GPT-6 Sol: What Is Verified and What Is Leaked

GPT-6 Sol is not a model you can call today. As of September 16, 2026, OpenAI's model catalog, pricing table, changelog, and documentation index list gpt-6-astra and gpt-5.6-sol. None of them lists gpt-6-sol or any other gpt-6-* tier.

The name circulates because two real names collide. GPT-6 is a released model, and Sol is a real tier of the previous family. Put them together and "GPT-6 Sol" sounds like the obvious next release, which is exactly why it spreads faster than the evidence supports.

This page does three things: it records what is verified, it shows what the leak actually contains, and it tells you which model ID to send today if what you wanted was a cheaper way to get GPT-6-class work done.

Executive answer

Why the name sounds real

The confusion is structural, not careless. OpenAI's two most recent naming layers overlap:

• GPT-5.6 is a three-tier family: gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna, plus a gpt-5.6-cyber route.
• GPT-6 shipped on September 3, 2026 as a single model, gpt-6-astra.

So the pattern "GPT-6 plus a Sol/Terra/Luna tier" is a reasonable extrapolation. It is also exactly the kind of pattern that produces confident-looking search results with no product behind them. A version number plus a tier name is a guess, not a model ID.

One more distinction matters when reading leaks. ChatGPT plan availability and API model IDs are different surfaces. A name appearing in a plan picker, an evaluation arena, or a monitoring tool is not evidence that the ID is callable through the API. Treat any screenshot as a lead, never as a specification.

Is there a GPT-6 Sol release date?

No release date exists, because no release has been announced. Anyone publishing a specific date is guessing, and the dates that circulate have not held: there is no model page, pricing row, or changelog entry to attach them to.

The only dated GPT-6 event in OpenAI's record is September 3, 2026, when gpt-6-astra shipped. If you are tracking the name, watch five surfaces instead of a rumour calendar: the model catalog, the pricing table, the changelog, a "using this model" migration guide, and the model list your own provider returns.

Prediction markets and aggregator blogs will keep producing dates. A date is only actionable once one of those five surfaces agrees with it.

What is verified: GPT-6 Astra

Astra is the only GPT-6 model in the API today. It was announced in the OpenAI changelog on September 3, 2026 as the most capable model for the hardest end-to-end work.

Two Astra details are widely missed and both are operationally relevant.

The first is endpoint support. Astra appears on both Chat Completions and Responses, but OpenAI's migration guidance states that tool calling requires the Responses API. An agent that works on a plain chat call can still fail the moment you attach tools.

The second is safety posture. The GPT-6 Astra system card states that Astra is the first OpenAI model to reach the "Critical" level of cybersecurity capability under the Preparedness Framework, and that it is significantly more robust to jailbreaks than GPT-5.6 Sol. In the same card's realistic-work-environment evaluation, Astra's overall misaligned-outcome rate is 3.4% against 18.8% for GPT-5.6 Sol. A stronger model is not automatically a safer deployment; here the capability jump and the robustness gain arrive together, but the capability threshold is the part that changes how much review your tool permissions deserve.

The full GPT-6 Astra review covers migration order, the pricing rule in detail, prompt caching math, and the Codex availability matrix.

What is verified: GPT-5.6 Sol

If you searched for "GPT-6 Sol" because you wanted the strongest GPT model and you already know the $5.6 tier name, this is the model you were looking for.

The differences that decide real work are narrower than the version numbers suggest:

• Effort range. Sol accepts none and defaults to medium. Astra has no none and starts at low. Clients that send none or minimal to Astra will error until you change them.
• Sampling. Sol still accepts the older sampling parameters in its documented surface; Astra does not support custom temperature, top_p, or logprobs.
• Knowledge cutoff. Astra's cutoff is about ten weeks later, which changes answers about anything that moved this spring.
• Price. Astra's direct input rate is 2.5x Sol's, and its output rate is 2.5x as well.
• Safety evaluation. Astra scores far lower on the misaligned-outcome evaluation described above.

The leak, itemised

Everything below is unverified. It is listed so you can see how thin the evidence is, and so you can tell the difference between a name sighting and a shipped product.

