[KongchangAI]
· 3 min read· 1,671 words

GPT-6.2 Dual-Model Release: Pricing, Benchmarks, and Controversy Around Astra and Sol

GPT-6.2 Dual-Model Release: Pricing, Benchmarks, and Controversy Around Astra and Sol

OpenAI releases the GPT-6.2 Astra and Sol dual models, with price cuts and cache discounts as highlights, though benchmarks and interaction experience remain controversial.

OpenAI launches the dual models GPT-6.2 Astra and GPT-6.2 Sol, with the former for high-difficulty projects and the latter for efficient everyday tasks; LLM 0.38 already supports both. Astra is priced at $8/$40 per million tokens (80% of 6 Astra and on par with Claude Fable 5.2), while cached pricing drops to $0.40; Sol remains at $2/$10. Simon Willison's pelican test shows Astra is more detailed at high effort tiers, but it does not clearly outperform Claude Fable 5.2 or Gemini 4 Carbon on the Yorkshire mecha task. Community feedback notes that Astra can block long tasks by asking incomprehensible internal questions. Overall, choose by task complexity and run small-scale A/B tests on real tasks.

OpenAI has officially launched two new models, GPT-6.2 Astra and GPT-6.2 Sol, targeting high-difficulty projects and everyday tasks respectively. Developer Simon Willison was quick to add support in the LLM 0.38 tool and conducted hands-on testing across three dimensions: pricing, benchmark performance, and community feedback. This article draws on his first-hand evaluation and publicly released information to outline the key highlights of this dual-model release.

GPT-6.2 Astra and Sol release information

Dual-Model Positioning: A Combination of Frontier and Efficiency

According to OpenAI's official description, GPT-6.2 Astra is a new frontier model aimed at "the hardest projects," while GPT-6.2 Sol is an efficient, versatile everyday model. They are available in the API as gpt-6.2-astra and gpt-6.2-sol respectively, and Simon Willison's LLM 0.38 has already added support for both.

This "flagship + lightweight" dual-track strategy has become increasingly common in recent model releases: Astra handles complex reasoning and long tasks, while Sol covers high-frequency, low-cost everyday calls. For developers, choosing between the models depends more on task complexity and cost budget than on a simple assumption that "the new version is always better."

The technical foundation behind the "flagship + lightweight" dual-track strategy is scale tiering within the same model family: the flagship model has more parameters and invests more compute at inference time (reflected in the effort tiers discussed below), making it suitable for multi-step reasoning, long context, and complex tool calls; the lightweight model trades a smaller activation scale for low latency and low cost, covering high-frequency scenarios such as classification, summarization, and simple code generation. The API interfaces of the two are typically kept consistent, so developers can switch model names based on the task without changing their calling code—which is also why LLM 0.38 in this article can support both models simultaneously. It is worth noting that the lightweight model is not a "compressed replica" of the flagship model; its gaps in long-chain reasoning and instruction following often only become apparent in complex tasks, so it is safer to select models based on task complexity rather than version recency.

Pricing Strategy: The Signal Value of Price Cuts

Pricing is one of the most noteworthy parts of this release. According to Simon Willison's records, GPT-6.2 Astra is priced at $8/MTok input and $40/MTok output, just 80% of the previous-generation 6 Astra ($10/$50), putting it at the same price as Claude Fable 5.2—which is likewise 80% of its predecessor's price. The cached price drop is even larger, at $0.40 per million tokens versus $1 per million tokens for the previous generation.

GPT-6.2 Sol maintains the $2/$10 price, on par with 6.1 Sol and 6 Sol, and half that of GPT-5.6 Sol ($4/$20).

These numbers reveal two things: first, the unit cost of frontier models is continuing to decline, and the steep discount on cached pricing is especially beneficial for long-running agentic sessions; second, the Sol series' pricing has remained stable across multiple generations, indicating that the pricing range for lightweight models is maturing.

Prompt caching is the key mechanism for understanding the price figures in this article: when multiple requests share the same prefix content (such as system prompts, long documents, tool definitions, or conversation history), the server can reuse the already-computed intermediate state and charge a price far below the standard input rate for those tokens. This is precisely why the cached price of $0.40/MTok is 20 times lower than the standard input price of $8/MTok. The defining characteristic of agentic sessions is that each turn resubmits an ever-growing history context, the vast majority of which is repeated prefix, so the caching discount directly scales to the total cost of the entire session—actual savings in long-task scenarios may be far higher than the headline per-unit price reduction suggests. When measuring the cost of such applications, "cache hit rate" should be treated as a variable just as important as input/output unit prices.

Benchmark Tests: Pelican Riding a Bicycle and the Yorkshire Mecha

Simon Willison's classic evaluation project is having the model draw "a pelican riding a bicycle." He provided side-by-side comparisons of GPT-6.2 Sol and GPT-6.2 Astra, showing differences across the high, xhigh, and max effort tiers.

Based on the comparison descriptions, gpt-6.2-astra produced the most detailed pelican at the Max tier: layered feathers, a fluttering red scarf, and one wing resting on the handlebars, at a cost of 88.71¢. The Xhigh tier (17.70¢) added a green hat and retained the seaside scene, while the High tier (11.55¢) simplified the feathers but preserved the riding posture. gpt-6.2-sol showed a convincing pedaling posture at the Max (13.69¢) and Xhigh (6.74¢) tiers, with one foot high and one foot low and wings extended forward; at the High tier (2.41¢), the bird's body sat lower and the handlebars became less recognizable.

