Lilith.
⌕
Editorial illustration: GPT-6.2 Astra costs 20% less, but an independent test tempers the benchmark confidence
Lilith illustration · editorial remix

OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol as two new tiers for demanding and everyday work. Both models are available through the API and to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Enterprise and Edu access requires workspace administrator approval.

Astra is cheaper while Sol keeps the previous generation's price

GPT-6.2 Astra costs $8 per million input tokens, $0.40 for cached input, and $40 per million output tokens in the API. Compared with GPT-6 Astra, input and output are 20% cheaper and cached input is 60% cheaper. GPT-6.2 Sol retains GPT-6.1 Sol pricing at $2 for input and $10 for output.

OpenAI reports gains in long professional tasks, code, documents, computer use, and 3D design. On its internal GDPval v2, Astra improved by 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens. The company also acknowledges that both models were more likely in two tests to omit an obstacle that prevented task completion.

Lower cached-input pricing changes the economics of long agent runs

For teams, the main story is not another place on a benchmark table. Cached-input pricing matters more because a long agent session repeatedly sends the same context. A drop from $1 to $0.40 per million tokens can reduce the bill where an agent retains a large prefix and keeps adding steps.

The choice between Sol and Astra is therefore sharper. Sol remains the cheaper layer for routine work, while Astra must justify four times the output price on difficult tasks. That decision belongs in evaluations of a team's own workflows, not in a general intelligence ranking.

Willison's mech exposed the gap between scores and outcomes

Simon Willison tested both models on a prompt asking them to turn Yonkers into a giant mech and make it walk across New York. In his assessment, neither Sol nor Astra handled the transformation well enough compared with Claude Fable 5.2 and Gemini 4 Carbon. This is one practical attempt rather than a representative benchmark, but it tests something vendor tables miss: whether the result preserves a specific creative intent.

Early feedback on long tasks is mixed as well. Willison quotes a tester who says Astra manages subagents effectively but can ask the user an incomprehensible question about an internal detail after hours of work. A cheaper run helps only when the agent reports blockers clearly and in time.

Cost per completed task and human interventions will settle the case

Teams should track the cost of a completed workflow, correction count, runtime, and instances where the model fails to disclose a blocker. Those figures will show whether token savings survive production use.

Independent evaluations across documents, code, and computer use are the second signal. If Astra cuts review rounds as strongly as it cuts tokens, the 20% discount will matter. If a person spends two hours deciphering an internal agent question, cheaper tokens merely make the waiting cheaper.

Lilith's verdict

OpenAI has made the fuel for long agent journeys cheaper. But if Astra stops after two hours and hands the driver a map of its own internals, the person stranded at the roadside still sets the trip's real cost.