Note what the leak is not. There is no OpenAI model page, no pricing row, no changelog entry, no documentation entry, no system card, and no migration guide. The public discussion also overstated its own reach: the forum thread that most coverage points back to had a few hundred views and four replies when we checked, and the screenshot posts all trace to the same image, which is a single sighting, not independent corroboration.

What would count as evidence

Before you re-plan a migration around a rumoured model, wait for at least two of these:

1. A model page under OpenAI's documentation with an exact ID and a specification table.
2. A row on the official pricing table with input, cached, and output rates.
3. A dated entry in the API changelog.
4. The ID appearing in a live model list from a provider you can actually call.
5. A "using this model" migration guide describing parameter differences from the model you are on.

A screenshot is none of these. A name in a leaderboard is none of these either.

Check any model ID yourself in about a minute

Do not take our word for it, and do not take a screenshot's word for it. Ask the API.

# Any OpenAI-compatible gateway: list model IDs and search for the exact name
set -o pipefail
curl --fail-with-body --silent --show-error https://api.omniakey.com/v1/models \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
| jq -r '.data[]?.id | select(startswith("gpt-6"))'

Run against OmniaKey on 2026-09-16, that returns exactly one model:

gpt-6-astra

The same list contains gpt-5.6-sol. There is no gpt-6-sol, no gpt-6-terra, and no gpt-6-luna. Run the equivalent against OpenAI's own endpoint if you have a key there, and check both: a gateway list tells you what your provider routes, while the vendor's list and documentation tell you what the model actually is.

Two rules keep this check honest:

• Match the whole ID. gpt-6 in a response object may be a family label or an alias, not a callable model. Search for the full ID you intend to send.
• A 200 response with the ID in the model list is evidence. A screenshot is not. If an ID is not in a list you can request, do not put it in a config file; you will get a 404 or an unsupported-model error at the worst possible moment.

If you wanted GPT-6 Sol because Astra is expensive

The recurring concern in the leak discussion was not capability. It was quota: people want Astra-class work without watching a usage limit, and they want a mid-tier option with a mid-tier price. That is a real problem, and it does not require an unannounced model to solve.

The most useful thing OpenAI has published on this is effort guidance rather than a new tier. As quoted in the community thread, OpenAI's guidance is that Astra at low reasoning performs better than GPT-5.6 Sol at high, and the recommendation for existing Sol high users is to move to Astra low or medium. The lever is reasoning.effort, not a smaller model name.

That gives you a routing table you can use today:

Budget with the real numbers instead of a hoped-for discount. On OmniaKey, gpt-6-astra is listed at $0.90 per 1M input and $4.50 per 1M output, and gpt-5.6-sol at $0.45 per 1M input and $2.70 per 1M output. Astra is roughly twice the input rate; a rumoured cheaper tier that has no price is not a saving you can plan against. Check the live catalog for current rates, and note that OmniaKey lists a single rate across the stated context window rather than adding a second tier above 272K input tokens.

When GPT-6 Sol (or Astra's siblings) actually ship

If OpenAI publishes a gpt-6-sol model page with a specification table and a pricing row, this article will be updated with the verified ID, rates, limits, and the gateway route, and the date at the top will change. Until then:

• We will not publish a "GPT-6 Sol review". A review of a model that has no model page is a review of a rumour.
• We will not add the ID to the catalog, and you should not add it to a config file.
• A leaked tier name is not a release date, and a release date is not a price.

Limits and common mistakes

"It appeared in a model list" is not generally available

Screenshots and scraped lists can come from canary deployments, internal tooling, third-party monitors, or plan pickers. None of them imply that you can call the ID, and none of them imply a stable price.

A rumoured tier does not inherit the current tier's economics

Assume nothing about a future tier's context rule, effort levels, endpoint support, or long-context multiplier. Astra itself shows how much these vary inside one generation: no none effort, no custom sampling, tool calling only on Responses, and cache writes billed at 1.25x input.

Do not pin an unannounced ID

model: "gpt-6-sol" today is a failed request, not a forward-looking configuration. Keep the ID in an environment variable so a change is a deploy-time edit rather than a code change, and validate it against the provider's model list in CI.

