
GPT-6.2 Sol Pro: The Rate Card Didn't Move. The Bill Fell 20%.
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 124 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 132 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 131 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 185 tok/s
Forty thousand output tokens to finish one evaluation task. That is what GPT-6.2 Sol spent on Artificial Analysis's suite at its highest effort setting, against 67,824 for GPT-6.1 Sol doing the same work — a 40% drop in tokens and a 20% drop in dollars per task, on a price list that did not change by a cent. The model shipped on 17 November 2026, alongside GPT-6.2 Astra, and the label people are searching for, GPT-6.2 Sol Pro, is that model with reasoning.mode set to pro in the Responses API. Pro mode bills at the base model's standard token rates. So the entire economics of the upgrade sit in how many tokens the model decides to spend — which is the one number OpenAI's rate card cannot tell you.
This is not a re-run of the "Pro is a mode" argument that has been made twice already on this blog, for the GPT-6 and GPT-6.1 tiers. The mechanism is the same and it is now well documented. What is new is the evidence: for the first time there are independent measurements of a Sol tier on the day after launch, and they say something more interesting than "cheaper per token". They say the improvement is concentrated in exactly the workloads OpenAI is selling this model for, and that one dimension got worse.
What shipped on 17 November, precisely
The changelog entry is two lines long and names two identifiers: gpt-6.2-astra and gpt-6.2-sol, both on the Responses API and Chat Completions, with tool calling requiring Responses. OpenAI's own summary of the release — "better professional outputs, coding, computer use, and efficiency at lower cost across ChatGPT Work, Codex, and the API" — is vendor framing, and it is worth holding at arm's length until the independent numbers arrive, which they now have.
The shape of the thing, from OpenAI's model page for gpt-6.2-sol:
• Context — 1,050,000 tokens total, 922,000 maximum input, 128,000 maximum output.
• Modalities — text and image in, text out. Audio and video are listed as unsupported, and there is no file-input modality in the specification.
• Knowledge cutoff — 1 June 2026, one month later than the 6.1 tier's 30 April 2026.
• Reasoning — always on, five effort settings from low through max, and the API rejects reasoning.effort: none outright rather than silently downgrading.
• Price, short context — $2.00 input, $0.10 cached input, $2.50 cache write, $10.00 output per million tokens. Identical to GPT-6.1 Sol.
• Price, above 272,000 input tokens — input and cache rates double and output goes to 1.5× for the whole request: $4.00 / $0.20 / $5.00 / $15.00.
• Service tiers — Batch at 50% of standard, Fast at 2×, Ultrafast at 3× through 28 February 2027 and 6× from 1 March 2027.
• New on this generation — discounted background completion windows in the Responses API public beta, 30% below standard rates for a two-hour window and 40% below for eight hours.
That last line is quietly the most useful thing on the release for anyone running batch-shaped agent work, and it is the kind of detail the launch post buries. A two-hour deferral that costs 30% less is a different product decision from a price cut, because it changes when you schedule the work rather than what you pay for it.

