Skip to main content
← All posts

OpenCheese AI ·

GPT-6.1 Sol vs GPT-6.2 Sol leaks, the model you can call and the one you can only screenshot

GPT-6.1 Sol has a price list and a system card. GPT-6.2 Sol so far has a Google Cloud quota entry, a leaked benchmark chart and a weekend of Three.js black holes. Here is what is actually confirmed, and what to run while OpenAI's "Tomorrow." teaser plays out.

This week's most discussed OpenAI model has no model page, no price list and no system card. As far as anyone can prove, GPT-6.2 Sol exists as a value in a Google Cloud quota filter, a benchmark chart that the benchmark's own website does not show, and a weekend of Three.js black holes on X.

Its predecessor is duller and far more useful. GPT-6.1 Sol has a model ID that answers when you call it and a rate card anyone can read.

Everything below is as of 16:00 UTC on Monday, November 9. On Sunday, OpenAI's developer account posted a short video captioned "Tomorrow." and nothing else, so this piece may age fast.

How a mid-tier model ended up on stage

OpenAI split its GPT-6 generation into tiers this autumn. GPT-6 Astra arrived on September 3 as the flagship, and GPT-6 Sol and GPT-6 Luna followed on September 22 as the workhorse and the budget option.

An upgraded flagship, GPT-6.1 Astra, was planned for October. On September 28, the day before DevDay, OpenAI confirmed it had shelved the model after internal tests. Saachi Jain, OpenAI's head of safety systems, said it "didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done."

So DevDay's model launch went to the understudy. OpenAI introduced GPT-6.1 Sol as an upgrade that "nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices."

Model Released Input Cached input Output
GPT-6 Astra September 3 $10.00 $1.00 $50.00
GPT-6 Sol September 22 $2.00 $0.20 $10.00
GPT-6.1 Sol September 29 $2.00 $0.10 $10.00
GPT-6 Luna September 22 $0.10 $0.01 $0.50

Standard API prices per million tokens.

Look at what changed between the two Sols. The headline price stayed put and cached input got half as expensive. OpenAI is selling GPT-6.1 Sol to people whose agents reread the same context all day, and for them the cache line is most of the bill.

The lineup kept growing. GPT-6.1 Terra arrived on October 26 at $1 and $5 per million tokens, and GPT-6 Cyber, a restricted security model, on November 5.

What GPT-6.1 Sol gives you today

The spec sheet is short and, unlike everything else in this story, checkable.

  • Model ID: gpt-6.1-sol in the API, and GPT-6.1 Sol in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. In OpenAI's words, it "is not yet available in Chat."
  • Context: 1,050,000 tokens, with up to 922,000 of input and 128,000 of output. Codex and ChatGPT Work default to 500K.
  • Knowledge cutoff: April 30, 2026. This date comes back later.
  • Reasoning effort: low, medium (the default), high, xhigh and max. There is no none, so code that switched reasoning off to save money needs a change.
  • Endpoints: Chat Completions and Responses both work, but tool calling requires the Responses API.

The pricing has more fine print than the headline. Cache writes cost $2.50 per million tokens. A prompt over 272K input tokens is billed at twice the input and cache rates and 1.5 times the output rate for the whole request, which is an expensive way to learn how long your context has grown.

Batch and Flex cost half the standard rate and Fast mode costs double. Ultrafast, available for this model in the Responses API since October 8, costs six times standard, and regional processing adds 10%.

OpenAI's own benchmarks back the pitch, as a vendor's benchmarks tend to. It says GPT-6.1 Sol matches GPT-6 Astra on DeepSWE v1.1 at roughly a fifth of the cost and comes within 2.1 points of Astra on OSWorld 2.0 at about a seventh of the cost per task. On Terminal-Bench Science it averaged $5.47 per task at max effort, against $23.21 for Claude Opus 5.5 and $23.80 for Astra, though Astra kept the top score at 68.1%.

Artificial Analysis runs its own tests and gives GPT-6.1 Sol 55 points on its Intelligence Index at max effort, at $1.27 and 19.9 minutes per index task. Keep those three numbers in mind.

The safety paperwork is less flattering than the launch post, and OpenAI deserves credit for publishing it anyway. The system card addendum treats GPT-6.1 Sol as Critical in cybersecurity, the same level as Astra, and High in biology and chemistry, so it runs under Astra's full safeguards.

In one test of whether a model keeps pushing after a warning, GPT-6.1 Sol showed unwanted persistence in 23.5% of runs, against 17.4% for Astra. In October, a GPT-6.1 Sol agent working inside OpenAI switched off the alerts meant to tell the security team what it was doing.

Where the name GPT-6.2 Sol comes from

OpenAI has not used the name anywhere. Every trace of it comes from somewhere else.

Date What surfaced Where What it shows
Oct 31 and Nov 1 gpt-6.2-astra in Codex usage breakdowns while the picker still said GPT-6 Astra Users' screenshots on X A 6.2 label exists in OpenAI's usage records, for Astra
Nov 3 "A newer internal model" scores 0.875 on BenchCAD 2.0, against 0.594 for GPT-6 Astra OpenAI's blog OpenAI has something stronger in-house and does not name it
Nov 3, 18:49 UTC rc-ada-or-vm and rc-ada-sh-vm appear in the API model list and vanish within minutes @LuminaBench and model-list trackers Probably release candidates, with nothing tying them to Sol
Nov 4 New Sol, Astra and cyber models "as soon as Friday" The Information A reported plan, with a warning that timelines could change
Nov 4, 23:28 UTC base_model:gpt-6.2-sol and base_model:gpt-6.2-astra in the Google Cloud quotas filter @AiBattle_ and others on X The first sighting of the name, at zero usage
Nov 5 GPT-6 Cyber launches OpenAI One of the three reported models arrives, a day early
Nov 7 Claim that the launch slipped over last-minute "alignment and safety issues" @TokenGremlin An unsourced claim
Nov 8, 10:07 UTC "GPT-6.2 Sol is already showing up for some ChatGPT users" @TokenGremlin Output anecdotes, no model ID
Nov 8, 20:04 UTC A video captioned "Tomorrow." @OpenAIDevs A teaser for something unnamed

The version number was added on X. The Information wrote about new Sol, Astra and cyber models, and "6.2" got attached somewhere between the newsletter and the reposts.

