A generated title card reading 'GPT-6.2 Sol vs GPT-6.1 Sol', subtitled 'Same price, same context, two index points apart — and one latency figure that has not been published', above three numbered rows: the rate card, where both models bill $2 per million input tokens, $0.10 cached and $10 output; the independent board, where GPT-6.2 Sol scores 57 and GPT-6.1 Sol 55 on Artificial Analysis Intelligence Index v5.0.1 at max effort, at $1.01 and $1.27 per index task respectively; and the open question, that GPT-6.2 Sol has no published time-per-task figure while GPT-6.1 Sol is measured at 23.1 minutes.
Guides & Insights

GPT-6.2 Sol vs GPT-6.1 Sol: Same $2, Two Index Points, and the One Number Nobody Has Published

Author

jinhao.song

Date Published

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

OpenAI shipped GPT-6.2 Sol on 17 November 2026 at exactly the price of the model it replaces. GPT-6.1 Sol has cost $2 per million input tokens and $10 per million output since 29 September; GPT-6.2 Sol costs $2 and $10. Cached input is $0.10 on both, a 30-minute cache write is $2.50 on both, the 1,050,000-token context window is identical, and both reject none as a reasoning effort. If you compare these two models the way most teams compare models — by the sticker on the rate card — there is nothing to compare. The difference is in what a finished task costs and what it takes to move an integration across, and only one of those two numbers is currently published.

The rate card did not move, and that is the actual news

GPT-6.2 Astra, the flagship released the same day, got a real price cut: input and output down 20% from GPT-6 Astra and cached input down 60%. The Sol line got nothing. OpenAI's launch post says it plainly — GPT-6.2 Sol "keeps GPT-6.1 Sol's API token prices" — and puts the entire argument for the new model on efficiency: better results on demanding tasks "with fewer output tokens, so you can get more done within your usage limits."

Here is the spec contrast that matters, one line per dimension, from OpenAI's own model pages:

• Price — GPT-6.2 Sol $2 input / $0.10 cached / $2.50 cache write / $10 output, per million tokens; GPT-6.1 Sol identical on all four.

• Context — 1,050,000-token window, 922,000 max input, 128,000 max output on both; text and image in, text out on both.

• Knowledge cutoff — 1 June 2026 on GPT-6.2 Sol versus 30 April 2026 on GPT-6.1 Sol. Thirty-two days of extra world knowledge, and the only capability difference either vendor page states outright.

• Long context — above 272,000 input tokens, input and cache rates double and output costs 1.5× for the whole request, on both models. The cliff is in the same place at the same steepness.

• Reasoning effort — low, medium, high, xhigh and max on both; none rejected by both, which is a change from GPT-6 Sol but not from 6.1.

• Endpoints — Chat Completions, Responses and Batch supported; Realtime, Assistants and fine-tuning not. Tool calling requires the Responses API on both.

• Tools — web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search, all supported on both.

• Cloud availability — the GPT-6.1 Sol page lists openai.gpt-6.1-sol on Amazon Bedrock and a Google Cloud private preview. The GPT-6.2 Sol page has no cloud-provider section at all, which reads as OpenAI's API plus ChatGPT Work and Codex for now.

• Cache write, 12-hour retention — $5 per million tokens on GPT-6.2 Sol for short-context requests at Standard pricing. GPT-6.1 Sol's page documents the 30-minute write at $2.50 and carries no 12-hour rate, so on that model the longer window is not priced.

What the two points cost per finished task

Artificial Analysis has already run both models, and its Intelligence Index is the only independent comparison available for a model this new. On Index v5.0.1 — nine evaluations spanning agentic knowledge work, terminal coding, open-source software engineering, document reasoning, physics and scientific workflows — the numbers are these:

• GPT-6.2 Sol, at max effort — index 57, eleventh of 87 models on the board, $1.01 weighted cost per index task, 60.9M output tokens to complete the suite.

• GPT-6.1 Sol, at max effort — index 55, seventeenth of 87, $1.27 per task, 63.7M output tokens.

Same token prices, two points of index, and a 20% lower cost per task. Those three facts do not sit together unless you accept that the per-task figure is a weighted blend of input, cached input, reasoning and answer tokens rather than a simple multiple of the output count — the published output-token drop is only 4.4%, from 63.7M to 60.9M. Both numbers are measurements, not projections; neither is a vendor claim.

