A generated title card for an article on GPT-6.2 compute allocation, with three cards headed The claim, The record and The check, and a footer stating that all figures are OpenAI-published and that no GPT-6.2 existed as of 16 November 2026.
Guides & Insights

GPT-6.2 and the Compute Excuse: What OpenAI Has Published About Where the GPUs Actually Go

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

On the night of 15 November 2026, @kimmonismus offered a theory for why GPT-6.2 has not appeared: OpenAI is putting scarce compute into research rather than into serving new models, because rapid progress makes intermediate checkpoints less compelling to release, and safety, product readiness and cost may be factors too. That claim has a checkable half and an uncheckable half. The checkable half is that GPT-6.2 does not exist in any artefact OpenAI publishes. The model catalogue lists six GPT-6 text identifiers — GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, GPT-6 Sol and the access-controlled GPT-6 Cyber — and nothing numbered 6.2. The rate card has no 6.2 row. The changelog's most recent model entry is GPT-6 Cyber on 5 November, eleven days ago. Artificial Analysis, whose index is independent of OpenAI even though it measures its own task set, has no GPT-6.2 entry. The uncheckable half is where the compute actually goes, and nobody outside the company can settle that from a tweet.

What can be done instead is narrower and more useful: read the four documents OpenAI has published about compute and model governance since late September, and ask what they say about allocation. Read together they describe something the tweet does not — compute being rationed inside research, by a documented control system with a published stop latency, while serving is described as a separate stack that is, in OpenAI's own phrasing, "already much more centralized by nature." That is the shape of the compute story as far as the public record supports it. Everything below is labelled by where it comes from, because most of it is a vendor describing itself.

A generated six-row scoreboard of GPT-6 compute facts, reading: catalogue IDs six, none numbered 6.2; latest model GPT-6 Cyber on 5 Nov 2026; Control Room stop bound 35 seconds; RL pause in September 11 days; Stargate added over 3GW in 90 days; Georgia Power curtailment up to 1,000 MW. Footer: all figures OpenAI-published, unreproduced; no GPT-6.2 scores exist.

What the catalogue says on 16 November

The absence is easy to verify and worth stating precisely, because "GPT-6.2 is coming" posts have been circulating for weeks. OpenAI's API model catalogue, fetched this morning, contains six GPT-6 identifiers and one image family, one voice model and one life-sciences model. The text line reads: GPT-6 Astra as "our most capable model for the most demanding work"; GPT-6.1 Sol as "near-Astra performance for complex work at a lower cost"; GPT-6.1 Terra for visual reasoning and computer use; GPT-6 Sol for coding and agentic workflows; GPT-6 Luna for high-volume work; and GPT-6 Cyber for vulnerability investigation through the Daybreak setup. There is no 6.2 in that list, and a request for a model page at the conventional documentation path returns a 404.

Screenshot of the OpenAI API model catalogue listing six GPT-6 text identifiers — GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, GPT-6 Sol and GPT-6 Cyber — with no GPT-6.2 entry.

The changelog fills in the shape of the gap. Since GPT-6 Astra shipped on 3 September, OpenAI has released along vertical lines rather than a single ascending text line:

• 22 September — GPT-6 Sol and GPT-6 Luna added as a coding and a volume tier under Astra

• 29 September — Astra and Sol offered on the Ultrafast service tier, at six times standard rates

• 8 October — GPT-6.1 Sol on Ultrafast, the first time a mid-tier model got the fast lane

• 16 October — nine new GPT-Live voices and prompt-to-voice; Ultrafast mode extended to the Agents API

• 22 October — the Decisions API generally available on GPT-6 Luna, with the per-request question cap raised from 200 to 400

• 26 October — GPT-6.1 Terra, aimed at visual reasoning and computer use rather than raw benchmark scores

• 30 October — GPT-6 Rosalind, biological reasoning, gated behind the trusted-access programme

• 2 November — outbound SIP calling on GPT-Live; 4 November — GPT-Image-2.6 Sunburst and Flare, with SVG and animated GIF output

• 5 November — GPT-6 Cyber, for vulnerability work, behind Daybreak onboarding

Eleven days of changelog since then contain no model at all: a wallet, business features, education guidance, a cyber-defence partnership, a physics post. If compute were being withheld from serving wholesale, this is not the pattern you would expect — images, voice, telephony, life sciences and a security model all shipped. What did not ship is a new frontier text checkpoint. That distinction is the whole argument, and it is one the tweet flattens.

