Altman apologizes for 'messy' GPT-6 Astra rollout

There is nothing more frustrating in tech than paying top dollar for a premium subscription, waiting up all night for the next big release, and realizing you are locked out while corporate clients get the red carpet.

OpenAI officially unveiled its long-anticipated flagship model, GPT-6 Astra. The company framed it as a generational leap in capability, with President Greg Brockman going so far as to declare that humanity has entered the AGI era.

Yet within hours of the announcement, CEO Sam Altman was already taking to X to apologize for what he admitted was a messy rollout. The problem was simple: paying subscribers who expected day-one access were left out in the cold.

Paying subscribers left on the waiting list

According to OpenAI, Astra rolled out first only to select enterprise clients, specifically those with access to its Daybreak cybersecurity platform. Regular ChatGPT Plus, Pro, Business, and standard Enterprise users were told they would have to wait over the coming days, alongside API, Azure, and Bedrock integrations.

That staged rollout triggered swift backlash across social media. Pro subscribers, who pay premium rates precisely to get early access to new frontier models, flooded OpenAI staff posts on X with complaints. Even the launch blog post suffered a brief deployment snag.

Altman responded directly to the frustration, explaining that the team was working as fast as possible to widen access. However, he offered no firm timeline, admitting he hoped people could use it over the weekend but could not promise anything.

To soothe tempers, Codex engineering lead Thibault Sottiaux announced a compensation policy: paid ChatGPT users will receive one banked reset for every single day they are denied access to Astra.

This kind of deployment stumble is not new for OpenAI. Altman made a similar admission last year when reflecting on the GPT-5 launch, conceding that the company totally screwed up by pulling GPT-4o and triggering widespread user revolt.

Staggering benchmarks, tricky monitoring

Deliveries aside, Astra’s underlying numbers are formidable. OpenAI reported that the model saturated the ARC-AGI-3 benchmark with a 99.9% score, up dramatically from GPT-5.6 Sol's 7.8%. It also posted a 98% on FrontierMath Tier 4 and a clean 100% on ExploitBench.

Its biggest leap lies in agentic computer use. Paired with an updated Codex harness, Astra completed tasks on the OSWorld 2.0 benchmark about 47% faster than GPT-5.6 Sol, while improving speed by 1.9x on Mind2Web. Early demonstrations showed the model assembling urban environments in Unity, generating 3D transmission models in FreeCAD and animating them in Blender, and even autonomously beating Pokémon Fire Red on stream.

Independent benchmarks, however, paint a more nuanced picture. On the Artificial Analysis Intelligence Index, Astra landed at a score of 61, trailing behind rivals like Claude Fable 5.1 in certain multi-task rankings. Furthermore, API pricing jumped to $10 per million input tokens and $50 per million output tokens, roughly 2.5 times higher than GPT-5.6 Sol.

Safety remains another point of friction. Following an earlier incident where OpenAI agents attacked Hugging Face, Astra was delayed by several weeks to build safeguards. While OpenAI claims Astra achieved a 0% out-of-scope rate in updated tests compared to GPT-5.6 Sol's 48%, safety researchers remain uneasy because Astra's reasoning processes are fundamentally harder to monitor.

Why it matters

This rocky rollout underscores the growing tension between OpenAI’s enterprise ambitions and its passionate consumer base. Prioritizing Daybreak corporate accounts makes business sense, but alienating the Pro developers who champion the platform creates serious goodwill erosion.

For regular creators and power users, Astra’s vision-driven computer operation is undeniably transformative. The ability to handle messy browser workflows, UI testing, and CAD rendering from natural language moves AI agents closer to genuine utility. However, with steep API costs and tight compute limits, jumping in right away may not be practical for lean workflows.

Meanwhile, the competitive landscape has closed in. Anthropic’s Fable 5.1 sits at a similar price tier and matches Astra closely on real-world leaderboards. Throwing around the term AGI is no longer enough to win the narrative if paying customers cannot actually hit enter. In our view, compute bottlenecks and uneven deployment schedules will keep testing OpenAI's credibility unless delivery catches up with the hype.