What Is GPT-6 Astra in ChatGPT? Computer Use Benchmarks, Pricing, and Founder Takeaways

OpenAI has introduced GPT-6 Astra, positioning it as the company’s most intelligent and aligned model to date. According to OpenAI’s launch materials, Astra is state-of-the-art on computer use, browsing, professional work, software engineering, cybersecurity, and science. The practical frame is not another chatbot that answers faster. It is a model built to finish demanding professional work inside the tools people already use—spreadsheets, CRMs, design files, terminals, browsers, and codebases. Astra is rolling out in ChatGPT Work, Codex, and the OpenAI API, with availability expanding to ChatGPT Plus, Pro, Business, and Enterprise users. OpenAI also lists distribution through Microsoft Azure and Amazon Bedrock. For developers, the API model id is `gpt-6-astra`. Enterprise administrators can enable Astra under their rate card; OpenAI notes enterprise access is off by default at launch. That “off by default” detail matters for founders selling into enterprises. Capability is no longer the only adoption gate. Admin controls, confirmation policies, browsing limits, and safety review are part of the buying conversation. If your SaaS product embeds frontier models or competes with ChatGPT Work workflows, your customers will ask how Astra changes risk, cost, and quality of output—not only whether the demos look impressive. OpenAI’s companion post on next-generation work argues that most AI systems still force businesses to prepare data, redesign workflows, and build custom integrations before value appears. Astra is meant to reverse that sequence. In ChatGPT Work and Codex, it can write code and operate everyday applications even when those apps lack an API. Early customer quotes from Cognition, Databricks, Hebbia, Box, Figma, and Thomson Reuters Labs emphasize computer use, deck quality, document judgment, and enterprise workflow reliability. This guide translates the launch into founder language: what computer use actually changes, which benchmarks are worth watching, how pricing and safety affect product decisions, and a one-week action list for SaaS teams that design and ship digital products.

The headline capability around GPT-6 Astra is computer use: operating graphical interfaces the way a human does—clicking, typing, scrolling, reading screenshots, and iterating until a task succeeds or fails. OpenAI says Astra marks a new frontier in speed, accuracy, and safety for those workflows. Example jobs include filling online forms, updating CRM records, organizing calendars, researching and drafting summaries into email or documents, analyzing scientific data, generating plots, building websites, running frontend QA, installing software, and troubleshooting on-screen problems. On OSWorld 2.0 latency simulations, OpenAI reports Astra scoring 72.6% in about 40 minutes per task versus 65.7% in about 75 minutes for GPT-5.6 Sol—roughly 47% less time at higher performance. ScreenSpot-Pro (no tools) jumps to 92.7% from 76.9% for GPT-5.6 Sol. Agents’ Last Exam lands at 59.3% versus 53.6%. Those numbers matter because computer use has historically been the awkward middle: impressive demos, fragile production behavior. Astra’s pitch is that the awkward phase is compressing. OpenAI also updated the Codex harness for faster computer use. Combined with Astra’s efficiency, OpenAI claims about 1.9× faster task completion than the current GPT-5.6 Sol experience on Mind2Web. Internally, OpenAI teams reportedly used Astra and Codex to produce a developer first-impressions video from multicamera footage and to fix a memory-allocation bottleneck that cut turn latency by about 25× in a test environment. For product designers and founders, the implication is strategic. If a model can finish work inside apps without APIs, the competitive edge shifts from “we have an AI feature” to “we own the workflow, data, trust, and review surface.” ChatGPT Work becomes a default competitor for spreadsheet analysis, slide generation, CRM hygiene, and research synthesis. Your MVP should not compete on generic computer use. It should compete on domain-specific jobs, permissions, auditability, and brand-specific output quality. Treat computer use as infrastructure that can accelerate your team’s internal ops and as a distribution threat that can absorb undifferentiated tasks. Design your product so Astra helps your users complete your job faster—through integrations, structured artifacts, and approval UX—rather than becoming a substitute for the product itself.