Three explanations, and only one with paperwork

The post names three or four causes and does not weigh them. Compute scarcity, safety, product readiness and cost are all plausible; only one of them has a published, inspectable mechanism attached, and it is not serving.

On 9 November, OpenAI published the Control Room: "a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise." The specifics matter more than the framing. It "halts model output within 35 seconds of a decision to intervene," and the company states the maximum time between an instruction to stop and the point where further progress is no longer possible is 35 seconds, with a stated goal of driving that to zero. It fails closed. It requires that "every research workload that does agentic inference" be "associated with an approved safety case," and that "all of our frontier reinforcement learning jobs run with a monitor."

That is a system for allocating and revoking research compute. It is not a system for scheduling releases, and the document says so almost in passing: "Our external model-serving stack is already much more centralized by nature." Read carefully, the sentence is doing a lot of work. It means the thing that would have to be rationed for the tweet's theory to hold — serving capacity — is described as already consolidated, and the thing that is being actively governed, throttled and stop-switched is research inference. The public record documents contention inside research. It does not document a queue of finished checkpoints waiting for serving capacity.

The supporting documents point the same way. On 28 September, OpenAI argued that "structured safety documentation should be required before continuing any frontier reinforcement learning training run" — a gate placed in front of training, not in front of release, with a pre-mortem dissent, senior-leader veto rights, and a mandate that monitoring and auto-pausing "fail closed." On 23 October it disclosed the incident that gave those rules teeth: a misconfigured DNS resolver on 20 September, training and evaluation of its most capable models paused, and reinforcement learning resumed on 1 October — "11 days after our pause began." A company that will stop its own frontier training for eleven days over a resolver is a company whose binding constraint on the research side is legible. Its binding constraint on the serving side is not documented anywhere.

The one number that is falsifiable

Compute scarcity is the part of the tweet that should be easiest to test, because OpenAI has published a great deal about physical capacity. On 29 April it said it had already surpassed the 10-gigawatt Stargate commitment it made for 2029, with "more than 3GW added in the last 90 days alone," and that "the only responsible way to meet it is to build more compute, faster." On 5 November, in a document about data-centre conduct, it disclosed that at its planned Effingham County campus "our agreement with Georgia Power allows the utility to reduce the power it delivers to us by up to 1,000 megawatts during periods of high demand."

Those two facts sit oddly beside "scarce compute" as an explanation for a delayed release. A company adding three gigawatts in a quarter and accepting a contractual right for a utility to curtail a gigawatt is not obviously short of megawatts; it is exposed to peak-demand events and to the cost of firm power. The 5 November document is also where the cost side of the tweet becomes concrete rather than rhetorical: OpenAI commits to paying the full cost of the electricity it uses, posting financial assurances before major spend so that other utility customers are protected if a project is delayed or cancelled, and — unusually — to not use non-disclosure agreements with public officials except for narrow security and trade-secret carve-outs. Cost and compute are, in that framing, the same constraint viewed from two ends, and both are being managed with published commitments rather than with release timing.

What a real GPT-6.2 launch would look like

Every OpenAI text model in the current line arrived with the same three artefacts, and their absence is a better predictor than any teaser: a row on the rate card, a page in the model catalogue, and a dated changelog entry describing what the model is for. GPT-6 Astra has all three. GPT-6.1 Terra has all three, dated 26 October. GPT-6 Cyber has all three, dated 5 November. When 6.2 ships, it will acquire a per-million-token price before it acquires a benchmark table — that is how this company has operated all year, and it is why the rate card is the place to watch, not X.

