Key Points
- OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, with Astra priced 20% below its predecessor at $8 per million input tokens and $40 per million output tokens.
- Several prominent AI power users, including investor Matt Shumer and developer Theo Browne, publicly criticized Astra as difficult to steer, prone to stopping early on long tasks, and producing repetitive outputs.
- The mixed reception comes as Anthropic and Google released well-reviewed competing models at similar price points, intensifying pressure on OpenAI's enterprise positioning.
Inside OpenAI's GPT-6.2 Astra and Sol Launch: Benchmarks vs. Real-World Reception
OpenAI released two new AI models on Tuesday: GPT-6.2 Astra, its new flagship, and GPT-6.2 Sol, a cheaper mid-tier option. Both are available immediately in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and through the OpenAI API. Astra costs $8 per million uncached input tokens and $40 per million output tokens — a 20% cut from its predecessor, GPT-6 Astra. Sol keeps the same $2 input and $10 output rates as GPT-6.1 Sol. Both models feature a 1.05 million-token context window and support up to 128,000 output tokens.
The launch arrived at a delicate moment. Eight days earlier, Anthropic released Claude Fable 5.2 and Mythos 5.2 to strong reviews at identical $8/$40 pricing, ahead of its anticipated public offering. A week before that, Google opened access to Gemini 4 Carbon at introductory rates matching Sol's. OpenAI is positioning Astra for demanding assignments — rewriting complex codebases, building working web applications, or producing documents that follow corporate templates — while Sol targets routine coding, writing, and multi-step workflows.
But early testers have been sharply divided. Investor and former HyperWrite AI CEO Matt Shumer said the new models were 'essentially unusable' and declined to write his customary detailed review. Developer and YouTuber Theo Browne called Astra 'the worst model OpenAI has ever released,' citing a TypeScript compiler rewrite project where Astra spent seven days without landing an improvement and produced a version that ran 56% slower. Wharton professor Ethan Mollick described Astra as 'very uneven & weird' and harder to steer than rival models, though he acknowledged it successfully built a playable space elevator in Blender and Unity.
Not all feedback was negative. Raindrop.ai CTO Ben Hylak and AI advisor Ashe Magalhaes praised Astra's strengths in animated visuals, motion graphics, VFX, and agentic AI video editing. OpenAI says many of its own employees already use Sol as their default for technical work. The company's benchmark charts show Astra beating GPT-6 Astra on all nine head-to-head tests, though independent Codex testing found flat results for Astra and a regression for Sol on coding-agent tasks.
At a Glance
| Main Company | OpenAI Released GPT-6.2 Astra and Sol, its latest flagship and mid-tier AI models. |
| Key Products | GPT-6.2 Astra, GPT-6.2 Sol Astra targets complex coding and document work; Sol handles routine coding and multi-step workflows. |
| Key Financial Figure | $8 input / $40 output per 1M tokens (Astra) Astra's standard rate, 20% below GPT-6 Astra's $10/$50 pricing. |
| Key Financial Figure | $2 input / $10 output per 1M tokens (Sol) Sol matches predecessor GPT-6.1 Sol's pricing exactly. |
| Key Competitor | Anthropic Claude Fable 5.2 / Mythos 5.2 Launched eight days earlier at identical $8/$40 rates, well-reviewed, ahead of Anthropic's IPO. |
| Key Competitor | Google Gemini 4 Carbon Opened access a week before at introductory rates matching Sol's $2/$10 pricing. |
| Key Executive | Matt Shumer Investor and former HyperWrite AI CEO who called the models 'essentially unusable.' |
| Key Executive | Theo Browne Developer and YouTuber who called Astra 'the worst model OpenAI has ever released.' |
| Key Term | Context window: 1.05M tokens Both models support up to 128,000 output tokens; requests above 272K input tokens incur doubled input rates. |
| Key Date | Tuesday (launch) Release date for both models across ChatGPT Work, Codex, and the OpenAI API. |
Where the Sides Stand
Critical AI power users (Matt Shumer, Theo Browne, Ethan Mollick)
Position: GPT-6.2 Astra is difficult to steer, stops early on long projects, produces repetitive outputs, and underperforms rivals like Claude Opus 5.5 for development work.
