What can you build with GPT-6 Astra that people will pay for?

What did GPT-6 Astra actually change for builders?

GPT-6 Astra changes the opportunity for builders mainly because it got much better at doing work inside real software.

When we compare OpenAI's evaluations with GPT-5.6 Sol, the improvement is surprisingly uneven. BrowseComp moves only from 90.4% to 91.5%, and DeepSWE from 72.7% to 74.1%. Those are small gains. Astra does not suddenly create a huge new opening for another research chatbot or coding autocomplete tool.

The bigger changes appear when the model has to operate tools and work through a process. AutomationBench jumps from 18.1% to 41.4%. OpenAI's internal database-migration benchmark goes from 42.7% to 63.9%. Terminal-Bench rises from 37.3% to 57.9%, while ScreenSpot-Pro climbs from 76.9% to 92.7%. On SRE-Bench, Astra reaches 88%, compared with 55.9% for GPT-5.6 Sol.

OpenAI also measured how quickly the models could complete computer-use tasks. On OSWorld 2.0, Astra scored 72.6% while taking roughly 40 minutes per task in OpenAI's latency simulation. GPT-5.6 Sol scored 65.7% and took around 75 minutes. Astra was therefore completing these tasks in about 47% less time.

That combination is where Astra starts to look different. Astra can navigate interfaces, use terminals, modify software, investigate failures, fill forms, work with business files and keep going through longer chains of actions. OpenAI even uses examples such as updating CRM records, installing software and running frontend QA in its launch material.

As of now, Astra is still rolling out gradually across ChatGPT, Work, Codex and the API. The exact product availability is moving quickly. But we already have enough evidence to identify where the model is unusually strong: work that involves several tools, several decisions and a finished outcome.

Evaluation	GPT-5.6 Sol	GPT-6 Astra	Improvement
BrowseComp	90.4%	91.5%	+1.1 pts
AutomationBench	18.1%	41.4%	+23.3 pts
Database migration tasks	42.7%	63.9%	+21.2 pts
Terminal-Bench 4.0	37.3%	57.9%	+20.6 pts
ScreenSpot-Pro	76.9%	92.7%	+15.8 pts
SRE-Bench	55.9%	88.0%	+32.1 pts

Can you trust GPT-6 Astra to finish real work today?

GPT-6 Astra is ready for workflows with clear limits and checks, but giving it an open-ended business process with no review is still too risky.

The model's computer-use results are good enough to change what we can build, yet they are nowhere near perfect. Astra scores 72.6% on OSWorld 2.0 and 59.3% on Agents' Last Exam, which tests complicated professional tasks inside real software. Even AutomationBench, where Astra more than doubles GPT-5.6 Sol's score, still ends at 41.4%.

The practical takeaway is simple. A system that succeeds on 70% of difficult computer tasks can save a huge amount of labor if failures are easy to detect. It becomes dangerous when one unnoticed mistake can corrupt a database, send money, submit the wrong legal statement or alter thousands of records.

The better product design is to make Astra handle the repetitive 90% and bring a person in for the strange 10%. A CRM migration agent, for example, can inspect the old schema, suggest mappings, move a test batch, compare record counts and show the user anything ambiguous before continuing. The customer could avoid hundreds of manual decisions without giving the agent unlimited freedom.

OpenAI's own safety results point in the same direction. In one internal test built around agents going beyond their authorized scope, GPT-5.6 Sol did so 48% of the time without production safeguards. Astra scored 0% in that evaluation. OpenAI still uses additional monitoring and confirmation systems around agentic deployments.

So we can already sell work done by Astra. We just need to design the job so mistakes become exceptions that are caught, rather than invisible failures that reach the customer.

Why would anyone pay for an Astra product if ChatGPT Work already has GPT-6 Astra?

Astra wrappers are in a much tougher position now because ChatGPT Work can already perform many of the generic computer tasks that used to justify a standalone AI app.

ChatGPT Work currently has its own cloud browser. It can read pages, click buttons, enter information into forms and work on supported websites where the user is signed in. Work can also run longer tasks, use connected apps and plugins, produce finished deliverables and pause when it needs an approval or more information.

Codex creates the same problem on the software side. Astra is being distributed directly inside OpenAI's coding product, with browser and computer-use capabilities available there as well.

That wipes out a lot of ideas that sounded good a year ago. “Research these companies and make me a deck,” “fill these web forms,” “update this spreadsheet,” “test my website,” and “build an app from this description” are all becoming things a customer can ask a general-purpose OpenAI product to do.

We therefore need more than a better interface around Astra.