There is also a version-number problem the tweet does not address. GPT-6.1 was not a capability jump over GPT-6 in the way the decimal implies; it was a cost-and-latency repositioning, with 6.1 Sol priced at $2 input and $10 output per million tokens against Astra's $10 and $50, and 6.1 Terra at $1 and $5. If 6.2 is another repositioning rather than a new frontier — and there is nothing in the public record saying which it is — then "prioritising research over serving" is not an explanation for its absence. There would be nothing withheld to serve.

The serving side, which is priced and measurable

Whatever is being allocated internally, the models that exist are sold on a published rate card, and that card is the honest basis for any cost comparison right now. These are OpenAI's own figures, per million tokens, standard tier:

• GPT-6 Astra — $10 input, $50 output; cached input $1.00; long-context requests above the threshold $20 and $75

• GPT-6.1 Sol — $2 input, $10 output

• GPT-6 Sol — $2 input, $10 output

• GPT-6.1 Terra — $1 input, $5 output

• GPT-6 Luna — $0.10 input, $0.50 output

• Service-tier multipliers — Batch and Flex bill at half the standard rate; Fast and Priority at double; Ultrafast at six times

Screenshot of the OpenAI API pricing page showing the GPT-6 text rate card: GPT-6 Astra at $10 input and $50 output per million tokens, GPT-6.1 Sol at $2 and $10, GPT-6.1 Terra at $1 and $5, and GPT-6 Luna at $0.10 and $0.50. No GPT-6.2 row appears in the table.

The interesting column is not price but price against measured output. On Artificial Analysis's index — independent of OpenAI, but measured on its own task set, so treat it as a second opinion rather than a verdict — GPT-6 Astra scores 56.85 at maximum reasoning effort and 54.85 at high, at a cost to run the index of $6.24 and $3.24 per task respectively. GPT-6.1 Sol scores 54.61 at maximum effort for $1.27, and 50.33 at high effort for $0.59. That is the arithmetic that makes a 6.2 interesting: Sol already reaches within two and a quarter points of Astra at roughly a fifth of the cost per task, so the marginal value of another frontier checkpoint is narrower than the version number suggests, and the marginal value of a cheaper one is obvious.

For scale from outside OpenAI: Claude Fable 5.2 scores 64.56 at maximum effort but costs $26.54 per task on the same measure, and 61.82 at high effort for $9.08. Gemini 4 Carbon scores 58.26 for $4.68. On this evidence OpenAI's gap to the leader is real, and closing it is a research problem, not a serving-capacity problem — which is, ironically, the part of the tweet that holds up best.

This is where routing does something concrete. Four of the six GPT-6 text identifiers run through OrcaRouter today — openai/gpt-6-astra, openai/gpt-6.1-sol, openai/gpt-6-sol and openai/gpt-6-luna — at OpenAI's own list prices, with no markup, so a vendor price change is live on our side the day it lands rather than at the next contract renewal. GPT-6.1 Terra and GPT-6 Cyber are not in our catalogue, and nothing here should be read as claiming they are. What the pass-through buys a team waiting on 6.2 is a way to price the fallback: run the workload against Astra at high effort and against 6.1 Sol at maximum effort on the same key, compare the two cost-per-task numbers against your own acceptance criteria, and set automatic failover so an availability wobble on one model does not become an incident. That is a decision available this week. Waiting for a checkpoint is not.

What would change this reading

Two things would, and both are observable. A GPT-6.2 row appearing on the rate card while the capability documentation lags would support the tweet's economics — compute going to research while a finished model waits its turn for serving. Conversely, a 6.2 launch with a fresh benchmark table and a price above Astra would suggest the opposite: that the constraint was capability, and the version number was waiting on the research, exactly as the governance documents describe.

Until one of those happens, the honest position is the uncomfortable one. The tweet's falsifiable half is confirmed — no GPT-6.2 exists, eleven days after the last model shipped — and its causal half is unverified, asserted by someone outside the company about a decision made inside it. What the public record does show is a company that has spent seven weeks publishing the machinery it uses to ration and stop research compute, down to a 35-second stop bound, and that has published almost nothing about how it decides which finished model gets served first. The compute is being governed. Whether it is being withheld from you is a different claim, and OpenAI has not made it.

Compared in this article4

Detected from this article · Benchmarks: Artificial Analysis · updated daily