OpenAI began rolling out GPT-6 Astra on September 3, 2026, pitching it as the engine behind the next version of ChatGPT and Codex and, according to the company’s own framing, its most capable model yet for coding, computer use, and multi-step professional work. The announcement landed across 9to5Mac, CNET, The Indian Express, PCWorld, PCMag UK, and Breitbart within hours, and by September 4 the rollout had reached Pro, Business Premium, and Enterprise seats in ChatGPT Work and Codex, plus the OpenAI API.
For a company that shipped GPT-5.6 Sol only weeks earlier, a full model-generation jump to “GPT-6” is a bold naming move. It also arrives while OpenAI is still managing fallout from safety reviews tied to Astra’s own capability testing. The two threads (a splashy product launch and a still-unfolding safety story) are now running in parallel, and that tension is arguably the most important thing to understand about this release.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What OpenAI actually announced on September 3
According to OpenAI’s own release notes, cited by 9to5Mac, GPT-6 Astra brings “improvements in coding, research, computer use, and complex, multi-step work.” The company says Astra can generate documents, spreadsheets, and presentations that follow a user’s existing templates, and that it can adapt mid-task when instructions change (something earlier models handled poorly, often restarting a task from scratch rather than adjusting course).
The rollout itself is staged, which is standard practice for OpenAI’s larger releases but still worth flagging because it affects who can actually test the model this week. Access started with “a limited set of organizations,” per the release notes, and was set to expand over the following days to ChatGPT Plus, Pro, Business, and Enterprise users, as well as the OpenAI API and AWS Bedrock. By September 4, OpenAI’s own account on X confirmed that Astra was live for all Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex, and in the API, with Plus and Business Standard users still waiting on a slower rollout.
Coverage from Breitbart and other outlets picked up on OpenAI’s own characterization of the release as the arrival of an “AGI era,” a phrase the company has used sparingly in the past and one that immediately drew skepticism from independent commentators given how loosely “AGI” gets defined across the industry.
The benchmark numbers OpenAI is leading with
OpenAI’s marketing for Astra leans heavily on three benchmark results: a 98% score on FrontierMath Tier 4, a 99.9% score on ARC-AGI-3, and a 100% score on ExploitBench, a cybersecurity-focused evaluation. Those numbers, as reported by 9to5Mac, are near-ceiling on tests that were specifically designed to be difficult for prior-generation models, which is part of why OpenAI is framing this as a discontinuous jump rather than an incremental update.
On the “computer use” side (models operating a screen the way a person would, clicking through forms, updating records, navigating between apps) reporting associated with the release cites an OSWorld 2.0 score of 72.6% for Astra against 65.7% for GPT-5.6 Sol, with Astra completing tasks in roughly 47% less time. OpenAI separately says ChatGPT is “nearly 2x faster at computer use” with Astra than with its immediate predecessor, and describes Astra as “the best model for software engineering to date.”
Those are OpenAI’s own benchmark selections and framing, not independent third-party audits, which is a distinction worth holding onto. Companies naturally publish the tests that make their own model look best, and a 100% score on any benchmark should prompt a question about whether the benchmark has been saturated (meaning it’s no longer difficult enough to distinguish between top models) rather than treated as proof of open-ended capability.
Codex gets a genuinely new memory mechanism
The most concrete technical change in this release, and the one most likely to matter to working developers, is in Codex rather than in general ChatGPT. Astra introduces a new way for Codex to preserve and retrieve context once a session’s context window fills up. Instead of compacting or summarizing older context (the standard approach, which tends to lose detail as conversations get longer), Astra keeps searchable notes that persist across a session without being compressed.
Practically, that should mean fewer situations where a long refactor or debugging session causes an agent to “forget” a constraint that was established fifty messages earlier. Using Astra in Codex requires Codex CLI version 0.153.0 or newer along with the latest ChatGPT desktop app, so teams running older CLI versions won’t see any of this until they upgrade their tooling. OpenAI has also kept GPT-5-codex, its GPT-5 variant tuned specifically for agentic coding, available alongside Astra in Codex, positioning Astra as the higher-capability but presumably pricier option for teams that need it.