A company buying a migration product wants its Salesforce data correctly moved into HubSpot, with the fields mapped, records checked, workflows recreated and failures explained. A finance team wants its monthly close prepared. A sales team wants its security questionnaire submitted. A product team wants to know whether a release is safe to ship.

Those jobs create room for a product because we can connect Astra permanently to the right systems, remember previous decisions, enforce approvals and measure whether the work was actually completed correctly.

The moat also becomes much clearer. If we have handled 20,000 migrations, our mapping rules, failure history and validation tests become valuable. If we process security questionnaires every day, we accumulate approved answers and know where submissions usually break. A competitor calling the same Astra API still has to reproduce all of that.

The safest assumption is that OpenAI will keep making the general agent better. We should build something that becomes more useful as Astra improves rather than something OpenAI can erase by adding one button to ChatGPT.

Is GPT-6 Astra cheap enough to build a profitable SaaS?

GPT-6 Astra is cheap enough for high-value SaaS workflows because even substantial model runs cost very little compared with an hour of professional labor.

The standard API price is currently $10 per million input tokens and $50 per million output tokens. That is expensive compared with smaller models, so sending every tiny request through Astra would be wasteful.

In absolute terms, however, the cost remains surprisingly manageable. A task using 50,000 input tokens and 5,000 output tokens costs about $0.75 in raw Astra tokens. A much heavier workflow using 200,000 input tokens and 20,000 output tokens costs around $3.

The economics become less friendly with huge contexts. Astra has a 1.05-million-token context window, but prompts above 272,000 input tokens move to higher rates: input becomes twice as expensive and output 1.5 times as expensive for the whole request. A 500,000-input, 30,000-output run works out to roughly $12.25 before tool charges.

We would therefore route the work. A cheaper model can classify files, extract simple fields and handle routine messages. Astra can take over when the task requires planning, computer use, debugging or judgment. Stable company context can also be cached instead of repeatedly sent at full price.

That architecture makes a lot more sense for a product that charges $200, $500 or $2,000 for useful work than for an unlimited $19 chatbot.

Illustrative Astra run	Input / output	Approx. raw token cost	Best fit
Small difficult task	50K / 5K	$0.75	Normal SaaS feature
Substantial workflow	200K / 20K	$3.00	Professional automation
Heavy pre-threshold workflow	250K / 50K	$5.00	High-value agent
Long-context workflow	500K / 30K	~$12.25	Expensive business outcome
Very large context	1M / 50K	~$23.75	High-ticket workflow

Is software migration the best GPT-6 Astra business right now?

Software migration is probably the strongest Astra business we found because customers already spend thousands of dollars on a job that closely matches the model's biggest improvements.

OpenAI's database-migration evaluation is particularly relevant here. Astra reaches 63.9%, compared with 42.7% for GPT-5.6 Sol. At the same time, its gains in terminal work and computer use give it more ways to handle the messy parts around a migration.

There is also obvious willingness to pay.

HubSpot currently charges a required $3,000 onboarding fee for Marketing Hub Professional and $7,000 for Enterprise. Those fees sit on top of the subscription price.

The HubSpot marketplace shows another piece of the market. Datawarehouse.io's HubSpot-to-HubSpot migration application starts at $599 and reaches $1,499 for its largest package. The marketplace currently shows roughly 5,000 installs for that application.

Astra lets us imagine a product that goes beyond copying data. The user connects the old CRM and the new one. The agent inspects both schemas, suggests how fields should map, finds incompatible objects, creates a test migration, checks the results and asks about ambiguous records. It can then move the rest, recreate selected settings and give the customer a final report showing anything that did not transfer cleanly.

We could start extremely narrow. “Move Pipedrive to HubSpot” is easier to sell and evaluate than “AI implementation consultant.” Once the workflow works, the same engine can expand into Salesforce, Zendesk, Intercom, Shopify, analytics tools or other business systems.

The outcome is also easy to understand. Customers can check whether their contacts, deals, fields and workflows are there. That makes a migration agent much easier to trust than an autonomous agent whose success is subjective.

Existing spend	Current public price	What an Astra product could handle
HubSpot Marketing Pro onboarding	$3,000	Setup, configuration and migration
HubSpot Marketing Enterprise onboarding	$7,000	More complex implementation
HubSpot portal migration software	$599–$1,499	Migration plus validation
Human implementation work	Often additional	Exceptions, cleanup and QA

Can GPT-6 Astra take over RFPs and security questionnaires?

GPT-6 Astra could turn RFP and security-questionnaire automation into a much more complete product because the model can now work through the actual submission process.

The willingness to pay is already unusually clear. Loopio currently starts at $20,000 per year. Responsive starts at $10,000. Conveyor's Business plan starts at $9,600 per year and includes questionnaire automation and RFP projects.

