A generated hero card headed '126 tokens a second, not 4.5x' with a subtitle reading 'The first public measurement of Ultrafast, on GPT-6.2 Sol', three cards reading 'Measured average: 126 tok/s', 'Measured peak: 252 tok/s' and 'Standard baseline: 85 tok/s', and a footer reading 'One user's 27-minute video edit, 18 November 2026; Standard baseline per Artificial Analysis'.
Guides & Insights

GPT-6.2 Sol Ultrafast, Measured: 126 Tokens a Second and No Sign of the 4.5×

Author

jinhao.song

Date Published

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

The first public number for GPT-6.2 Sol running with Ultrafast switched on is 126 tokens per second, averaged across a job that took 27 minutes of wall clock and about 14 minutes of actual model response, with a peak burst of 252 tokens per second. It comes from a video editor on X who asked the model to cut an eight-minute video — a real task, reported in four lines, with no harness, no repeat runs and no Standard-mode control. Set that average against the figure Artificial Analysis publishes for the same model on Standard, 85 output tokens per second, and the tier OpenAI says generates tokens "at up to 4.5 times the speed of Standard" landed at roughly 1.5× in this one report. GPT-6.2 Astra and GPT-6.1 Sol, the other two models on the Ultrafast list, sit lower still on that independent board at 60 and 49 tokens per second, which is why the ratio matters more here than the raw number.

Four numbers, and the fifth one nobody posted

The post is unusually specific for a social-media speed claim, which is what makes it worth taking apart. It gives the task (an eight-minute video edit), the total time (27 minutes), the portion spent waiting on the model (about 14 minutes), the average generation rate (126 tokens per second) and the best single burst (252). OpenAI documents the tier as Ultrafast; the post calls it "Ultra Fast". Both names describe the same thing — the fastest speed option in the picker in ChatGPT Work and the Codex app.

What the post does not give is the half that would make it a benchmark:

• Which surface it ran on — ChatGPT Work, the Codex app, or the API. All three expose the tier; the numbers are not comparable across them.

• The reasoning effort setting. Sol at Light and Sol at Pro are different amounts of thinking before the first token, and thinking time is inside that 14 minutes.

• How many tokens were actually generated. It can be inferred from the average and the duration, and the inference is worth doing — see below — but it is an inference.

• What Standard would have taken on the same task. This is the number the piece is really about, and it is the one number the post cannot contain, because it is a single run with nothing to compare against.

• Whether the edit was usable first time. "It cut an eight-minute video" is a success claim, not a quality claim, and a re-run would have changed the total time substantially.

The claim, stated exactly, and what the tier costs

OpenAI's speed documentation is precise about what Ultrafast promises: "On GPT-6.2 Astra and GPT-6.2 Sol, Ultrafast generates tokens at up to 4.5 times the speed of Standard on the same model." Note the two qualifiers. It is a claim about token generation, not about the time to first token and not about the time a whole task takes. And it is "up to", which is a ceiling rather than a typical value — a distinction that matters enormously when the first independent reading comes in at roughly a third of that ceiling.

On the API it is a flag, not a model: set model to gpt-6.2-sol and service_tier to "ultrafast". OpenAI recommends WebSockets for agentic work so that connection overhead does not eat the gain, and Ultrafast carries its own rate limits, separate from Standard and Fast. For the GPT-6.2 models those limits are not published — the docs tell you to check your organisation's own settings — while GPT-6.1 Sol Ultrafast publishes 1,000,000 tokens per minute on the Build tier and GPT-6 Astra publishes 500,000. Inference residency for the GPT-6.2 tier is United States only. In ChatGPT, Pro $200 gets Ultrafast on Sol alone and Pro $500 gets both models.

The price is the part with a deadline on it. Ultrafast on GPT-6.2 Sol is $6.00 per million input tokens, $0.30 cached input and $30.00 output during the promotional period, and $12.00 / $0.60 / $60.00 at list — exactly 3× and 6× the Standard rate of $2.00 / $0.10 / $10.00. OpenAI dates the switch: "Ultrafast costs 3 times the Standard rate through February 28, 2027, and 6 times the Standard rate starting March 1, 2027 (Pacific time)."