Pricing: what Astra costs on the API
For developers building on the API rather than using ChatGPT directly, Astra is priced at $10 per million input tokens and $50 per million output tokens, according to details reported by Kingy AI‘s review of OpenAI’s API documentation. Cached input tokens are billed at $1 per million, a significant discount for workloads that repeatedly reuse the same context, while cache writes run $12.50 per million. OpenAI also offers a “Fast” mode at double the standard rate: $20 per million input tokens and $100 per million output tokens.
Inside ChatGPT Work and Codex, usage is measured in credits rather than raw token pricing: 250 credits per million input tokens, 25 credits per million cached input tokens, and 1,250 credits per million output tokens, with the Fast multiplier again running at 2.5 times the standard rate. API rate limits scale by usage tier, starting at 500 requests per minute and 500,000 tokens per minute at Tier 1, up to 15,000 requests per minute and 40 million tokens per minute at Tier 5.
Who gets Astra first, and who’s still waiting
Access to Astra is not uniform across OpenAI’s subscription tiers, and the differences are more granular than past rollouts. Pro subscribers paying $200 a month reportedly get 200 GPT-6 Pro messages per week in Chat, while the $100 Pro tier gets 50 messages per week, shared with GPT-5.6 Sol Pro usage. Business Premium seats get 50 included messages per week, while Business Standard seats are limited to 15 included messages per month. Plus subscribers are excluded from GPT-6 Pro in Chat entirely at launch, though they do get limited Astra access in ChatGPT Work and Codex, with the option to buy additional credits once that allowance runs out.
Enterprise customers face an additional wrinkle: Astra access is reportedly off by default when it first reaches Enterprise accounts, requiring administrators to explicitly turn it on. That default-off posture, unusual for a headline product launch, lines up with the safety monitoring OpenAI says it built into this release specifically because of how capable the model tested.
| Plan | Astra in Chat | Astra in Work / Codex | Default status |
|---|---|---|---|
| Plus ($20/mo) | Not included at launch | Limited allowance, credits available | Rolling out over following days |
| Pro ($100/mo) | 50 msgs/week (shared w/ Sol Pro) | Full existing allowance | Live |
| Pro ($200/mo) | 200 msgs/week | Full existing allowance | Live |
| Business Standard | Not included | 15 msgs/month | Rolling out |
| Business Premium | 50 msgs/week | Full existing allowance | Live |
| Enterprise | Admin-controlled | Admin-controlled | Off by default at launch |
The safety story running underneath the launch
OpenAI says Astra ships with additional safety monitoring designed to catch situations where an agent may have misread a user’s instructions. When that happens, the company says a conversation can be automatically paused for the user to review before the agent continues acting. That’s a meaningful design choice given the benchmark numbers involved: a model that scores 100% on ExploitBench and clears what OpenAI has previously called a “critical” threshold in cybersecurity capability is also a model capable of causing real damage if it acts on a misread instruction inside a connected system.
OpenAI has also built a separate, more heavily gated cybersecurity-capable configuration referred to in reporting as “Daybreak,” which requires Enterprise-level access and is off by default for the first two weeks after launch. Figures tied to that configuration describe a jump in proof-of-concept exploit completion from a baseline of 2.4% up to 92% when the Daybreak configuration is enabled, and cyber red-teaming task completion rising from 7.4% to 76.9%. Even with Daybreak active, completion of arbitrary, open-ended cyber requests reportedly stays capped around 3.5%, which OpenAI appears to be using as evidence that the model’s most dangerous capabilities remain bounded rather than open-ended.
These figures come from third-party reporting on OpenAI’s own disclosures rather than an independent audit, and they should be read with that caveat. Still, the gap between a 2.4% baseline and a 92% figure under a gated configuration is large enough that it explains why OpenAI built an opt-in, admin-controlled access model around this specific capability rather than shipping it open by default the way it has with most past releases.
How Astra compares with GPT-5.6 Sol
GPT-5.6 Sol, which OpenAI had positioned as its most capable model prior to this release, supported a 1-million-token context window, multi-agent orchestration, and an “Ultrafast” mode capable of running up to 14 times faster than standard processing in certain API tiers. Astra doesn’t appear to compete on raw throughput in the same way; instead, OpenAI’s own materials frame the upgrade around capability and reliability on complex tasks rather than a straightforward speed claim, apart from the roughly 2x computer-use speedup and the 47%-less-time result on OSWorld 2.0 benchmarking.