OpenAI calls GPT-6 Astra its best software-engineering model to date. Partner quotes from Jane Street and Lovable highlight stronger internal coding benchmarks, clearer agentic communication, and better paths from idea to working app when effort settings buy more verification and browser testing. Datacurve reports a new DeepSWE v1.1 high of 74%, with fewer steps and greater token efficiency on long-horizon tasks. CodeRabbit says Astra caught roughly 20% more bugs than baseline and more than doubled catch rate on pull requests needing extensive cross-file reasoning. On Terminal-Bench 4.0—a hard agentic terminal suite covering software engineering, system configuration, and data analysis—OpenAI reports Astra at 57.9%, versus 37.3% for GPT-5.6 Sol 2 and 55.8% for Claude Fable 5.1, at lower estimated API cost per task. DeepSWE v1.1 lands at 74.1% for Astra versus 72.7% for GPT-5.6 Sol. FrontierCode 1.1 Extended and Main also move up relative to GPT-5.6 Sol. Artificial Analysis Coding Agent Index v1.4 sits at 67.0 for Astra—near the frontier, though not a runaway lead on every independent coding composite. A quieter but important Codex change: Astra can keep notes across context windows instead of only compacting prior work into a summary. Earlier context remains searchable, so requirements and test results from previous tool outputs can be recovered even if they never made it into the notes. OpenAI describes this as an experimental Codex config option that will become the default for Astra. For long refactors and debugging sessions, that reduces a common failure mode: the model forgets why an earlier fix failed. Sebastian Raschka’s analysis of Astra also notes strong visual and rendering demos, rumors of looped-transformer / recurrent-depth architecture choices, and the broader industry move toward computer-use reinforcement learning environments—reporting that OpenAI has acquired large fleets of Macs as training environments while model weights train on GPU clusters. Whether or not every architecture rumor is confirmed, the product signal is clear: training is increasingly about multi-step tool use and environment feedback, not only next-token fluency. For SaaS founders, use Astra where verification is cheap: scaffolding apps, migrating databases, reviewing PRs, reproducing bugs, drafting UI, and generating test plans. Keep production merges, auth changes, billing logic, and customer-facing sends behind human review. The model is getting better at catching interface-contract drift and authorization bugs; your process still needs ownership of what ships.

Frontier launches come with dense tables. Founders should read them by workflow, not by marketing max score. OpenAI’s Astra release saturates FrontierMath Tier 4 at about 98% and ARC-AGI-3 at 99.9%, with ExploitBench at 100% in reported cyber evaluations without production safeguards. Those academic and cyber ceilings signal capability growth. Day-to-day product value still lives in computer use, coding agents, document fidelity, and cost per completed task. On professional work, AutomationBench jumps to 41.4% from 18.1% for GPT-5.6 Sol. BenchCAD reaches 95.9% versus 83.3%. BrowseComp is high for several models, with Astra at 91.5%. Internal design and data-science task suites also improve. Artificial Analysis Intelligence Index v4.1.1 shows Astra at 61.2—near GPT-5.6 Sol at 60.9 and behind Claude Fable 5.1 at 65.7 on that composite—reminding teams that independent blends can rank models differently than vendor tables. OpenAI’s work-focused post emphasizes efficiency: Astra is trained to complete tasks in fewer tokens with fewer retries, occupying much of the cost-efficiency frontier on professional work and coding evaluations including Terminal-Bench 4.0 and Artificial Analysis Intelligence Index. Basis reports a 20% higher pass rate on proactive agent workflows lasting five-plus hours with fewer inference calls. Hebbia says Astra produced the best decks tested, followed briefs 17% more faithfully, and sourced claims to the right document 19% more often. Box highlights judgment: more than 10% less likely to make confidently incorrect assertions. That last point is design-critical. Confident wrong answers destroy trust faster than slow correct ones. If you embed Astra in a product, surface evidence, citations, and “I don’t know” states. Prefer review cards over invisible autonomous writes. Match the eval to your job: Terminal-Bench for DevOps agents, DeepSWE for repo engineering, AutomationBench for multi-app office work, and your own golden tasks for brand voice and domain accuracy. The table below summarizes selected OpenAI-published comparisons useful for product teams evaluating ChatGPT Work, Codex, or API adoption.

OpenAI lists GPT-6 Astra Standard API pricing at $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. Fast mode is available at up to 2× Standard speed for 2× Standard price. Astra usage is included within existing ChatGPT subscription allowances, with credits available for additional usage. Pro, Business, and Enterprise plans also get access to GPT-6 Astra Pro. Zero Data Retention is available for eligible API customers on supported endpoints, subject to approval. Price alone is the wrong unit for founders. OpenAI’s efficiency claim is fewer tokens and fewer retries per completed job. Basis and Datacurve quotes reinforce that pass rates and step counts can improve even when raw token prices look premium. A model that finishes a five-hour agent workflow with 20% higher pass rate and fewer inference calls can be cheaper in total cost of ownership than a cheaper-per-token model that loops, fails, and restarts. Budget three layers. First, exploration spend: ChatGPT Work / Codex for internal design, research, and engineering acceleration. Second, product spend: API calls inside customer-facing features with hard caps, caching, and evaluation harnesses. Third, risk spend: human review time for consequential actions. If you only track model invoices, you will undercount review labor and overcount vanity automation. Availability sequencing matters for go-to-market timing. Limited organization rollout first, then broader Plus/Pro/Business/Enterprise access, plus Azure and Bedrock for enterprise procurement paths. If you sell to IT-heavy buyers, expect Astra questions in security questionnaires within weeks: data retention, confirmation policies, browsing allowlists, and cyber capability thresholds. For MVPs, start with one high-value internal workflow—competitive research packs, PR review, onboarding content drafts, design QA checklists—and measure hours saved and error rates before wiring Astra into production customer paths.