Those products prove that companies already spend five figures solving this problem. We do not need to invent demand.

What remains frustrating is that the work often crosses several formats and systems. A questionnaire might arrive as an Excel file, a Word document, a PDF or a procurement portal containing hundreds of fields. Answers may depend on the company's security policies, previous responses, product documentation, legal language and several employees.

Astra can potentially connect these steps. We could give the system the company's approved answer library and supporting documents. When a new questionnaire arrives, the agent identifies each question, finds the best supported response, highlights anything that needs human approval and handles the surrounding file or browser work.

The useful product ends with a completed submission rather than a page containing 300 generated answers.

It also gives us a clean place for human review. If Astra knows that 270 answers closely match approved material while 30 introduce new commitments, the security or legal team can spend its time on those 30.

For a SaaS company where security review regularly sits between a signed buyer and a closed deal, that is a very easy value proposition to understand.

Can GPT-6 Astra make compliance evidence collection a real business?

Compliance evidence collection is already a huge software market, and Astra could make the tedious operational layer much more automatic.

Vanta gives us the clearest indication of how much companies spend here. Earlier this year, Vanta said it had passed $300 million in ARR and 16,000 customers. The journey from $200 million to $300 million took only nine months, and Vanta said its growth rate had increased for four consecutive quarters.

The more interesting detail is how the product itself has changed. Vanta now talks about continuously updated controls, vendor relationships and evidence across more than 400 integrations. Its agent can work across compliance, audits, vendor risk and questionnaires. In one example published by Vanta, Samsara consolidated 820 controls across 10 frameworks into roughly 260 common controls and cut its vendor-review process by 50%.

That tells us how high the bar already is. A new compliance startup offering “ask AI questions about SOC 2” would arrive several years too late.

There is still room lower down in the workflow. Many companies have niche systems that established compliance platforms do not deeply understand. Employees still collect screenshots, investigate failed controls, find supporting files, request evidence from other teams and prepare material for auditors.

An Astra product could specialize in one messy part of that process. For example, it could continuously gather evidence from a specific vertical software stack, explain failed checks, prepare remediation steps and assemble the auditor package.

We would keep material compliance judgments with a person. The automation opportunity comes from eliminating the hours spent chasing evidence and moving it between systems.

Can GPT-6 Astra become a QA engineer companies will actually pay for?

GPT-6 Astra can power a strong QA product, but we would use the model to create and investigate tests rather than make every test run depend on an expensive AI agent clicking around.

Astra is clearly better at visual computer work. OpenAI shows it running frontend QA, and ScreenSpot-Pro rises to 92.7% from 76.9% for GPT-5.6 Sol.

The testing market gives us a pretty good clue about how to productize that ability.

QA Wolf currently promotes an agent that explores an application and writes real Playwright or Appium code. The company says those generated tests run 12 times faster than computer-use agents and remain deterministic, meaning the same test follows the same path each time.

Autify's current pricing also confirms that smaller teams already pay for AI-assisted testing. Its Core plan is $99 per month when billed annually, while Team is $450 per month.

We would push the idea further into the release process. Astra could inspect a new feature, work out which user flows might break, create or update deterministic tests and run them. When something fails, it can review logs, reproduce the issue and decide whether the failure looks like a bug, bad test data or a broken test.

The product could then open a bug with the exact reproduction and, for straightforward cases, prepare a code fix and rerun the failed test.

That gives the customer something much easier to buy: confidence that a release has been checked before production.

Can GPT-6 Astra automate month-end close without scaring accountants?

GPT-6 Astra can automate a meaningful part of month-end close today, especially the reconciliation and investigation work that accountants still do manually.

The freshest evidence here is Rillet. The AI-native accounting company now says it has more than 600 customers, and recent reporting says its annualized revenue rate doubled in a single quarter. It recently raised $100 million at a $1 billion valuation.

The customer results show why businesses are paying.

Found used to spend five or six days reconciling cash and needed 12 to 14 days for month-end close. According to Rillet's case study, cash reconciliation dropped to one day and the company moved toward a five-day soft close with 90% to 95% confidence in the numbers.

Found also automated a monthly process that involved pulling data from BigQuery, transferring it into another template and manually creating 15 to 20 partner emails and Excel reports. Those reports now go out on day three rather than day ten.

Astra could attack exactly this kind of work without trying to become the company's general ledger on day one. It can retrieve invoices, investigate unmatched payments, compare data across billing and CRM systems, prepare proposed classifications, explain unusual variances and assemble a review package.