Where "Pro" comes from, and why it is not a model ID
OpenAI's reasoning guide states it plainly: GPT-5.6 and GPT-6 models "support standard and pro reasoning modes in the Responses API. standard is the default." Mode and effort are independent axes — the guide is explicit that mode "selects standard or pro execution, while reasoning.effort controls how much reasoning the model applies within that mode". You can run pro mode at low effort; you can run standard mode at max effort.
What pro execution does is aggregate several passes of model work into a single answer and bill those tokens at "the selected model's standard token rates". No premium multiplier, no separate meter. Existing Pro identifiers — GPT-5.5 Pro and its predecessors — keep their old behaviour and their old prices, which is why "Pro" now means two different things depending on which string you name.
So when a catalogue lists gpt-6.2-sol-pro, it is describing a configuration and inventing an identifier to hang it on. OpenAI's own docs 404 on the string, and the only identifier on the model page is gpt-6.2-sol. That distinction matters for more than pedantry: a model ID you can look up has a documented context window and a published rate, and a configuration has whatever your application passes at call time.
The independent read: up 2.7 points, 20% cheaper per task
Artificial Analysis has already run the model on its own harness, and the numbers are the first non-vendor measurements of the tier. On Intelligence Index v5.0.1, GPT-6.2 Sol at max effort scores 57.3 against 54.6 for GPT-6.1 Sol at max effort. Cost per Index task falls from $1.27 to $1.01. Both figures are measured over the identical suite, which is what makes them comparable: same tasks, same harness, same scoring, only the model changes.
Those two numbers moving in the right direction together is the real claim of this release. A generation that gets smarter and cheaper per task at a fixed rate card is an efficiency gain, not a repositioning. A generation that gets smarter while costing more is a trade. This one is the former, by a wide enough margin to notice.
Where the gain lands is more interesting than the composite. Against its own predecessor at max effort:
• Terminal-Bench 4.0 — 60.1% vs 56.1%.
• Terminal-Bench Science — 67.6% vs 58.1%, a 9.5-point jump and the largest single move on the board.
• HLE Diamond — 64.0% vs 58.8%.
• FrontierCode 2.0 — 44.4% vs 40.6%.
• AutomationBench — 71.3% vs 64.9%.
• AnalystAgent — 53.8% vs 50.0%.
Every one of those is a tool-using, multi-step, verification-heavy task. This is not a model that got better at trivia; it is a model that got better at finishing things, which is consistent with the token count falling rather than rising. Fewer wrong turns means fewer tokens spent recovering from them.
One caveat about what these numbers are. Artificial Analysis measures effort levels, not reasoning modes, so its "max" row is standard mode at maximum effort. Nobody has published a pro-mode measurement of any GPT-6 model, because the mode multiplies an unknown quantity of model work. Treat the max-effort row as the closest available ceiling, not as a pro-mode benchmark.
Two numbers that moved the wrong way
AA-Omniscience is the part of the suite that tests whether a model knows what it does not know, and it is where this release gives something back. GPT-6.2 Sol at max effort scores 40.4 on the Omniscience index, down from 42.0. Accuracy is flat — 61.6% against 61.8%, which is noise — but the hallucination rate rises from 51.9% to 55.4%.
Read that carefully before deciding it is a scandal. Both models hallucinate on more than half of the questions they get wrong, and a 3.5-point move on a suite of this size is not a catastrophic regression. But it is a direction, and it points the opposite way from the capability numbers. The tier got better at doing work and slightly less careful about asserting things it cannot support.
For a coding agent that runs tests before declaring victory, that trade is close to free. For a research assistant that answers without a verification step, it is the exact failure mode you were paying the previous generation to avoid. If your workload is in the second category, the composite index is the wrong number to shop on, and the 20% saving is smaller than the cost of one confident wrong answer.
What pro mode does to the meter, in tokens
Because pro mode bills at standard rates, the only lever it pulls is token count, and the independent data lets you size that lever for the first time. At max effort, GPT-6.2 Sol averages 40,360 output tokens per task, of which 22,800 are reasoning tokens and 17,561 are visible answer — reasoning is 56% of the output. GPT-6.1 Sol at the same setting averaged 67,824 output tokens, 49,213 of them reasoning. The newer model thinks for less than half as long and answers in about the same length.
That is the mechanism behind the 20%, but it is not the whole arithmetic and the other half is worth knowing. Decomposing the per-task cost: output falls from about $0.72 to $0.40, while input rises from about $0.55 to $0.61 — more cache reads and cache writes per task, not fewer. The two roughly cancel into the net 20% saving, which means the upgrade is a bet that thinking less is worth more than the extra context it re-reads. On the suite, it was.
Run pro mode on top of either and the multiplier is unknown — OpenAI documents that pro "performs more model work than standard mode, increasing token usage and cost" but publishes no factor. What you can say with confidence is that the baseline it multiplies is 40% smaller on the output side, so the same pro multiplier lands on a cheaper floor.
Practical consequence: if you budgeted pro-mode calls against 6.1-era token counts, re-measure rather than rescaling. The token distribution moved enough that a linear adjustment will overstate your bill.

Where it is callable today, and where it is not
OpenAI's API is live: Responses and Chat Completions, batch, prompt caching, structured outputs, the hosted tool set including computer use, apply_patch, hosted shell and MCP, and Zero Data Retention for approved organisations on eligible endpoints. ChatGPT Work and Codex users got it on day one, per the release note. That is the GA picture as of 18 November 2026.
On OrcaRouter's own catalogue the 6.2 generation has not landed yet — a check of the public model list on 18 November returns openai/gpt-6.1-sol and openai/gpt-6-sol but no gpt-6.2-sol. Two things follow from that, and the second is the one worth planning around. First, nothing about the upgrade is available here today, and this article will not pretend otherwise. Second, OrcaRouter passes provider list price through with 0% markup, so when a tier does appear the vendor's own $2/$10 is what you pay — no gateway multiplier on top of the rate card, and a vendor price change is live the same day rather than on a reseller's schedule.

In the meantime, the tier this one replaces is openai/gpt-6.1-sol, which is on the catalogue at list price, and the migration path that makes sense is to keep production on the 6.1 tier while you evaluate 6.2 directly, then move the route rather than the code.
What to watch
Three things will decide whether 17 November looks like a step or a wobble. The calibration regression either persists across the next AA refresh or shrinks to noise. The Ultrafast multiplier doubles on 1 March 2027, from 3× to 6× standard, which triples nothing but does double the premium on the fastest path — if your architecture depends on Ultrafast, that date is in your budget. And the token efficiency has to hold under pro mode, where the one number nobody publishes is the one that decides your bill.
Until then, the honest summary is narrow and checkable: same price, 2.7 index points, 20% cheaper per task, and a measurably looser grip on what it does not know.
Compared in this article2
Detected from this article · Benchmarks: Artificial Analysis · updated daily