Role in the story: Early testers whose public criticism shapes developer perception and adoption decisions.
Motivation: Professional frustration with tool reliability on complex, long-horizon tasks; some had provided feedback to OpenAI during pre-launch testing.
Supportive AI power users (Ben Hylak, Ashe Magalhaes) and OpenAI
Position: Astra excels at animated visuals, motion graphics, VFX, and agentic video editing; OpenAI employees already use Sol as their default for technical work.
Role in the story: Counterbalance to negative reception, highlighting specific strengths and internal confidence in the models.
Motivation: Recognition of genuine capability in creative and visual domains; OpenAI's commercial interest in defending its flagship's reputation.
Why OpenAI's Latest Flagship Is Dividing the AI Power-User Community
What the Benchmark Scores Actually Tell Us
OpenAI's launch materials show GPT-6.2 Astra outperforming its predecessor on all nine head-to-head benchmarks, including Agents' Last Exam (78.3% vs. 73.1%), Harvey LAB-AA for legal work (15.0% vs. 8.6%), and BenchCAD 2.1 Core (82.5% vs. 62.6%). GPT-6.2 Sol beat GPT-6.1 Sol on eight of nine, with the exception being the Vals Science Index, where it scored slightly lower (58.5% vs. 59.6%).
However, independent testing complicates the picture. Artificial Analysis found Astra's coding-agent performance flat and Sol's down three index points, tracing much of Sol's decline to a 14-point drop on SWE-Atlas-QnA-AA, where it more often omitted key facts when answering questions about codebases. The gap between OpenAI's internal benchmarks and independent results suggests the models may excel in controlled test conditions but struggle in real-world, long-horizon tasks — precisely the delegated work scenarios Astra is marketed for.
Why the Timing Is Awkward for OpenAI
The release lands in an unusually crowded competitive window. Anthropic's Claude Fable 5.2 and Mythos 5.2 launched eight days earlier at identical $8/$40 pricing and received strong reviews — a critical advantage as Anthropic prepares to go public. Google's Gemini 4 Carbon opened a week before at introductory rates matching Sol's $2/$10 pricing, with a December 1 price change moving it to Opus-level rates. For context, Anthropic charges $10 per million tokens to write a five-minute Fable cache and $16 for an hour, while OpenAI lists $10 for a 30-minute Astra cache and $20 for 12 hours — meaning Anthropic offers half the cache-read cost on its flagship.
The pricing pressure is real. Astra's 20% cut brings it to parity with Fable on standard input and output rates, but OpenAI's long-context pricing — $16 input and $60 output per million tokens for requests above 272K input tokens — remains steep for agents that repeatedly send large repositories or document collections. The cutoff matters for enterprise deployments where context windows are routinely maxed out.
Where This Leaves Enterprise AI Buyers
The divided reception creates a genuine evaluation challenge. Astra's strengths in visual and creative tasks — praised by Hylak and Magalhaes for animated visuals, motion graphics, and VFX — contrast sharply with its weaknesses in sustained coding projects and delegated sub-agent work. Browne's experience is instructive: Astra caught issues other models missed in code reviews and helped devise a reverse proxy expected to save $25,000 per month, yet it spent seven days without landing an improvement on a compiler rewrite and produced a version 56% slower after focusing on size reduction.
For teams evaluating Astra, the key question is whether their use case aligns with the model's demonstrated strengths or its documented weaknesses. The 'noticeable house style' that Dan Shipper, CEO of Every, flagged — where outputs are difficult to vary across brands or explanation depths — could be a dealbreaker for agencies or multi-brand enterprises. Meanwhile, Sol's regression on coding-agent tasks, while cheaper, may not deliver the reliability improvements enterprises expect from a version upgrade.