That’s a different kind of pitch than prior OpenAI launches, which tended to lead with either raw benchmark gains or dramatic price cuts. Astra’s positioning is closer to “does more kinds of professional work reliably” than “is simply faster or cheaper,” and the pricing (higher than Sol’s published rates in several public comparisons) reflects that. For teams already invested in GPT-5.6 Sol, Terra, or Luna workflows, Astra looks like an optional upgrade tier rather than a wholesale replacement, at least during this staged rollout period.
Competitive pressure: Claude, Gemini, and the coding-assistant market
The timing matters. OpenAI’s push into “best model for software engineering” territory comes directly after Anthropic’s own comparisons showing Claude Opus 4.8 beating GPT-5.6 Sol on coding benchmarks, a gap tech-insider.org previously covered in detail. Astra’s FrontierMath and ARC-AGI-3 scores are OpenAI’s most direct attempt yet to reclaim the “best coding model” narrative, and the Codex-specific context-preservation feature is aimed squarely at the kind of long, multi-file agentic coding sessions where Anthropic’s Claude Code and Cursor’s own agent tooling have picked up developer mindshare over the past year.
Google’s Gemini line remains the other major pressure point, particularly on pricing. Astra’s $10-per-million-input, $50-per-million-output API rate sits well above the aggressive per-token pricing Google has used to undercut OpenAI on Gemini Flash-tier models. OpenAI appears to be betting that enterprise buyers will pay a premium for a model marketed as safer and more reliable on agentic tasks, rather than competing purely on cost per token, a dynamic already visible in how ChatGPT, Claude, and Gemini free tiers currently stack up.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Source framing |
|---|---|---|---|
| FrontierMath Tier 4 | 98% | Not disclosed at this level | OpenAI benchmark release |
| ARC-AGI-3 | 99.9% | Not disclosed at this level | OpenAI benchmark release |
| ExploitBench | 100% | Not disclosed at this level | OpenAI benchmark release |
| OSWorld 2.0 (computer use) | 72.6% | 65.7% | Reported comparison, per-task time down ~47% |
| Context window | ~1.05M tokens | 1M tokens | Public API documentation |
Historical context: how fast OpenAI is moving
The pace here is notable on its own. GPT-5 launched in August 2025. GPT-5.6 Sol followed as an update within the GPT-5 family. GPT-6 Astra now arrives as a full generational jump only weeks after Sol, which is a faster cadence than the roughly year-plus gaps that separated GPT-3 from GPT-4, and GPT-4 from GPT-5. Whether “GPT-6” is a meaningful architectural leap or largely a branding decision to reset the narrative after a bruising few months of safety headlines is something outside reporting hasn’t settled, and OpenAI hasn’t published the kind of technical detail that would let independent researchers verify it either way.
It’s also worth noting that this launch follows a period in which OpenAI reportedly paused aspects of its own training process over internal alarm about a prior model’s capabilities, and separately faced state-level regulatory inquiries tied to unrelated security incidents. Astra’s staged, permission-gated rollout, especially the admin-controlled Enterprise default and the separate Daybreak cyber configuration, reads as a direct response to that scrutiny rather than a coincidence. Prediction-market traders had already been pricing in an aggressive capability jump before launch, with one Manifold market on ARC-AGI-3 performance swinging sharply on GPT-6 Astra chatter in the days before the announcement.
Market and industry reaction
No stock-specific market reaction to the Astra announcement had been confirmed in reporting as of publication; OpenAI remains privately held, and public reactions have so far focused on rival AI labs and vendors rather than tradeable securities. That said, the launch adds to a broader pattern this year of AI-model announcements moving share prices for companies exposed to the AI infrastructure buildout, from GPU makers to cloud providers, even when the announcement itself doesn’t directly involve those companies.
Forbes contributor coverage, per Forbes, characterized the initial announcement process itself as unusually messy, noting confusion in the early hours of the rollout about exactly what had shipped and to whom. That kind of stumble, even a minor communications one, tends to get magnified when a company is simultaneously trying to claim its release marks the start of an “AGI era.” It also lands just as reports of two unreleased Claude models leaking ahead of Anthropic’s reported $100B IPO push put every major lab’s release calendar under extra scrutiny.