A screenshot of OpenAI's ChatGPT speed documentation describing Ultrafast mode, showing the statement that on GPT-6.2 Astra and GPT-6.2 Sol Ultrafast generates tokens at up to 4.5 times the speed of Standard, the recommendation to use WebSockets, and the note that Ultrafast has separate rate limits from Standard and Fast.

The baseline that makes 126 legible

A tokens-per-second figure means nothing without a reference, and there is exactly one independent reference for this model family. Artificial Analysis measures GPT-6.2 Sol at max reasoning effort at 85 output tokens per second, a 68-second time to first token, 7.9 minutes for a completed index task, and $1.01 per task, against an Intelligence Index of 57. The same board puts GPT-6.2 Astra at 60 tokens per second and an Index of 62, at $4.00 per task, and GPT-6.1 Sol at 49 tokens per second and an Index of 55, at $1.27 per task.

Those are harness medians on a fixed evaluation set, and the video edit is one person's afternoon. They are not the same measurement, and the honest way to use them together is as a rough scale factor rather than a verdict:

• Measured average, 126 tokens per second — about 1.5× the Standard baseline of 85.

• Measured peak, 252 tokens per second — about 3.0× that baseline, on the single fastest burst.

• The 4.5× claim applied to 85 — a little over 380 tokens per second, which nothing in the report approaches.

• Artificial Analysis's own GPT-6.2 Sol page contains no mention of Ultrafast at all, so its 85 is a Standard-mode figure by default rather than by assertion.

Why four and a half times the tokens is not four and a half times the job

This is the part that gets lost when a speed tier is sold as a multiplier, and the video post accidentally contains the proof. Fourteen minutes of model response out of a 27-minute job means 52% of the clock was generation and 48% was everything else — upload, transcoding, the editor's own steps, tool calls, whatever the pipeline does between model turns. Treat that split as the shape of the workload and the arithmetic becomes unforgiving:

• The generation time implies roughly 106,000 output tokens, at 14 minutes and 126 tokens per second.

• At the Standard baseline of 85 tokens per second, those tokens take about 21 minutes, so the job lands near 34 minutes.

• At the full 4.5× claim — about 383 tokens per second — they take under 5 minutes, so the job lands near 18 minutes.

• The run actually took 27 minutes, almost exactly halfway between the two.

Even if the tier had delivered every bit of the 4.5×, the end-to-end gain on this workload would have been about 1.5×, because the model was only ever half the job. Amdahl's law does not care how fast the fast part gets. That is not an argument against Ultrafast; it is an argument for measuring the tier against your own pipeline shape rather than against the multiplier, because an interactive chat turn is nearly all generation and a media pipeline is not.

What those 14 minutes actually cost

Apply the promotional rate card to the inferred 106,000 output tokens and the price of the measurement is legible in one line:

• Standard, $10.00 per million output tokens — about $1.06.

• Ultrafast at the promotional rate, $30.00 per million — about $3.18.

• Ultrafast at list from 1 March 2027, $60.00 per million — about $6.35.

Output is only half the bill, and on a video-editing agent it is probably the smaller half. Every frame description, transcript and tool result goes back in as input at $6.00 per million during the promotion, and an agent looping over the same context is exactly the workload the cache-write rows exist for — $7.50 per million for a 30-minute cache and $15.00 for a 12-hour one, against $0.30 for a cache read. A pipeline that re-sends its context instead of caching it can spend more on input than on the output it was trying to speed up. The post does not report token counts, so this cannot be totalled, only bounded: the tier multiplied the generation half of the bill by three, and will multiply it by six in fourteen weeks.

Which model to put the tier on

The same 4.5× applied to three different baselines gives three very different destinations, and the cheapest one is not the fastest one.

• GPT-6.2 Sol — 85 tokens per second at Standard, so the claim implies about 383. Index 57, $1.01 per task.

