone who did. AT&T took a third, in which a vendor keeps the base model current and AT&T owns the specialization layer on top, the network data and the definition of a good answer. Forever is split between the two. Microsoft handles the base and AT&T handles what makes the model its own. For most companies with a high-volume domain task, this middle door will be the right one, and few of them know it exists.

When to build vs buy AI, asked as a question about forever

Executives keep asking the old question: can we build this? In November 2026 the answer is almost always yes, and cheaply. So the useful questions are different, and there are three.

The first is whether you will still want to own this thing in year three. Replit wanted monitoring "in-house to reduce costs and gain more control." Control is a good reason when the system sits close to your product and you want to change it on your own schedule. It is a poor reason when the system is generic plumbing that a vendor improves every month for free.

The second is whether your volume makes cost per task the dominant number. Microsoft's case only works because Copilot is enormous. Say your company runs 2,000 coding tasks a month. On Microsoft's published promotional figures, the gap between $0.309 and $6.306 a task comes to about $12,000 a month, which is real money. Suppose instead you run 200 a month: the gap is about $1,200, which will not pay for anyone to think about model selection. At low volume, buy the best model and stop optimizing.

The third is whether the thing is your wedge. Twilio built an internal AI sales-coaching tool called Jarvis. Asked about outside products like it, chief executive Khozema Shipchandler said in reporting relayed by Rews on September 30, 2026, that Twilio "probably won't use it, because we've already built our own." The same piece cites a McKinsey survey of 1,719 executives and managers in which about a third of organizations had skipped at least one software purchase because AI let them build the product instead. It also reports that West Monroe, a Chicago consultancy, built eight internal tools in place of vendor offerings and put the avoided vendor cost at more than $300,000 a year.

Twilio's sentence is the one to study. Once you build, you stop being a prospect for that category, permanently. That is a fine outcome when the tool encodes how your sales team wins and you want it to keep diverging from what competitors can buy. It is a bad outcome when you have shut yourself out of a category that vendors are about to improve faster than you can.

The forever test, made concrete

For illustration, here is how I would run the numbers on something like DataCat in a mid-sized company. Say the agent build costs $3,000, as Replit's did. Say keeping it alive takes a fifth of one engineer's time for patching, upgrades, on-call and feature requests, at a loaded cost of $200,000 a year. That is $40,000 a year of forever. Suppose agent-run maintenance missions halve the labour and add $5,000 a year in agent spend. Forever then costs about $25,000 a year.

Now compare that with the vendor invoice. If the hosted platform costs you $150,000 a year and grows with your data, building wins easily, and the $3,000 is a rounding error in either direction. If it costs $30,000, building saves you almost nothing, and you have taken on a 3 a.m. pager for the privilege. Neither verdict depends on the build cost. Both depend on the forever line, which most build or buy memos still leave blank.

Where the money will actually leak

Three failure patterns are already visible, and each comes from pricing the build while ignoring what follows it.

The first is the orphaned tool. A product manager uses an agent platform to replace a $40,000 SaaS contract over a weekend, then moves teams in March. The tool keeps running. Nobody owns it. The first sign of trouble is a security review in October that finds a dependency with a known vulnerability and a database nobody can identify. The build was a success. The ownership was never assigned.

The second is the model-selection tax paid at low volume. A team reads the MAI Code 2 numbers, builds a routing layer to send cheap tasks to cheap models, and spends a quarter maintaining evals for a workload that was never big enough to matter. Microsoft can afford a model team because twenty times cheaper at its scale buys a lot of people. At small scale, the routing layer is a hobby with a payroll.

The third is subtler. A company builds where it should have taken the third door. It trains or fine-tunes from scratch on its domain data and then watches the base model it started from go stale while the vendor's licensed weights get better every quarter. AT&T avoided this by owning the specialization and leaving the base to Microsoft. Owning a base model means owning its obsolescence, and that obsolescence comes on a schedule the frontier labs set.

The renewal is now a negotiation you can win

There is a quieter consequence of the Replit post that every CFO should act on before the next renewal cycle. Vendor pricing in categories like monitoring rested partly on the cost and slowness of the alternative. That alternative now takes two days and costs $3,000, and Replit has open-sourced a self-hostable version of DataCat on GitHub.

You may never build your own. You should still be able to show your vendor a credible build plan with a forever line on it. A buyer who has priced the forever properly negotiates differently from one who has only heard that building is possible. The cheap build is most valuable as a credible threat, even in the many cases where you never carry it out.

The architecture is the decision

Everything above points to one conclusion that a tool purchase cannot deliver. When to build vs buy AI is no longer a decision you make once per category and file away. It is a standing policy that needs an owner, because both sides of the comparison move every month. Microsoft's promotional pricing changes on May 17, 2027. New models will land in Copilot and Foundry before then. Agent platforms will get better at maintenance, which will shrink the forever line for some systems and leave it unchanged for others. The vendors you keep will reprice when they notice you have alternatives.

A company that handles this well has three things. It has an inventory of AI-touched work sorted by volume, so it knows which tasks are frequent enough for cost per task to matter. It has a forever ledger for everything it builds, with a named owner, a budget and a retirement date for each system. And it has a routine for looking at the third door, licensed and tuned weights, before anyone commits to training from scratch or to a three-year contract.

None of that can be bought. Microsoft and Replit sell excellent components, and each published honest caveats this week about what those components do not prove. Putting the components together into a policy your company can live with for years is architecture work, and it needs someone who has no product to sell you and has seen the orphaned tools and the stale fine-tunes before.

That is the work Agor AI Advisory does. We map your workloads by volume and by how they are graded. We price the forever line before you commit to a build. We find where licensed weights beat both building and buying. And we design the ownership model so that the system you build in two days is still owned by someone in year four. The companies that set this policy now will renegotiate every renewal from strength. The ones that don't will learn what the $3,000 left out, one 3 a.m. outage at a time.

Sources

Two points of accuracy, twenty times the bill

This verifies the post's claim that the pass rates differ by about two percentage points while the costs differ by a factor of twenty. In 15 seconds the reader sees that the gap between models is large on cost and small on accuracy, so cost per task dominates at high volume.

  • Two points of accuracy cost a factor of twenty.
  • At Microsoft's volume, cost per task is the number that decides. At low volume, buy the best model and stop optimizing.
  • MAI Code 2's cost reflects a 50% promotional price. At list price it is about $0.62 per task, still under a tenth of Opus.
Tasks passedAvg cost per taskCost vs MAI Code 2
MAI Code 2Lowest pass rate of the three, at the lowest cost. Uses a 50% promotional price, and Microsoft says it is not yet tuned for the hardest long-horizon work.63.2%$0.3091x (baseline)
GPT-6.1 SolPasses about one more task in a hundred than MAI Code 2 for roughly two and a half times the cost.64.1%$0.761about 2.5x
Claude Opus 5.5Highest pass rate, about two points above MAI Code 2, at about twenty times the cost per task. This is the premium that only pays off where the hard tail matters.65.3%$6.306about 20x

Source: Microsoft's MAI Code 2 benchmark figures averaged across four coding benchmarks, as reported by Unite.ai (https://www.unite.ai/microsoft-ai-releases-mai-code-2-and-mai-code-2-flash-for-enterprises/). The cost multiples are computed from the reported per-task costs, using MAI Code 2's promotional price. · verified · as of 2026-11-17

Want this kind of automation working for your business?

Agor AI designs and ships the systems these posts describe, scoped in weeks, not quarters.

Book a Free Strategy Call