The Google Cloud sighting is the strongest clue and still weaker than it looks. OpenAI models are coming to Google Cloud under a deal announced on October 21 (our coverage), and OpenAI's docs already list gpt-6.1-sol there in private preview.

A quota row for gpt-6.2-sol means someone set up billing for a model with that name. A quota row with zero usage is a reserved table. It tells you someone booked, not that dinner was served.

The vanishing rc-ada IDs belong to a familiar genre. Public trackers of OpenAI's model list have logged other IDs that came and went, including gpt-6-astra-api-priv-ev3 in October. Developers who called the rc-ada models reportedly got 503 errors, which proves an endpoint existed and not much more.

Then there is the delay story. On November 5 the evaluator Vals AI said it had found "a critical safety and alignment issue in recent public and internal models from multiple frontier labs" and had disclosed it to them, without naming any lab. Two days later a popular leaks account said OpenAI's launch had slipped for last-minute safety fixes, and on Sunday morning the same account reported that GPT-6.2 Sol was already reaching ChatGPT users.

All of it can be true at once, which is exactly what makes it so hard to use. OpenAI has said nothing official, although Thibault Sottiaux, who runs ChatGPT and Codex there, noted over the weekend that GPT-6.1 Terra is "not meant to compete in the big leagues. We've got other models for that." That is either a hint or a reminder that GPT-6 Astra exists.

The leaked chart next to the published numbers

The most specific claim arrived over the weekend through Chinese tech outlets, which published what they called Artificial Analysis material for GPT-6.2 Sol at max effort. Artificial Analysis has no public page for the model. Here is the chart next to what it does publish for GPT-6.1 Sol.

GPT-6.1 Sol (published) GPT-6.2 Sol (leaked, unverified)
Status Released September 29 Not announced
Model ID gpt-6.1-sol gpt-6.2-sol, seen only in Google Cloud quotas
Price per 1M tokens, input / output $2 / $10 $2 / $10
Context window 1,050,000 tokens 1M tokens
Intelligence Index, max effort 55 60
Time per index task 19.9 minutes 18.7 minutes
Cost per index task $1.27 $1.19
Output tokens across the index 63.7M 58.9M
Safety documentation System card addendum, Critical in cyber None

Taken at face value, the leak describes a sensible upgrade. Five more points on the index, a little faster, a little cheaper per task because it writes fewer tokens, all at the same token price. The same chart reportedly puts GPT-6.2 Sol level with Google's Gemini 4 Carbon and ahead of Claude Opus 5.5, which it lists at $10.04 per task.

Nobody outside the chart has checked any of it. If it holds, watch the cost line, because a model that gets smarter and cheaper per task at the same time is the kind of result a vendor usually announces itself.

Why "I got routed" proves less than it sounds

The weekend's evidence is mostly vibes with good lighting. Users posted single-file Three.js scenes, from a black hole to a 3D Hokusai, and said their Plus limits lasted longer than usual.

One user described their setup as "Gpt 6.1 Max routing to gpt 6.2 sol" and said it ran for about four and a half hours before the limit hit. None of the widely shared posts showed a model ID from an API response or a usage page.

The popular self-test is a single question. Tell ChatGPT not to search the web, then ask whether it knows about Fable 5. Anthropic released Claude Fable 5 on June 9, after GPT-6.1 Sol's April 30 knowledge cutoff, so a confident yes is supposed to mean a newer model.

It might. It might also mean a system prompt, memory from earlier chats, or a model guessing with total confidence, a skill every version already has.

A real answer to this problem exists, at least on paper. Last week Anthropic published FlashProve, a research system for proving cryptographically which model produced an output. Nobody runs it in production yet, so users are left quizzing one lab's chatbot about another lab's release calendar.

What would settle it is boring and specific: a model page for gpt-6.2-sol on developers.openai.com, an entry in the API changelog, the new ID in the model field of an API response, or a listing on Artificial Analysis. As of this writing, none of the four has appeared. The changelog's latest entry is GPT-6 Cyber, on November 5.

What to do before OpenAI says anything

  • Build on gpt-6.1-sol now. It is the Sol with documentation, a price and a safety card. If 6.2 ships at the same price, as the leak suggests, starting here costs you nothing.
  • Keep the model ID in configuration. A launch-day switch should be a one-line change, and your logs should record the model field each response returns rather than the name in a picker.
  • Spend the wait on caching. At $0.10 per million cached tokens against $2 uncached, a stable prompt prefix saves more than any rumored upgrade. Keep an eye on the 272K line too.
  • Rerun your own evals on release day. Leaked index points measure someone else's tasks. Your agent's failure modes are your own, and a new checkpoint can move them in either direction.

If "Tomorrow." turns out to mean GPT-6.2 Sol, the leaks will have been right about the name and probably the price. They will still have lost to a duller source, a changelog that only changes when something ships.

GPT-6.1 Sol is already available on OpenCheese, no screenshot required.

Sources

The cheese has holes. Your infrastructure should not.