For a workload of any size, that 20% is the whole business case. It is also the reason a team that budgets by dollars-per-million-tokens will conclude the two models are indistinguishable and miss it entirely.

A generated two-column comparison scoreboard. Left column 'GPT-6.2 Sol' reads: Index 57, $1.01 per task, 60.9M output tokens, time per task not published, $2 in and $10 out, knowledge cutoff 1 June 2026. Right column 'GPT-6.1 Sol' reads: Index 55, $1.27 per task, 63.7M output tokens, 23.1 minutes per task, $2 in and $10 out, knowledge cutoff 30 April 2026. A footer line states that board figures are per Artificial Analysis Intelligence Index v5.0.1 at max effort and prices are per OpenAI.

The number that is missing

One figure on the GPT-6.1 Sol entry has no counterpart on the new one. Artificial Analysis measures GPT-6.1 Sol at 23.1 minutes per Intelligence Index task, running at 49 output tokens per second with 156.21 seconds to the first answer token. For GPT-6.2 Sol the performance panel reads N/A: the board has scored the model and published its cost, but not how long it takes.

That gap matters more than its size suggests. GPT-6.1 Sol's 23.1 minutes is already the slowest figure in its tier — the board calls it "notably slow per task" — so the question a latency-sensitive team needs answered is whether the successor is faster, slower, or the same, and nobody has answered it yet. A 20% cheaper task that takes 40% longer is a worse trade for anything sitting in front of a user.

Two related cautions on reading that page today. The Artificial Analysis chart still carries a "No publicly available API" placeholder beside the intelligence panel for GPT-6.2 Sol, even though the model has been on OpenAI's API since launch — the summary panel above it holds the real numbers, and the placeholder is stale. And every score on the board is the max-effort configuration, so the index figures describe the top of the effort dial, not the default your application will use.

Screenshot of the Artificial Analysis page for GPT-6.2 Sol, showing an Intelligence Index of 57 at max effort with the model ranked eleventh of 87, a weighted cost per index task of $1.01, 60.9M output tokens used for the full suite, and a performance panel whose time-per-task, output-speed and time-to-first-token figures read N/A — beside a stale 'No publicly available API' placeholder.

Ultrafast is the one line item with a clock on it

There is a second price difference hiding in the fine print, and it is time-limited rather than cheaper. Ultrafast — OpenAI's accelerated inference tier, available on the API for both 6.2 models at launch — is billed at 3× standard rates on GPT-6.2 Sol through 28 February 2027, then 6× from 1 March 2027. That is the only Ultrafast rate the Sol line currently publishes: the GPT-6.1 Sol model page carries no accelerated tier in its price table, so there is no documented 6.1 number to set against it, only the 6× that the 6.2 rate reverts to.

On a 100,000-token request with a 4,000-token answer, standard is $0.20 of input plus $0.04 of output, or $0.24 on either model. The same request on GPT-6.2 Sol's Ultrafast tier is $0.72 until the end of February and $1.44 from March. The honest reading is not that the new model is cheaper on this tier — nothing published says that — but that OpenAI has put a four-month discount on the accelerated tier and dated it. If your reason for looking at 6.2 is interactive latency and you would pay the premium to get it, the window is worth using before it closes.

The migration bill that is not on the rate card

Moving a production integration from 6.1 Sol to 6.2 Sol is not a model-string change, even though the price is a non-event. OpenAI's GPT-6.2 migration guide lists four things that will break or degrade, and only one of them is loud enough to be caught in a smoke test.

The quiet one is reasoning state. Starting with GPT-6.2, the API drops reasoning items whenever it detects a mid-conversation change to the history — a prefix edit, a model downgrade, an end-user change, an organisation change. Lose those items and "answer quality can decline and the model may use more tokens to repeat earlier work." A team that edits a system instruction between turns, or that replays a conversation under a different end-user ID for evaluation, will not see an error. They will see a slightly worse answer and a larger bill, which is the hardest kind of regression to attribute.