The accountant keeps approval authority over entries that require judgment.

That division makes the product much easier to sell. We are removing hours of investigation while keeping the person who signs off on the books in control.

Can GPT-6 Astra turn analyst work into a product people buy?

GPT-6 Astra makes narrow research-and-analysis services much easier to turn into software, although a generic “AI analyst” is becoming harder to defend.

Astra currently supports a 1.05-million-token context window and can work directly with documents, spreadsheets and presentations. OpenAI has also trained it to follow existing templates instead of producing a generic-looking deliverable every time.

That combination opens some genuinely useful businesses.

Take commercial real estate. A buyer could provide a property address and data room. The product retrieves relevant public records and comparable transactions, reads the documents, checks the numbers, updates the buyer's financial model and produces the standard investment memo.

A procurement team could give the system a potential vendor. It gathers the vendor's security documents, commercial terms and outside information, updates the company's scoring model and produces the packet used for approval.

A small investment team could use the same idea for one specific type of acquisition. The product extracts the operating history, checks inconsistencies across the data room, performs outside research, updates a standard model and prepares the first investment-committee draft.

We would go very narrow here because ChatGPT Work is already capable of broad research and business-document creation. The product needs to know how one particular decision is made, which sources are trusted, which calculations matter and what the finished deliverable should look like.

“Analyze any company” is easy to substitute. “Produce our complete first-pass diligence package for a self-storage acquisition” has much more substance around the model.

Is customer support already too crowded for a GPT-6 Astra startup?

A general customer-support agent powered by Astra looks unattractive right now because the incumbents have already moved from answering questions to taking actions and charging for outcomes.

Intercom is the clearest example. Fin currently costs $0.99 per successful outcome over chat and email, and Intercom says Fin resolves an average of 76% of conversations across its customers. The system can also take actions in external tools and hand the conversation to a person when needed.

Intercom has even extended this model into sales. A successful qualification currently costs $9.99, while a disqualification or simple resolution costs $0.99.

The pricing itself tells us how mature this category has become. Customers are already being asked to pay when the AI finishes useful work.

A new “Astra customer support agent” would therefore compete with companies that already own the inbox, help center, conversation history, routing, reporting, integrations and human handoff.

A vertical version could still work. Imagine technical support for one category of laboratory equipment where the agent understands manuals, machine telemetry, replacement parts and the manufacturer's service procedures. That is much harder for a horizontal support platform to cover deeply.

For normal SaaS support, however, we would look elsewhere.

Is building another GPT-6 Astra coding app a trap?

A general coding product built around GPT-6 Astra is one of the hardest markets to enter today because the leading products already have enormous usage, and they can use Astra too.

Lovable said in June 2026 that it had passed a $500 million annualized revenue run rate. The company also said users were creating around one million new projects each week and had created more than 50 million projects in total.

OpenAI's own Astra launch includes a testimonial from Lovable's CTO describing better performance on fresh application builds when the model is allowed to spend more effort iterating, verifying through browser testing and running code.

Devin is attacking the same market from another direction. Cognition recently changed Devin's self-serve pricing, with Pro starting at $20 per month and more expensive tiers available for heavier users. Codex then adds OpenAI itself to the list of direct competitors.

The problem for a new builder is pretty simple: any Astra coding improvement can quickly reach several large competitors at the same time.

A narrower coding business can still make sense. We could build something that creates and maintains one very specific type of internal application, integration or industry workflow. The value then comes from everything we know about that job.

But we would avoid competing on “tell us what software you want and Astra will build it.” That battle is already crowded with extremely well-funded companies.

Is legal AI still open to new GPT-6 Astra startups?

Legal AI is still open to new Astra products, but the opportunity has moved toward specialized workflows as the general legal-assistant market gets much more competitive.

Harvey now shows just how large the category has become. The company raised at an $11 billion valuation earlier this year and said customers were already running more than 25,000 custom agents. The latest reporting puts Harvey at roughly $350 million in annual recurring revenue and more than 200,000 lawyers using the product.

Something even more interesting has happened lately: large law firms are starting to build more of their own AI workflows.

Recent Financial Times reporting found that firms including Kirkland & Ellis, Freshfields and Goodwin Procter are investing in customized AI systems. Kirkland has committed $500 million to its own platform, while Freshfields is working with Anthropic on legal tools. The FT estimated that roughly 20% of large law firms are now personalizing or building AI technology.

That leaves a different opportunity from launching another Harvey clone.

A smaller Astra company could build the workflow layer for one practice area or one kind of repetitive legal process: preparing a particular filing, checking a specific transaction package, organizing one type of discovery work or generating documents from a firm's own playbook.