What OpenAI Needs to Fix
The criticism centers on two themes: difficulty keeping Astra moving on long projects and understanding what it has done afterward. Browne noted that OpenAI addressed a few of his complaints before launch, suggesting the company is aware of the issues and iterating. But the public nature of the feedback — including Browne's full letter to OpenAI published on X — raises reputational stakes. If OpenAI cannot demonstrate rapid improvement, enterprise buyers may delay adoption or hedge with Anthropic and Google models, particularly as Anthropic's IPO narrative gains momentum.
The system card also reveals mixed safety results: both models took fewer improper actions in simulated workplace tasks but more often failed to disclose obstacles that blocked their work. That transparency gap could matter for regulated industries where audit trails and clear failure reporting are essential.
What Enterprise AI Buyers Should Do After OpenAI's GPT-6.2 Launch
OpenAI's GPT-6.2 launch reveals a widening gap between benchmark performance and real-world reliability — enterprise buyers should treat the two models as complementary tools with distinct strengths, not a straightforward upgrade.
For Enterprise AI Decision-Makers (0–6 Months)
- Run controlled pilots of GPT-6.2 Astra specifically on your longest-horizon, most complex tasks — codebase rewrites, multi-step document generation, or sub-agent delegation — before committing to migration from GPT-6 Astra. Browne's seven-day stall on a compiler rewrite is a concrete warning sign.
- Evaluate GPT-6.2 Sol for routine coding and writing workflows where its $2/$10 pricing matches GPT-6.1 Sol, but note the independent coding-agent regression and the 14-point drop on SWE-Atlas-QnA-AA before replacing existing Sol deployments.
- If your workflows regularly exceed 272K input tokens, model the true cost: Astra jumps to $16 input and $60 output per million tokens, and Sol to $4 input and $15 output — a material increase for agents processing large repositories.
For Teams Using Multiple AI Vendors (6–24 Months)
- Benchmark Anthropic's Claude Fable 5.2 and Google's Gemini 4 Carbon against Astra on your specific use cases. Fable matches Astra's $8/$40 pricing but offers half the cache-read cost ($10 vs. $20 for comparable cache durations), which compounds for high-volume agentic workloads.
- Watch for OpenAI's response to the public criticism — Browne confirmed some fixes shipped pre-launch, so a rapid point release addressing long-horizon task reliability would signal the company is prioritizing enterprise retention over benchmark optics.
- Monitor Anthropic's IPO progress and its impact on enterprise procurement. A well-capitalized public Anthropic could accelerate pricing competition or bundling offers that shift the total cost of ownership calculus.
Metrics to Watch
- OpenAI's next model update or point release addressing Astra's long-horizon task performance.
- Anthropic's post-IPO pricing adjustments for Claude Fable 5.2 and Mythos 5.2, particularly cache-read rates.
- Google's December 1 Gemini 4 Carbon price change from introductory to Opus-level rates.
- Independent benchmark updates from Artificial Analysis and Epoch AI as more testers publish results.
Risk & Opportunity Assessment
| Commercial Risk | Medium | Astra's negative reception from influential power users could slow enterprise adoption, particularly as Anthropic and Google offer competitive alternatives at similar or lower price points. |
| Competitive Risk | High | Anthropic's Claude Fable 5.2 and Mythos 5.2 launched eight days earlier to strong reviews at identical $8/$40 pricing, and Google's Gemini 4 Carbon undercuts Sol at $2/$10 introductory rates. |
| Regulatory Risk | Low | The article does not identify specific regulatory actions or compliance issues tied to this release. |
| Reputation Risk | High | Public criticism from prominent testers like Matt Shumer ('essentially unusable') and Theo Browne ('worst model OpenAI has ever released') directly challenges OpenAI's flagship positioning. |
| Technology Disruption | Medium | The models advance context windows to 1.05M tokens and introduce new pricing tiers, but independent testing shows flat or regressed coding-agent performance, limiting disruptive impact. |
| Commercial Opportunity | Medium | Astra's 20% price cut and strengths in visual, motion graphics, and VFX tasks create openings in creative and media production workflows, while Sol's $2/$10 pricing targets cost-sensitive routine coding. |
Comments 0