Do not plan a budget around a discount that does not exist

Every rate in this article is either on OpenAI's pricing page or on an OmniaKey catalog page on 2026-09-16. A rumour has no rate, and extrapolating "Sol was 20% cheaper than the previous flagship, so the next one will be too" is not a budget.

Capability is not permission

Astra's Preparedness Framework classification and its tool support are different facts. A model that can use a shell, a browser, and MCP connections will do exactly what your permissions allow, and the interesting failures in agent work are permission failures. Review what the agent may touch before you review which model runs it.

Frequently asked questions

Is GPT-6 Sol real?

Not as a callable model. The name has appeared in an unverified screenshot and in community discussion, but it is absent from OpenAI's model catalog, pricing page, changelog, and documentation as of September 16, 2026. Two real models are behind the confusion: gpt-6-astra and gpt-5.6-sol.

What is GPT-6 Sol?

Nothing that OpenAI has documented. "GPT-6" is the family that contains gpt-6-astra, and "Sol" is a tier name from the older GPT-5.6 family. The two names describe different generations, so "GPT-6 Sol" is a combination no model page defines.

What is GPT-6 Sol's release date?

There is none to report, because no release has been announced. Any date you see is speculation. OpenAI's changelog dates the only GPT-6 release so far at September 3, 2026, and that model is Astra.

Is there a GPT-6 Sol Ultra?

Not as a model ID or a reasoning-effort value. OpenAI's reasoning guide lists none, minimal, low, medium, high, xhigh, and max, and no model in the catalog carries an "Ultra" suffix.

The name almost certainly comes from Ultrafast mode, which is real: a documented API service tier for GPT-5.6 Sol announced on August 13, 2026 that runs up to 14x faster than Standard processing, initially in limited preview. If you were looking for a faster Sol, that is the feature to ask your provider about — not a new tier.

Will there be GPT-6 Terra and Luna versions?

Not announced. The expectation is understandable: the GPT-5.6 family had three tiers, and earlier generations added smaller variants shortly after a flagship launch. OpenAI has published no GPT-6 tier beyond Astra, so the honest answer is that nobody outside OpenAI knows.

Is GPT-6 Sol the same thing as GPT-5.6 Sol?

No. gpt-5.6-sol is a shipped model with a specification table and published rates. "GPT-6 Sol" is a name from an unconfirmed leak. They are different families, and only one of them is in the API.

How does GPT-6 Sol compare with Claude Fable 5?

There is nothing to compare, because one of the two does not exist. If what you actually want is the Fable-versus-Sol decision, that comparison is published with the same evidence standard: Claude Fable 5 vs GPT-5.6 Sol for coding.

Should I wait for GPT-6 Sol before migrating off GPT-5.6 Sol?

No. If you are moving today, migrate to gpt-6-astra and control cost with reasoning.effort and prompt caching. Waiting for an unannounced tier means staying on the model you already decided was not the best fit.

Which model should I use for a GPT-6-class task right now?

gpt-6-astra is the only GPT-6 model available. Start at low or medium effort, use the Responses API for tool workflows, and only raise effort when a test shows a gap.

Does OmniaKey route GPT-6 Sol?

No, because there is no such model to route. OmniaKey currently lists gpt-6-astra and gpt-5.6-sol as separate live routes with separate rates.

How can I verify a model ID without trusting an article?

Request your provider's model list and search for the exact full ID, then compare it against the vendor's own model page and pricing table. The commands are in the section above.

Official sources and leak sources

Official, verified on 2026-09-16:

• OpenAI: model catalog
• OpenAI: GPT-6 Astra model page
• OpenAI: GPT-5.6 Sol model page
• OpenAI: API pricing
• OpenAI: API changelog (source of the Ultrafast mode announcement)
• OpenAI: reasoning guide (the full list of reasoning-effort values)
• OpenAI Deployment Safety Hub: GPT-6 Astra system card
• OmniaKey: GPT-6 Astra route and OmniaKey: GPT-5.6 Sol route

Unverified discussion of the leaked name, listed so the leak can be traced rather than repeated:

• OpenAI Developer Community: "GPT-6 Sol, Terra, and Luna?"
• OpenAI Developer Community: Astra low/medium as the replacement for Sol high