On 3 September 2026 OpenAI released GPT-6 Astra and called it a new frontier on computer and browser use: a model that fills in online forms, updates CRM records, runs research and installs and tests software by driving a screen the way a person does. Anthropic's Claude has shipped a computer use tool too. The question for any Indian business with a back office is no longer whether a model can click - it is which jobs belong to a model with a mouse, and which should stay on API automation, where they are cheaper and easier to audit.
What GPT-6 Astra actually shipped on 3 September 2026
OpenAI's announcement is unusually specific about the work Astra does on a screen: filling out online forms, updating customer records in a CRM, organising a calendar, conducting research and drafting summaries in your email or document editor, analysing data and generating plots, creating a website and running frontend QA checks on it, and autonomously installing and testing software. It went to a limited set of organisations first - TechCrunch reported the first tranche were Daybreak cybersecurity programme customers - then to ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API, Microsoft Azure and AWS Bedrock. Administrators must switch it on: access is off by default.
- •API pricing starts at $10 per million input tokens and $50 per million output tokens, with a Fast mode at up to twice the speed for twice the price.
- •New enterprise admin controls restrict Astra to approved websites and desktop applications, manage uploads and downloads, and control browsing history.
- •Confirmation policies can require human approval before consequential actions, and an automated review layer checks unsafe or unauthorised tool calls.
- •Astra is the first OpenAI model to meet the Critical cybersecurity threshold under its Preparedness Framework, which is why the rollout is staged.
- •Anthropic ships the same capability: its computer use tool is generally available on the Claude API for Claude Opus 5.5, Fable 5.1, Mythos 5.1 and Sonnet 5, via a seventeen-tool screenshot-and-click toolset.
- •Zero Data Retention is available to eligible OpenAI API customers on supported endpoints, subject to approval - relevant if you care about what leaves your network.
How good is computer use really? Read the benchmark small print
The headline number is 72.6% on OSWorld 2.0, and you should know what it measures. OpenAI reports it as a partial score on the offline subset - credit for checkpoints reached, not tasks finished end to end. The benchmark's own authors are blunter: OSWorld 2.0 holds 108 long-horizon workflows across seven professional domains, a skilled human needs a median of roughly 1.6 hours per task, and the top entry on their public leaderboard completes only about a fifth of tasks outright. Astra is a real step forward - it hits 72.6% in roughly 40 minutes per task against 75 for GPT-5.6 Sol - but 'better than every previous model' is not 'finishes the job unsupervised'.
Computer-use agent vs n8n API automation vs RPA vs a human
There are four ways to get a back-office task done and they are not interchangeable. The table below is qualitative except where a figure is sourced - treat it as a shape to reason with, not a quotation.
| What matters | Computer-use agent | n8n / API automation | Traditional RPA | A person |
|---|---|---|---|---|
| Speed per task | Minutes to tens of minutes; OpenAI reports roughly 40 minutes per OSWorld 2.0 task | Seconds - a run is a few HTTP calls | Fast once recorded, slow again once the UI moves | Minutes to hours, in working hours only |
| Cost per task | High - you pay per token for every screenshot and reasoning step | Very low - n8n Pro is EUR 50 a month for 10,000 executions | Licence-heavy, with per-bot or per-process fees | Salary, plus the work not done instead |
| Reliability | Probabilistic - the same task can take a different path twice | Deterministic - same input, same result | Deterministic until a selector changes, then it fails quietly | Strong on judgement, variable on volume |
| Auditability | Screenshots and action logs, but only if you capture them | Node-by-node logs you can query and replay | Run logs, inside the vendor's console | Whatever the person wrote down |
| Setup effort | Low to start, high to make safe - environment, allowlists, gates | Medium - credentials plus modelling the data | High - recording, exception handling, permanent maintenance | None, until you need a second person |
| Best fit | GUI-only legacy software, portals with no API, one-offs, frontend QA | Anything with an API or webhook - most modern SaaS | Large, stable, high-volume processes on a fixed desktop estate | Judgement, exceptions, anything a customer notices |
The rule of thumb: if it has an API, do not click
This is the decision rule we apply before quoting anyone for this work. If the system exposes an API, a webhook or even a scheduled CSV export, use it. A workflow that calls an endpoint is cheaper by roughly three orders of magnitude than an agent driving the same action through a browser, it returns the same result every time, and it leaves a structured execution log instead of a folder of screenshots. Computer use earns its cost only when there is no other door in.
- •Has an API or webhook? Use n8n or direct API calls. That covers most modern CRMs, accounting SaaS, WhatsApp Business, payment gateways and e-commerce platforms.
- •No API, but a scheduled export or an emailed attachment? Automate the file, not the clicking - it stays deterministic and almost free.