• GPT-6.2 Astra — 60 at Standard, so the claim implies about 270. Index 62, $4.00 per task, four times the cost for five more index points.

• GPT-6.1 Sol — 49 at Standard, so the claim implies about 220. Index 55, $1.27 per task, and its Ultrafast tier is priced flat at $12.00 input and $60.00 output per million with no promotional row at all.

Two consequences follow. First, Ultrafast on Astra is still slower than Ultrafast on Sol, because it is starting from a lower base — if the goal is throughput, the stronger model does not get you there. Second, GPT-6.1 Sol's Ultrafast is already priced at what GPT-6.2 Sol's Ultrafast will cost after February, which makes it a fair preview of the bill rather than a discount. OpenAI also notes that Ultrafast on GPT-6 Astra and GPT-6.1 Sol is available to all API users, while the GPT-6.2 tier points you at your organisation's rate limits — worth checking before you build a queue around it.

A screenshot of the Artificial Analysis model page for GPT-6.2 Sol, showing an Intelligence Index of 57, an output speed of 85 tokens per second, a time to first token of 68 seconds, 7.9 minutes per index task, and a cost of $1.01 per task.

Getting your own number instead of borrowing this one

The measurement above is a single report and should be treated as a signal, not a specification. The way to resolve it is a controlled A/B on your own workload — same prompt, same task, three runs on Standard and three with service_tier set to "ultrafast", recording end-to-end minutes, total tokens and the time to first token separately. Report the peak as well as the average, because the gap between 126 and 252 in the original post is the difference between a tier that helps and a tier that occasionally helps.

OrcaRouter does not serve GPT-6.2 Sol or GPT-6.2 Astra today; the catalogue carries openai/gpt-6-astra, openai/gpt-6-sol, openai/gpt-6.1-sol and openai/gpt-6-luna, and the GPT-6.2 model pages do not resolve. So this is not a suggestion that you route the new model through us. What is true today is narrower and still useful for exactly this decision. GPT-6.1 Sol is on the catalogue, it is the closest model to the one measured, and running your evaluation set against it through a single key is the cheapest way to learn whether a speed tier changes your product before you commit a budget to a tier that doubles in March. Pricing is the provider's list rate passed through with no markup, so the 1 March 2027 increase arrives on our side the same day it arrives on OpenAI's — no repricing cycle, no renegotiation.

There is also a routing argument specific to this tier. Ultrafast has its own rate limits, separate from Standard and Fast, and OpenAI publishes no default tokens-per-minute figure for the GPT-6.2 tier — you find out your ceiling by hitting it. A pipeline that treats Ultrafast as the only path inherits that unknown; a pipeline that can fall back to Standard on the same model degrades in speed instead of failing. That is the one structural advantage of putting a speed tier behind a router rather than hard-coding it, and it is worth having before the tier is on a critical path.

A generated single-column scoreboard for GPT-6.2 Sol Ultrafast with rows reading 'Measured average: 126 tok/s', 'Measured peak: 252 tok/s', 'Standard baseline: 85 tok/s', 'Vendor claim: up to 4.5x Standard', 'Output price, promo: $30 per MTok' and 'Output price, list: $60 per MTok', with a footer noting that the measured figures are one user's task report, the baseline is per Artificial Analysis, and the claim and prices per OpenAI.

What would settle it

One number would move this from anecdote to evidence: a Standard-versus-Ultrafast A/B on a fixed task, published with tokens and wall clock. Until that exists, the defensible reading of 18 November 2026 is that Ultrafast on GPT-6.2 Sol is real and materially faster than Standard — 126 against 85 is not noise, and a 252-token burst is not a rounding artefact — but that the published multiplier is a ceiling nobody has come close to in public, and that a task-level speedup depends as much on what fraction of your pipeline is token generation as on the tier itself. The person who posted the measurement got a video cut in 27 minutes instead of an estimated 34, and did it on a launch-week tier that will cost twice as much in March. That is a good outcome and a modest multiplier, and both halves are worth knowing before anyone budgets on 4.5×.

Compared in this article1

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