The loud ones are ordinary API hygiene: omit temperature, top_p, logprobs and top_logprobs from Chat Completions requests, and drop temperature, top_p and top_logprobs from Responses requests, removing message.output_text.logprobs from include. If you are migrating from GPT-5.5 or earlier in the same pass, replace prompt_cache_retention with prompt_cache_options.ttl set to "30m". Zero Data Retention is supported on 6.2 for approved organisations on eligible endpoints; US and EU data residency are supported for Standard, Fast and Batch, but Ultrafast is US-only. None of that is expensive. All of it is easy to miss if the migration plan is "swap the string and watch the error rate."

Two boards disagree about whether 6.1 Sol is retired

Artificial Analysis marks GPT-6.1 Sol deprecated and recommends the successor. OpenAI's own lifecycle table tells a different story: it still lists gpt-6.1-sol as the recommended replacement for gpt-5, gpt-5-pro, o3 and o3-pro, all of which shut down on 11 December 2026. A model being handed a wave of migrations three weeks from now is not a model on its way out.

The practical reading is that GPT-6.1 Sol is being superseded in the catalogue sense while remaining a supported destination in the migration sense. That makes it a legitimate fallback rather than a legacy liability — which is the whole reason a rollout of the new model can be reversible.

Running the pair side by side without a second contract

The reason a side-by-side trial is cheap here is that the expensive thing — the invoice — does not change. GPT-6.1 Sol is on OrcaRouter's catalogue at the vendor's own $2 in, $0.10 cached and $10 out, passed through with 0% markup on top, so the September model stays reachable on the same key while you evaluate the November one. Automatic failover is what turns the rollout into a routing decision instead of an incident: pin the workload to GPT-6.2 Sol, point the fallback at GPT-6.1 Sol, and a request that the newer model refuses or fails degrades to the model you already trusted instead of returning an error.

One honest caveat about the new model's home. GPT-6.2 Sol is not on our catalogue yet — the 6.2 generation arrives on OpenAI's own API first, and any claim that you can call it here today would be wrong. So the trial has a shape: route GPT-6.1 Sol through us, call GPT-6.2 Sol on the vendor's API, and keep the two on one routing policy so the comparison is a policy change rather than a rewrite. When the 6.2 models land in the catalogue, the swap is the same kind of edit, in the other direction.

That split is also the sane answer to the cyber question. OpenAI shipped both 6.2 models with tighter safeguards for higher-risk cybersecurity assistance, and its own launch post concedes the trade: the protections "reduce misuse risk, but also restrict some legitimate vulnerability research and security testing workflows." Approved defenders currently get reduced refusals on GPT-6 Astra through Daybreak and GPT-6 Cyber on Daybreak Red; OpenAI says it "will work to bring GPT-6.2 models to Daybreak." Until that lands, a security workload on 6.2 Sol is one that can be refused, and a 6.1 fallback is what keeps it running.

Screenshot of the OrcaRouter model page for GPT-6.1 Sol, showing a 1,000,000-token context window, 128,000 max output tokens, text, image and file input, and pass-through pricing of $2.00 per million input tokens and $10.00 per million output tokens.

Switch, wait, or split

If your workload is batch, background, or anything where a task taking minutes is already priced in, move to GPT-6.2 Sol. You are getting two index points and a 20% lower cost per task on an unchanged rate card, plus a discounted Ultrafast tier until the end of February, and the migration work is the checklist above rather than an architecture change.

If your workload is interactive and your latency budget is tight, wait for one more number. GPT-6.1 Sol's 23.1 minutes per index task is a known quantity you have already built around; its successor's is unpublished, and no amount of per-task savings compensates for an answer that arrives after the user has left. Run 6.2 on a sample, measure your own p50 and p95, and let that decide rather than a board that has not published the figure.

If your workload carries cyber or dual-use risk, split deliberately. Keep 6.1 Sol as the path that answers and treat 6.2 Sol as the path that is better when it answers — because the refusal behaviour, not the price, is the variable that changed.

The thing to watch is unglamorous and specific: Artificial Analysis publishing a time-per-task figure for GPT-6.2 Sol. Everything else about this upgrade is already settled — the price is the same, the context is the same, the index moved two points and the per-task cost moved 20% in the right direction. Only the speed is unknown, and until it is known, "same price, strictly better" is a claim with one unanswered clause in it.

Compared in this article2

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