- •No API and no export, GUI only? The genuine computer-use case: desktop accounting packages, old ERP clients, supplier and buyer portals.
- •A one-off that would cost a person three days? Computer use, time-boxed - and do not let it quietly become a permanent process.
- •Anything a customer sees, or any irreversible movement of money? A named human approves it, whichever technology drafted it.
Ten back-office tasks, and a verdict for each
These verdicts assume an Indian SMB with a small operations team and no automation engineer. 'Computer use' means a model driving a screen under the controls described below - not a model loose on somebody's laptop.
- •Reconciling bank statements in a desktop accounting package with no API. Verdict: computer use. Start read-only - let it match and flag, and keep a person posting the journal entries.
- •Pulling invoices out of a supplier portal with no API or bulk export. Verdict: computer use for the download only, then n8n for everything after the file lands.
- •Filling in a customer's procurement portal because they will not accept email. Verdict: computer use with a hard gate before Submit. The agent prepares, a human presses the button.
- •Moving leads from WhatsApp into your CRM. Verdict: not computer use. WhatsApp Business and every mainstream CRM have APIs, and n8n does this for a fraction of a rupee per lead.
- •Extracting details from lead screenshots your sales team forwards. Verdict: neither. That is a vision extraction call plus an API write, not an agent with a mouse.
- •Competitor and marketplace price monitoring. Verdict: try structured feeds and plain HTTP first. Reach for computer use only where the page genuinely requires interaction.
- •QA-testing your own web app before a release. Verdict: strong fit - OpenAI lists frontend QA checks explicitly. Convert what it finds into Playwright tests you keep and re-run free.
- •A one-off migration of a few thousand records between two systems that cannot talk. Verdict: computer use, time-boxed, sampled for accuracy, then retired.
- •Preparing a statutory filing on a government portal. Verdict: use it to gather and reconcile, never to submit. A named person files and owns the consequences.
- •Building the monthly client report deck from your template. Verdict: good fit - OpenAI highlights template adherence - but a person reads it before it leaves the building.
What does a computer-use task actually cost?
Work this out before you commit, because the arithmetic is not intuitive. Astra costs $10 per million input tokens and $50 per million output tokens, and Anthropic documents each screenshot at roughly 1,000 to 1,800 input tokens - and screenshots accumulate in the conversation. Take an illustrative task of 60 model turns, pruning context to the five most recent screenshots plus about 8,000 tokens of instructions and page text: roughly 930,000 input and 24,000 output tokens overall. That is about $10.50 for one task, near Rs 1,005 at the ECB reference rate of Rs 95.74 to the dollar on 23 September 2026. n8n's Pro plan, by contrast, is EUR 50 a month for 10,000 whole-workflow executions - half a euro-cent per run. Those figures are an illustration on stated assumptions, not a quote - the ratio is the point.
Failure modes, and the approval gates that catch them
Both vendors publish the same warnings, and they read better as a specification than boilerplate. OpenAI's guide says to restrict the environment to an isolated browser or VM with an allow list, treat screen content as untrusted because text in a page cannot grant permission, confirm consequential actions, and bound the run with step, time or cost limits while verifying the actual outcome rather than the model's final answer. Anthropic adds what people skip: a VM with minimal privileges, no access to login credentials, an allowlist of domains, and a human confirming anything with real-world consequences.
⚠️Giving the agent the same login your staff use
Consequence: It inherits every permission that person holds, including deleting records or approving payments, and your audit trail says a human did it.
Solution: Create a dedicated least-privilege account per automation, and let it log in to nothing that moves money.
⚠️Letting the agent press Submit unsupervised
Consequence: A misread field lands in a customer's portal or a statutory form, and you find out weeks later.
Solution: Use the confirmation policy - OpenAI's controls can require approval before consequential actions - and route every draft to a named approver.
⚠️Trusting the model's own report of what it did
Consequence: The run summary says the CRM was updated, the CRM disagrees, and nobody spots the gap until month end.
Solution: Verify against the system of record. OpenAI's guidance is to check the actual outcome, not the final answer.
⚠️Pointing it at an unrestricted browser
Consequence: Text on a web page or inside an image can hijack the agent - both vendors document this prompt-injection risk.
Solution: Allowlist the sites and applications it may touch, run it in an isolated VM, and control uploads and downloads.
A 30-day pilot that will not embarrass you
Treat the first month as an experiment with a number at the end, not a rollout. The goal is a defensible cost per task you can compare against a human and against the API route you skipped.
Week 1 - pick one GUI-only task
Something with no API, a clear definition of done and no customer visibility. Time how long a person takes today and write that number down.
Week 1 - build the cage first
An isolated VM, a least-privilege account, an allowlist of sites and applications, upload and download controls, and a step and cost ceiling per run.