The customer already has access to powerful general models. What they increasingly want is software that understands how their particular legal work gets done.

Is cybersecurity GPT-6 Astra's biggest opportunity or a bad place to start?

GPT-6 Astra has extraordinary cybersecurity capabilities, but cybersecurity would be a difficult first Astra business for a small team because the safety and trust burden is much higher than in ordinary workflow automation.

OpenAI now classifies Astra at the Critical cybersecurity capability level under its Preparedness Framework. In its launch evaluations, Astra scored 100% on ExploitBench versus 78.5% for GPT-5.6 Sol. OpenAI also tested vulnerabilities disclosed during a recent three-month period to reduce the chance that benchmark contamination explained the result; Astra scored 39% there compared with 5.5% for GPT-5.6 Sol.

During those evaluations, OpenAI says Astra discovered two previously unknown zero-day vulnerabilities.

This is a genuine capability jump. It also explains why OpenAI applies additional restrictions and monitoring around advanced cyber use. Some higher-risk workflows remain restricted even when the intended use is defensive.

For builders, the safer commercial opportunity sits around tightly authorized defensive jobs: reviewing code for security problems, investigating suspicious configurations, checking patches, prioritizing vulnerabilities or helping security teams understand failures.

The upside could eventually be enormous. We simply would not choose autonomous penetration testing as the easiest way to launch an Astra company this week.

How should a GPT-6 Astra product charge customers?

GPT-6 Astra products should usually charge around completed work when the result is easy to measure.

We can already see that shift in current AI products. As seen above, Intercom charges for completed Fin outcomes. Conveyor describes its pricing around AI work and usage instead of adding a fee for every user. Devin meters heavier autonomous work because the underlying compute cost varies enormously from one task to another.

The same logic works for the opportunities we found.

A migration product can charge per migration or according to record volume and complexity. An RFP product can charge per completed questionnaire or include a defined number of submissions in a subscription. A QA product can price around applications, tested flows or releases. A finance agent can charge according to entities or monthly closes.

This is better than exposing token costs to the customer. Nobody wakes up wanting 800,000 Astra tokens. They want their CRM migrated before Monday or their security questionnaire back to the prospect today.

Seat pricing can still work when the product is genuinely collaborative. But the more work Astra performs on its own, the less closely seat count tracks the value being created.

What can you actually build with GPT-6 Astra that people will pay for?

The best GPT-6 Astra businesses right now are narrow agents that take responsibility for expensive work, with software migration, RFP operations and QA standing out as the strongest places to start.

Software implementation and migration would be our first choice. We have strong evidence that customers already spend thousands of dollars on the outcome, Astra has made a particularly large jump on migration and terminal tasks, success can be checked objectively and a founder can start with one very narrow integration pair.

RFP and security-questionnaire operations come next. The existing software routinely costs five figures per year, the work directly affects sales, and Astra's computer use allows a product to go further than drafting answers by working across files and procurement portals.

QA also looks strong. The model can understand an interface, write code and investigate failures, while deterministic test execution keeps recurring inference costs and reliability problems under control.

Finance operations and compliance evidence collection could become larger companies, although we would expect longer sales cycles and a higher trust burden.

Vertical diligence and analyst workflows are easier to launch. We would pursue them only where we can define a very specific decision process and finished output because broad research is rapidly becoming part of ChatGPT Work itself.

Customer support, generic coding tools and general browser agents sit at the bottom of our list. Strong companies already own those categories, and every improvement to Astra strengthens those incumbents too.

So if we were building on Astra today, we would start with a job that currently costs a company hundreds or thousands of dollars, narrow it until success can be checked automatically, let Astra handle most of the execution, and build human review around the few decisions where mistakes are expensive.

That gives us something people already know how to buy: completed work.

Opportunity	Why Astra fits	Difficulty	Our verdict
Software migration & implementation	Computer use, coding, database work, clear validation	Medium	Best first bet
RFP & security questionnaires	Documents, browser work, company knowledge, approvals	Medium	Excellent
QA release agent	Computer use + coding + reproducible checks	Medium	Excellent
Compliance evidence collection	Repetitive cross-system work with clear exceptions	High	Strong larger business
Finance close operations	Reconciliation, investigation and document work	High	Strong larger business
Vertical diligence	Research, long context, models and deliverables	Medium	Good if very narrow
Vertical legal workflows	Complex work and high willingness to pay	High	Attractive niche
General support agent	Mature incumbents already own the workflow	High	Weak entry point
General coding/app builder	Intense competition using the same frontier models	Very high	Avoid
Generic Astra wrapper	Little workflow ownership	Low to build	Avoid