GPT 6.2 Sol Ultrafast Delivers 150 Tokens/s but Burns Through Limits 6x Faster
OpenAI has released GPT 6.2, but after an hour of testing, the author of Tips AI saw none of the promised improvements. The model still misreads intent and runs unnecessary checks. What's new is Ultrafast mode in the $200 Pro subscription: about 150 tokens per second, with limits draining 6x faster than in standard mode (previously 8x). That's handy for an urgent fix, but Claude remains ahead as a primary agent.
AI-processed from TG: @tips_ai (Tips AI / IT & AI); edited by Hamidun News
OpenAI has released GPT 6.2, and developers on X are already arguing over whether it is actually better. After his first hour with the model, Tips AI's author says that, globally, nothing has changed.
The exception is the new Ultrafast mode in the $200 Pro subscription (the post calls it Pro 200). It is a sped-up version of GPT 6.2 Sol: it generates about 150 tokens per second (a token is a chunk of a word, and response speed is measured in them) and uses 6x more of your subscription limits than standard mode.
An example makes it simple: a service goes down in production, you need a quick fix, and you switch to Ultrafast. The answer arrives noticeably sooner, but if a regular request cost, say, one unit of your limit, the fast one costs six. If you pay for OpenAI or Claude subscriptions and build chains of several models, these numbers can help you decide where you need GPT 6.
2 and where you don't.
What hasn't changed in GPT 6.2
The author adds a caveat: the release was an hour ago, so these are first impressions, not benchmark results. Still, some authors on X had early access, and their criticism matches his own experience. The main complaint: for yet another generation in a row, OpenAI has not fixed the known problems of its models.
In GPT 6.2 Sol / Astra he barely noticed anything of what OpenAI's blog described. The model hasn't become more pleasant to talk to, shows very little initiative and doesn't grasp intent: you ask for one thing, and it reads the task narrowly and literally. Its old habits remain too: the model gets stuck in loops and runs a lot of unnecessary tests and checks.
Ultrafast: the only truly new thing
The author calls the arrival of Ultrafast in Pro 200 great, because there is no equivalent. Claude's subscription has no fast Opus, and other players' Fast models are available only via API, where you pay for every token, and that is expensive.
Numbers from the post:
- in the OpenAI subscription, Ultrafast consumes 6x more limits than standard mode; the multiplier used to be 8;
- the author measured the speed himself: 6.2 Sol Ultrafast runs at about 150 tokens per second.
For generation speed, 150 tokens per second is not a record. But the model spends tokens very economically: it needs less text to reach a result. That makes it feel fast in use, not just on the counter. When something needs fixing urgently, the author calls this mode indispensable. The downside is obvious: not everyone has $200 for a subscription, especially just to keep Ultrafast on hand in case.
Where GPT is useful: as a subagent, not the main agent
The most practical takeaway concerns how to structure a workflow. The author uses OpenAI models as subagents for Claude on Explore (scouting a codebase), Backend and Review (code review) tasks. They worked for him and his Opus there before, too. After the update, these subagents will get a bit smarter and will spend fewer limits and less time, so he is not upset with the release overall.
This is an example of an approach where you don't pick one model for everything but assign roles. You keep a proactive orchestrator on top and hand narrow, well-defined tasks to a model that handles them cheaply and carefully. Where you need diligence rather than initiative, GPT's weak proactivity doesn't get in the way.
The author calls using GPT directly in Codex unbearable: the models feel "inhuman" and are unpleasant to work with. As a primary agent, he sees no competitors to Claude, and losing all his Claude accounts, by his own account, would drive him mad. He hopes OpenAI will someday ship a proactive model that is pleasant to work with and strong at frontend, but he isn't sure that will happen.
What to do in practice
This is the subjective view of one experienced user in the first hour after release. There are no independent quality measurements, and the author measured the 150 tokens per second himself. Whether a model is "pleasant to work with" depends on your tasks and habits: some people may find GPT's literalness and meticulous checks a good fit.
If you build agent chains, test GPT 6.2 in the scout and reviewer roles rather than making it the main agent. If you have the $200 Pro plan, try Ultrafast on urgent edits and work out whether burning 6x more limits pays off. As for a revolution in conversation and proactivity, judging by the first reactions, don't expect one from this release.
- Nov 18, 2026StarSling Cuts Mastra's Slowest Build from 29 Minutes to 5
- Nov 18, 2026AI Agents Shipped 228 Game Releases Without Human Review; Tests Sent Back 25% of Tasks
- Nov 18, 2026JetBrains' Mellum Runs Locally in an IDE: 5 Out of 10 on a MacBook M3 Pro
- Nov 18, 2026OpenAI Releases Legal AI: 95% Fewer Fabricated Facts Than Claude Opus 5.5
Key Questions Answered
What hasn't changed in GPT 6.2?
The author adds a caveat: the release was an hour ago, so these are first impressions, not benchmark results. Still, some authors on X had early access, and their criticism matches his own experience. The main complaint: for yet another generation in a row, OpenAI has not fixed the known problems of its models. In GPT 6.2 Sol / Astra he barely noticed anything of what OpenAI's blog described.
What are the key details on Ultrafast: the only truly new thing?
The author calls the arrival of Ultrafast in Pro 200 great, because there is no equivalent. Claude's subscription has no fast Opus, and other players' Fast models are available only via API, where you pay for every token, and that is expensive. Numbers from the post: in the OpenAI subscription, Ultrafast consumes 6x more limits than standard mode; the multiplier used to be 8; the author measured the speed himself: 6.2 Sol Ultrafast runs at about 150 tokens per second.
Where GPT is useful: as a subagent, not the main agent?
The most practical takeaway concerns how to structure a workflow. The author uses OpenAI models as subagents for Claude on Explore (scouting a codebase), Backend and Review (code review) tasks. They worked for him and his Opus there before, too. After the update, these subagents will get a bit smarter and will spend fewer limits and less time, so he is not upset with the release overall. This is an example of an approach where you don't pick one model for everything but assign roles.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.