Another test asked the models to use Three.js and OpenStreetMap data to create an animation of "Yorkshire transforming into a giant mecha, stomping Manhattan, Queens, and Brooklyn." Simon's assessment was that both models did poorly on the "Yorkshire transforming into a mecha" segment—at least compared with Claude Fable 5.2 and Gemini 4 Carbon. He also used GPT-6.2 Astra (Ultra tier) in the Codex desktop app to stitch the two models' animations into a comparison video and attached the exported conversation records.

This comparison suggests that the new models are steadily improving on image-generation-style tasks, but on complex 3D animation and spatial reasoning tasks, they have not yet opened a clear gap over competitors.

The effort tiers (high / xhigh / max / ultra) that appear repeatedly in this article are "thinking budget" control parameters for reasoning models: they do not change the model weights but determine how many reasoning tokens and how much internal computation the model is allowed to consume before giving a final answer. The higher the tier, the more the model tends to engage in multi-step self-checking and try more approaches, and output quality and detail usually improve accordingly—but latency and cost also rise roughly linearly. This is precisely why the same task is quoted from 2.41¢ to 88.71¢ across different tiers. A reasonable practice is to first run the workflow at a low tier, then retry failed samples individually at higher tiers rather than using the highest tier globally. In addition, "pelican riding a bicycle" has long been used by the LLM community as a capability probe, because the scene simultaneously tests spatial relationships, object-part correspondence, and common-sense constraints—making it hard to fake with a single metric.

Community Feedback: A Mismatch Between Capability and Communication Style

Early reactions to the release were mixed. According to Simon Willison's observations, GPT-6.2 Astra did not surpass Claude Fable 5.2 on multiple benchmarks, and some users complained that its communication style was confusing.

Theo Browne's feedback letter to OpenAI during the pre-release phase provides a typical example: when handling long tasks, the model can dispatch many sub-agents and manage them reasonably well, but it will suddenly ask the user about internal implementation details that "you could not possibly understand under any circumstances on Earth, let alone have an opinion on." He gave an example: after issuing "/goal make a new level for Outer Wilds," the task was "blocked" two hours later because the model wanted to ask him, "Does the reviewer for r24 write its clips to 24_g521.mp4 or 24_p11.mp4?"

This case highlights a real pain point in agentic workflows: while sub-agent management capabilities improve, the model's judgment about "when to ask and what to ask" is still immature, making it easy to throw internal details at users and cause unnecessary blocking. For developers, this means that when using Astra for long tasks, clearer boundaries for autonomous decision-making need to be set.

Sub-agent orchestration refers to the main model decomposing a long task, spawning multiple subtask executors with independent contexts to handle them in parallel or serially, and then aggregating the results. This architecture can significantly expand the manageable scale of a single task, but it also introduces new failure modes: the main model must judge which decisions it can make autonomously and which must be confirmed with the user. When internal implementation details (such as intermediate file naming) are misjudged as matters "requiring user decision," the entire task stalls waiting for a meaningless answer. Mitigation measures include explicitly defining autonomous decision boundaries in the system prompt, setting default value rules for sub-agent outputs, and configuring automatic fallback decisions after a timeout for "blocking-type questions." This also shows that when evaluating agent capabilities, interaction quality metrics such as "invalid question rate" should be considered alongside task completion rate.

Model Selection Recommendations

Overall, GPT-6.2 Astra is suitable for long-task scenarios that demand deep reasoning and can tolerate higher unit prices and some communication friction; GPT-6.2 Sol covers everyday calls at a stable low price. The sharp drop in cached pricing is an underappreciated highlight of this release and has real cost implications for continuously running agent applications. As for benchmark-level controversy, developers are advised to run small-scale A/B tests on their own real tasks rather than making decisions solely based on public leaderboards.

One-line Summary

OpenAI releases the GPT-6.2 Astra and Sol dual models, with price cuts and cache discounts as highlights, though benchmarks and interaction experience remain controversial.

Paragraph Summary

OpenAI has launched the dual models GPT-6.2 Astra and GPT-6.2 Sol, with the former targeting high-difficulty projects and the latter focused on efficient everyday tasks; LLM 0.38 already supports both. On pricing, Astra is $8/$40 per million tokens, 80% of the previous-generation 6 Astra and on par with Claude Fable 5.2; cached pricing drops to $0.40, a significant reduction; Sol remains at $2/$10, half that of GPT-5.6 Sol. Simon Willison's pelican-on-a-bicycle test shows Astra produces richer detail at high effort tiers, but on the Yorkshire mecha animation task it does not clearly outperform Claude Fable 5.2 or Gemini 4 Carbon. Community feedback notes that Astra, during long tasks, raises internal implementation questions that users cannot understand, causing blocking. Overall, it is recommended to select models by task complexity and run small-scale A/B tests on real tasks.

#GPT-6.2 Astra#GPT-6.2 Sol#OpenAI pricing#model benchmarks#prompt caching#AI agent
Share:

Related articles