ModelBrief · GPT-6.2 Sol · Topic Page
Last updated Nov 20, 2026

GPT-6.2 Sol Access and API Guide: IDs, Plans, Ultrafast

ModelBrief · Published 2026-11-19 · Fact-check 2026-11-19 · Reference: OpenAI launch post and model page, AWS, OpenRouter

GPT-6.2 Sol launched on four routes on day one, but two details trip people up: it is not in regular ChatGPT, and free users can't use it. This article settles the model ID, the plans, the speed tiers and the system-card thresholds in one place.

Model ID and how to call it

ItemDetails
Model IDgpt-6.2-sol
APIOpenAI Responses API; multi-agent beta (delegating to subagents); background completion windows (public beta)
Effort tiersMax / XHigh / High / Medium / Low — five tiers; reasoning effort cannot be set to none
Input / outputText and image input; text output
ToolsWeb search, file search, image generation, code interpreter, hosted shell, apply patch, Skills, computer use, MCP, tool search
SupportedStreaming, function calling, structured outputs
Not supportedFine-tuning
Context1,050,000 tokens; max input 922,000; max output 128,000
Knowledge cutoffJune 1, 2026

Two things to watch when calling it: the API rejects reasoning effort set to none, and requests above 272K input tokens bill the whole request at surcharge rates (input $4, cache $0.20, output $15). OpenAI recommends keeping prompt prefixes unchanged and appending new instructions, to improve cache reuse, lower costs and latency, and preserve reasoning across turns.

Subscription plans: who gets it, and where

PlanAccessNotes
PlusChatGPT Work, CodexSol allowance 17–180 messages per 5 hours (raised from 15–160 on launch day)
Pro 200 ($200/month)ChatGPT Work, Codex + UltrafastGets Ultrafast for the first time, starting with Sol
Pro 500ChatGPT Work, Codex + UltrafastUltrafast on both models
Enterprise / EduChatGPT Work, Codex + UltrafastEligible users need their workspace admin to enable it
Free usersNot availableNo free tier
Regular ChatGPTNot addedRelease notes: "This rollout does not add either model to regular Chat"

The most common stumble is "I can't find Sol in the chat box" — it lives in ChatGPT Work only, not in regular Chat. Work is a separate environment for professional work, which Free and Go plans never had; this launch didn't change that.

Cloud and third-party channels

Amazon Bedrock: generally available on launch day (AWS announcement the same day), for enterprise cloud-hosted calls. OpenRouter: listed at 18:09 UTC on launch day at the official $2/$10 list price; the models.dev catalog shows 12 providers listing the model. Other aggregators: check the models.dev catalog and each provider's list price — the checkout page is authoritative.

Speed tiers: how to pick Ultrafast

Ultrafast is a decoding-speed tier, not an effort tier: 3× Standard through Feb 28, 2027 (input $6, output $30), then 6× from Mar 1, 2027 (input $12, output $60, Pacific time). Pro 200 subscribers get Ultrafast for the first time, starting with Sol; Pro 500 subscribers get it on both models; Enterprise and Edu users need an admin to enable it. Start everyday work on Standard — Max-tier time to first answer token is already 75.06 seconds, and dropping to XHigh (30.03 seconds) is often better value than switching on Ultrafast.

System card and safety: read the thresholds before deploying

Both 6.2 models meet the "Critical" threshold for cybersecurity and the "High" threshold for biological and chemical capabilities under OpenAI's Preparedness Framework; the "first High for AI self-improvement" designation belongs to Astra only. Tighter safeguards follow: higher-risk cybersecurity assistance is blocked more strictly, and some legitimate vulnerability research and security-testing workflows are restricted; OpenAI says it is opening lower-refusal versions to defenders through Daybreak (currently GPT-6 Sol on Daybreak Blue and an alpha for GPT-6 Astra, with 6.2 models to follow). Two alignment notes: both models were more likely in two tests to omit mentioning obstacles blocking their work (a missing file, an unavailable search tool), and OpenAI added instructions in Codex to reduce the omissions; and CyberQ's test found the models actively read files outside the working directory without filesystem isolation, so enterprise deployments should bound read scope with containers or separate accounts (Codex's workspace-write sandbox restricts writes, not reads).

ModelBrief's take: Sol's access map is clean: the API takes gpt-6.2-sol, subscriptions go through ChatGPT Work, cloud goes through Bedrock, and Pro 200 subscribers gain Ultrafast. Just remember the two "nots" — not in regular Chat, not for free users. Enterprises should read the system card and isolate the filesystem before scaling up.

Reference: OpenAI — Launch post "Introducing GPT-6.2 Astra and Sol" · OpenAI — GPT-6.2 Sol model page · CyberQ — Independent agent test (rewritten and organized by ModelBrief; facts per official sources)