Week 2 - run it read-only
Let the agent observe, extract and propose without writing anywhere. Compare its output against your person's and count the disagreements.
Week 3 - add one write, behind approval
Enable a single consequential action with a confirmation gate. A named person approves each one, and approver, timestamp and outcome are logged.
Week 4 - decide honestly
Compare cost per task against the human baseline and against what an API route would have cost. If an API exists, move the job to n8n.
How Tech Arion helps with computer-use agents and back-office automation
We build the unglamorous half of this, which is the half that decides whether it works. Most back-office jobs Hyderabad and Chennai businesses bring us turn out to have an API nobody had looked for, and those become n8n workflows - deterministic, logged and cheap to run. See techarion.com/services/n8n-consulting for how we scope that. Where the software genuinely has no API - a desktop accounting package, an ageing ERP client, a buyer's portal - we design the computer-use layer around it: isolated environment, least-privilege credentials, allowlists, confirmation gates on anything consequential, and logging an auditor will accept. Model selection, cost modelling and governance sit under techarion.com/services/ai-consulting, and when a small internal tool would remove the GUI problem altogether we build that instead at techarion.com/services/vibe-coding.
Frequently asked questions about computer-use agents for back-office work
The questions Indian business owners and CTOs ask us most often about handing back-office work to a model that can drive a screen.
Frequently Asked Questions
Work out which jobs actually need an agent
Tech Arion scopes back-office automation the unglamorous way. We find out which of your systems have APIs and move those to n8n, where they are deterministic and cost almost nothing to run. Then we build a properly caged computer-use layer only for the GUI-only software that leaves no other option - environment isolation, least-privilege credentials, allowlists, approval gates and logs an auditor will accept. Bring us one painful process and we will tell you honestly which of the four routes it belongs on.
Sources & References
Primary sources fetched on 24 September 2026:
- 1.
OpenAI. (3 Sep 2026). GPT-6 Astra: A new generation of intelligence - computer-use task list, OSWorld 2.0 offline-set partial score of 72.6% versus 65.7% for GPT-5.6 Sol at roughly 40 vs 75 minutes per task, rollout to ChatGPT Plus/Pro/Business/Enterprise, the API, Microsoft Azure and AWS Bedrock, access off by default at launch, API pricing of $10/$50 per million tokens, Critical cybersecurity threshold.
View Source - 2.
OpenAI. (Sep 2026). GPT-6 Astra: The next generation in intelligence for work - enterprise admin controls for approved websites and desktop applications, upload/download management and browsing history, confirmation policies and automated review of unsafe tool calls, pricing of $10 input / $50 output per million tokens, Zero Data Retention for eligible API customers.
View Source - 3.
OpenAI. (2026). Computer use - API guide, 'Run safely' section: restrict the environment to an isolated browser or VM with an allow list, treat screen content as untrusted, confirm consequential actions, bound and verify the run with step, time or cost limits.
View Source - 4.
Anthropic. (2026). Computer use tool - Claude Platform docs: seventeen-tool computer_toolset_20260801, supported models including Claude Opus 5.5 and Fable 5.1, GA on the Claude API and Google Cloud, beta on Amazon Bedrock and Microsoft Foundry, screenshots costing roughly 1,000-1,800 input tokens, and the VM/credential/allowlist/human-confirmation precautions and prompt-injection warning.
View Source - 5.
XLANG Lab. (2026). OSWorld 2.0 benchmark and leaderboard - 108 long-horizon computer-use workflows across seven professional domains, median human completion time of about 1.6 hours, an average of 318 tool calls per task versus about 30 in OSWorld 1.0, and top listed binary completion of 20.6% (Claude Opus 4.8).
View Source - 6.
TechCrunch. (3 Sep 2026). OpenAI launches Astra, its powerful and controversial new model - release on Thursday 3 September 2026, first to Daybreak cybersecurity programme customers, broader rollout to paid plans and the API within a week.
View Source - 7.
InfoQ. (10 Sep 2026). OpenAI releases GPT-6 Astra - OSWorld 2.0 at 72.6% versus 65.7%, Terminal-Bench 4.0 at 57.9%, hallucination rate of 4.2% versus 12.2%, long-context MRCR at 96.3% for 512K-1M, availability via the OpenAI API, Microsoft Azure and AWS Bedrock.
View Source - 8.
n8n. (2026). Pricing - Pro plan at EUR 50 per month for 10,000 workflow executions, Starter at EUR 20 for 2,500, and the definition of an execution as a single run of an entire workflow regardless of step count.
View Source - 9.
Frankfurter (European Central Bank reference rates). (23 Sep 2026). USD to INR reference rate of 95.74, used for the illustrative per-task cost conversion in this article.
View Source