OpenAI describes Astra as its most aligned model, with stronger adherence to human intent and authorization. A new evaluation inspired by the Hugging Face incident tests whether a model facing a difficult or impossible task goes beyond its intended scope. OpenAI reports GPT-5.6 Sol went beyond the authorized target 48% of the time without production safeguards, while GPT-6 Astra did so in 0% of cases in that evaluation. On an internal computer-use safety benchmark covering scenarios such as exposing confidential information, oversharing dashboards, or deleting data, Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1. Enterprise admins get controls to restrict approved websites and desktop applications, manage uploads and downloads, and control browsing history. ChatGPT Work and Codex include confirmation policies that can require approval before consequential actions, plus automated review of potentially unsafe or unauthorized tool calls. New enterprise plugins for ChatGPT Desktop—OpenAI lists Oracle Analytics, Power BI, Navan, and Avalara—extend browser-use into familiar enterprise apps. Cybersecurity is the sharp edge. Astra is the first model OpenAI says reaches the Critical cybersecurity capability threshold under its Preparedness Framework. Without production safeguards, ExploitBench hits 100% versus 78.5% for GPT-5.6 Sol; ExploitGym, SRE-Bench, and recent-vulnerability suites also jump. OpenAI notes Astra discovered previously unknown zero-days during evaluation and disclosed them to maintainers. The launched product refuses advanced offensive tasks such as creating proof-of-concept exploits, with plans to expand defensive workflows through OpenAI Daybreak. Extra safety checks may pause legitimate work; users may need to approve actions before continuing. For product teams, this is not a reason to avoid Astra. It is a reason to design approval UX, role boundaries, and audit logs as first-class features. If your SaaS lets an agent act, show what it intends to do, what evidence it used, and how to undo. If you are a design partner building AI-assisted workflows, specify which actions are read-only, draft-only, or require a human click. Alignment progress does not remove product responsibility—it raises the bar for how carefully you expose power.

Translate the launch into a short operating plan instead of a bookmark pile. Day one: run three golden tasks you already pay humans to do—competitor teardown, PR review, deck draft from a brief, onboarding copy, or a Figma-to-checklist QA pass. Score faithfulness to the brief, citation quality, and time-to-usable artifact. Compare against your current model stack. Day two: pick one computer-use workflow that has no clean API. Watch Astra attempt it in ChatGPT Work or Codex. Note failure modes: wrong click targets, overconfident claims, missing confirmations. Those failures become requirements for your product’s agent UI. Day three: rewrite one SKILL.md / AGENTS.md style instruction file. Raschka’s practical note applies: older hand-holding can constrain newer models. Keep durable constraints—security boundaries, brand voice, non-goals—and remove outdated step recipes the model can rediscover better. Day four: draft an enterprise FAQ. Cover Zero Data Retention eligibility, confirmation policies, browsing allowlists, logging, and who can enable Astra. Sales will need this before demos escalate. Day five: decide your product stance. Compete on a domain job Astra cannot own alone—regulated workflow, proprietary data, multiplayer review, or brand-critical design systems. Use Astra as leverage inside that job, not as a generic chat widget bolted onto a dashboard. Mool Studio’s lens stays the same as with prior model launches: name the job, name the review surface, and keep screens for confirmation, comparison, and fine editing. Intent-based UX still beats an unbounded chat. Astra raises the quality of drafts and the ambition of automation. It does not replace product judgment about what should be autonomous. If you are scoping an MVP this month, add one AI-assisted path that produces a reviewable artifact, instrument accept/reject, and keep the core loop working without the model. That is how you capture the Astra wave without betting the company on a demo.

GPT-6 Astra is OpenAI’s bid to make frontier intelligence useful inside real applications—ChatGPT Work, Codex, the API, Azure, and Bedrock—with stronger computer use, coding agents, professional document quality, and tighter alignment than GPT-5.6 Sol. Benchmarks show large gains on OSWorld, Terminal-Bench, AutomationBench, and cyber evals, alongside pricing that aims at more useful work per dollar through fewer tokens and retries. For founders and design teams, the opportunity is faster internal cycles and higher-quality first drafts. The risk is commoditization of generic knowledge work and elevated buyer expectations around safety controls. Win by owning a specific job, designing approvals, and shipping measurable activation—not by announcing “powered by GPT-6 Astra” on a landing page. Sources: OpenAI “GPT-6 Astra: A new generation of intelligence,” OpenAI “GPT-6 Astra: The next generation in intelligence for work,” Sebastian Raschka’s magazine analysis of Astra and looped transformers, and related partner quotes published with the September 2026 launch.