Key takeaways

  • OpenAI’s official RSS dates the GPT-6.2 Astra and Sol announcement to November 17, 2026 at 17:45 UTC and names ChatGPT Work, Codex and the API.
  • The direct API IDs are gpt-6.2-astra and gpt-6.2-sol. Both references list a 1,050,000-token context, 922,000 maximum input tokens and 128,000 maximum output tokens.
  • Standard short-context prices are $8 input/$40 output for Astra and $2 input/$10 output for Sol, per million tokens. Requests above 272K input tokens incur higher rates for the full request.
  • Both references document Responses API tools and a Multi-agent beta; neither accepts reasoning effort none. AccessAllGPT has not called or benchmarked these paid models.
01

What OpenAI announced on November 17

OpenAI announced GPT-6.2 Astra and Sol on November 17, 2026 at 17:45 UTC, according to its official news feed. The feed describes better professional outputs, coding, computer use and efficiency at lower cost across ChatGPT Work, Codex and the API. Those improvement claims are OpenAI’s, not results from AccessAllGPT testing.

The accessible API references identify gpt-6.2-astra as the model for the hardest projects and gpt-6.2-sol as the everyday-work option for coding and agents at scale. OpenRouter also lists openai/gpt-6.2-astra and openai/gpt-6.2-sol, alongside Pro and batch variants. A listing establishes a published route, not proof that every account can successfully call it.

The full OpenAI announcement page was blocked by a security challenge during this review. We therefore use the directly retrieved official feed for the announcement date and product-channel summary, and the developer references for specifications. We cannot establish ChatGPT Work or Codex plan eligibility, rollout completion or detailed launch benchmark results from the feed alone.

02

The API specifications builders can use

Both model references list text and image inputs and text output, a 1,050,000-token context window, a 922,000-token maximum input and a 128,000-token maximum output. The total context number is not a promise that a request can contain 1,050,000 input tokens. Audio and video are marked unsupported as model modalities.

OpenAI documents streaming, function calling and structured outputs, but not fine-tuning. Its Responses API tool list includes web search, file search, code interpreter, hosted shell, apply patch and computer use. Image generation appears as a supported tool; that does not change the base model’s text-output modality.

Both pages describe delegating work to subagents with the Multi-agent beta and discounted background completion windows in public beta. They explicitly say requests setting reasoning effort to none are rejected. Existing integrations that assume a no-reasoning mode need a configuration check before migration.

Rate limits remain usage-tier dependent. These are documented capabilities and limits, not an account-level access guarantee, tool reliability measurement or production service-level commitment.

03

Pricing depends on input length and processing tier

For Standard short-context requests, OpenAI lists Astra at $8 per million input tokens, $0.40 per million cached-input tokens and $40 per million output tokens. Sol is $2 input, $0.10 cached input and $10 output on the same basis. Thirty-minute cache writes cost $10 for Astra and $2.50 for Sol per million tokens; twelve-hour writes cost $20 and $5 respectively at Standard short-context pricing.

For requests with more than 272K input tokens, the references say input and cache rates double and output rates rise to 1.5 times the short-context rate for the full request. That makes Standard long-context input/output $16/$60 for Astra and $4/$15 for Sol per million tokens. Do not apply the surcharge only to tokens beyond the threshold.

Batch costs 50% of Standard rates and Fast costs twice Standard for the same context length. The pages list Ultrafast at three times Standard through February 28, 2027, rising to six times from March 1, 2027, Pacific time. Processing mode belongs in a cost estimate alongside model ID, input length, cache behavior and generated tokens; the headline token rates alone are not a complete agent-run bill.

04

What the launch changes—and what remains unverified

Our analysis: this launch creates a practical two-model shortlist, with Sol as the lower-price starting point and Astra as the candidate for tasks where additional capability may justify its higher rates. Equal context limits and broadly similar documented tools make a controlled comparison easier, but do not establish equal quality, latency or reliability.

OpenAI claims improvements across professional work, coding and computer use. We did not retrieve the full launch evaluation, reproduce a benchmark, run a coding task or measure tool success. The registry’s benchmark fields are not used here as independent evidence. These paid models do not qualify for this job’s free-model hands-on workflow.

The next step is a small, non-sensitive evaluation against your current model: keep the task set and tool permissions fixed, record correctness and successful completion, and compare total billed usage and latency. Start with the exact API ID and a supported reasoning configuration. Keep the existing route available until the replacement meets your own acceptance criteria.

05

A first migration check

Suggested evaluation setup, not a test we performed.

Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].

Begin with gpt-6.2-sol; compare Astra on the same tasks rather than assuming the flagship is always necessary.

Remove reasoning effort none; verify your required tools and actual account rate limits.

Record processing tier, cache writes and reads, input length, output usage and tool charges; flag requests above 272K input tokens.

Compare successful task completion, correctness, latency and total cost with the current route before switching traffic.

Primary sources

  1. OpenAI news RSS — Introducing GPT-6.2 Astra and SolOpenAI · Reviewed: Matching item title, canonical link, November 17 17:45 UTC pubDate and description · Retrieved · Supports: Announcement date, model-family names, named product channels and vendor improvement claims. Not the full announcement or a plan-eligibility matrix.
  2. GPT-6.2 Astra API model referenceOpenAI · Reviewed: Model ID and description; reasoning restriction; token limits; text pricing and context/tier qualifications; modalities; features; Responses tools; rate limits · Retrieved · Supports: Direct-API Astra identity, documented capabilities, limitations, beta features and pricing. No independently measured performance.
  3. GPT-6.2 Sol API model referenceOpenAI · Reviewed: Model ID and description; reasoning restriction; token limits; text pricing and context/tier qualifications; modalities; features; Responses tools; rate limits · Retrieved · Supports: Direct-API Sol identity, documented capabilities, limitations, beta features and pricing. Does not prove individual account access.
  4. OpenRouter public model registryOpenRouter · Reviewed: GPT-6.2 Astra, Sol, Pro and batch listing IDs; context fields and non-zero token rates · Retrieved · Supports: Registry availability of the named OpenRouter routes; not a completed inference call, vendor benchmark validation or proof that Pro is a distinct base model.

Limitations

No model calls, authenticated ChatGPT Work or Codex trial, hands-on tests, independent benchmarks, latency measurements or billing verification. The full canonical announcement returned a 403/security challenge. Announcement claims are limited to the official RSS description; specifications and prices come from accessible developer references. OpenRouter was inspected as a registry only. Account access, complete rollout scope, beta eligibility and real workload reliability remain unverified.

Disclosures

AccessAllGPT received no payment, private access or vendor review for this briefing. OpenAI and OpenRouter did not review or endorse it. AccessAllGPT Research is operated by NeuralArc; publication-wide relationships are listed on the disclosures page.

Further AccessAllGPT guidance

  1. Model selection without leaderboard chasing
  2. What AI models cost
  3. Models and APIs
  4. How this publication is made