What developers should actually do this week
For teams that want to test Astra in Codex, the practical first step is simply updating tooling: Codex CLI needs to be at version 0.153.0 or newer, and the ChatGPT desktop app needs to be current, before Astra shows up as an option at all. Azure customers on usage Tier 5 or 6 reportedly get default quota access to Astra, while lower tiers need to explicitly request quota, according to Kingy AI‘s review of the rollout documentation.
Given the credit-based pricing inside ChatGPT Work and Codex versus the flat per-token API pricing, teams running high-volume automated workflows will likely want to budget separately for API access rather than relying on ChatGPT plan allowances, which are explicitly described as limited for anyone below the top Pro and Business Premium tiers.
Five predictions for the next 90 days
1. Independent benchmarking will chip away at the headline numbers. A 100% ExploitBench score and a 99.9% ARC-AGI-3 score are the kind of results that invite third-party researchers to look for edge cases and benchmark contamination, and that scrutiny typically arrives within a few weeks of any major model launch.
2. Plus-tier access will lag longer than the “coming days” framing suggests. Staged rollouts that start with Pro and Enterprise tiers have historically taken several weeks, not days, to fully reach Plus subscribers in past OpenAI launches.
3. Anthropic and Google will respond with their own coding-focused updates. Given the direct framing around “best model for software engineering,” expect competitive responses aimed specifically at Codex’s context-preservation feature and the OSWorld 2.0 computer-use numbers.
4. The Daybreak cyber-configuration gating will draw regulatory attention. A model that can jump from 2.4% to 92% completion on proof-of-concept exploits under a specific configuration is likely to attract questions from policymakers already scrutinizing AI cyber capability, regardless of how tightly OpenAI has gated access.
5. Pricing will compress within two quarters. OpenAI’s pattern with prior frontier models has been to launch at a premium price and cut rates significantly once a next-generation model is ready to take over the top tier; expect Astra’s $10/$50 per-million-token rate to look expensive by mid-2027.
The bottom line
GPT-6 Astra is a real, shipping product change, not vaporware: it’s live for Pro, Enterprise, and Business Premium users right now, with concrete pricing, a specific Codex CLI version requirement, and benchmark numbers that OpenAI is willing to put its name behind. But the gap between OpenAI’s “AGI era” framing and what’s actually verifiable from outside the company remains wide. The most durable, checkable change in this release is the Codex context-preservation feature; everything else, from the benchmark ceiling scores to the safety-monitoring claims, will need independent verification over the coming weeks before it’s fair to take at face value.
Frequently asked questions
What is GPT-6 Astra?
GPT-6 Astra is OpenAI’s newest model, announced September 3, 2026, and positioned as an upgrade to ChatGPT and Codex focused on coding, computer use, research, and complex multi-step tasks.
When does GPT-6 Astra become available?
Rollout began with a limited set of organizations on September 3, 2026. By September 4, OpenAI confirmed availability for all Pro, Enterprise, and Business Premium users in ChatGPT Work, Codex, and the API, with Plus and Business Standard users still waiting on a gradual rollout.
How much does GPT-6 Astra cost through the API?
Standard API pricing is $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million and a Fast mode priced at double the standard rate.
Do I need to update Codex to use Astra?
Yes. Using Astra in Codex requires Codex CLI version 0.153.0 or newer, along with the latest version of the ChatGPT desktop app.
How does GPT-6 Astra compare to GPT-5.6 Sol?
OpenAI positions Astra above GPT-5.6 Sol in capability, citing an OSWorld 2.0 computer-use score of 72.6% versus 65.7% for Sol, plus new benchmark results on FrontierMath Tier 4 and ARC-AGI-3 that OpenAI says exceed what Sol achieved.
Is GPT-6 Astra available on Plus plans?
Plus plans do not include GPT-6 Pro in Chat at launch but do get limited Astra access in ChatGPT Work and Codex, with the option to purchase additional usage credits.
What safety measures does Astra include?
OpenAI says Astra includes monitoring that can pause a conversation automatically if the system detects that an agent may have misinterpreted a user’s instructions, prompting a manual review before the agent continues.
Is this the start of “AGI,” as OpenAI has suggested?
That framing comes from OpenAI’s own announcement language, as reported by outlets including Breitbart. Independent researchers have not verified the claim, and “AGI” remains a term without a single agreed-upon technical definition across the industry.
