Field noteNov 18, 20265 min read1 source

GPT‑6.2 Astra and GPT‑6.2 Sol

In one minute

OpenAI has introduced GPT-6.2 Astra for its hardest projects and GPT-6.2 Sol for everyday work, with both models accessible through its API. Astra’s listed input price is 20% lower than its predecessor, and its cached-input price falls from $1 to $0.40 per million tokens—an important cost change for teams running long agent sessions.

What happened

OpenAI describes GPT-6.2 Astra as a frontier model for difficult projects and GPT-6.2 Sol as an efficient, versatile model for everyday work. The release makes them available in the API under the names gpt-6.2-astra and gpt-6.2-sol. Simon Willison reports that he added support for both models in version 0.38 of his LLM command-line tool.

Willison reports Astra pricing of $8 per million input tokens and $40 per million output tokens. That is 80% of the previous GPT-6 Astra’s $10/$50 rates. Cached input costs $0.40 per million tokens, compared with $1 for the older model. The lower cached-input rate is particularly relevant to agent sessions that repeatedly send shared context. GPT-6.2 Sol keeps the same $2/$10 per-million-token prices as GPT-6.1 Sol and GPT-6 Sol; Willison contrasts those rates with GPT-5.6 Sol’s $4/$20 pricing.

The post includes a visual comparison of pelican-on-a-bicycle images generated by Astra and Sol at several effort levels. It also describes a coding test: Willison gave both models the same prompt in Codex CLI, using Ultra effort, to create a Three.js animation in which Yonkers becomes a mech and stomps around Manhattan, Queens and Brooklyn using OpenStreetMap data. He says neither model produced a good enough transformation compared with Claude Fable 5.2 and Gemini 4 Carbon.

Willison also reports an operational concern from Theo Browne’s feedback to OpenAI during pre-release. On a long task, Astra could manage subagents but then block to ask about a low-level implementation detail that a user might not be able to answer. The example involved a request to make a new level for Outer Wilds; roughly two hours into the task, the model asked which of two filenames a reviewer wrote clips to. This is a reported experience, not a measured frequency or a formal release limitation.

What the archive adds

New Runtime’s earlier coverage on August 15, 2026 examined Astra’s capability thresholds as release controls. This is a new model release in a continuing Astra story, with a concrete new pricing fact: cached input is now $0.40 per million tokens.

Trend signal

This release supports agent economics moves to completed work, though only as a pricing signal: cheaper cached context can reduce one component of long-running agent work, but the source does not measure cost per verified completed task. The reported need for user intervention also underscores that token prices alone do not describe the economics of an agent completing a job.

Why now

The new model names and API prices are part of OpenAI’s current release, making this immediately relevant to teams comparing model costs or testing the models in existing API workflows. The announcement’s timing matters most for builders deciding whether Astra’s lower cached-input rate changes the economics of sessions that reuse substantial context.

Why it matters

For teams using agents over long tasks, cached input can be a material part of the bill. Astra’s $0.40 per-million-token cached rate, versus $1 for the previous Astra, gives builders a specific price change to test against their own context reuse and completion costs. Sol’s unchanged $2/$10 rates offer a stable price point for everyday workloads.

But lower token costs are not equivalent to lower costs per completed task. Willison’s coding comparison did not show Astra or Sol outperforming his cited alternatives on that task, and the reported blocking question illustrates how a model can require human attention even while orchestrating subagents. Teams should measure both spend and interruption rate on representative work before shifting production traffic.

Limits and uncertainty

The performance examples are individual tests, not controlled or independently replicated benchmarks, and the post does not quantify task-success rates, latency, total cost per completed task or the frequency of Astra’s reported blocking behavior. The reported comparison concerns one prompt and one kind of coding task; it cannot establish general model rankings. Pricing details here are reported by Willison, while the post does not establish how rates or access vary across all API arrangements. The effect of cached-input savings depends on how much context a particular workflow reuses.

Skill to build now

Run one representative long task on Astra and Sol with the same prompt, tools and context. Record input, cached-input and output tokens, whether the task finishes, and every point where a human must intervene. Use Astra’s $0.40 cached-input rate in the cost calculation, but compare cost per accepted result—not just the token bill. If an agent can block on a question your team cannot answer, add an explicit escalation path before expanding its autonomy.

Reader takeaway

GPT-6.2 Astra’s cached-input price has fallen sharply, making it worth testing for context-heavy agent work; Sol retains its previous price. Treat the launch as a reason to benchmark your own tasks, not as proof of better outcomes: measure completion, cost and human interruptions together.

Evidence / sources

  1. 1simonwillison.net/2026/Nov/17/gpt62-astra-and-sol
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicClaudeExplore the claude topic hub.
  2. 02topicCodexExplore the codex topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract