
GPT-6.2 Sol Is Live at $2/$10: The Third Sol in Eight Weeks
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 184 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 228 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 131 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 77 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 69 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 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 · 208 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
OpenAI shipped GPT-6.2 Sol on 17 November 2026, and the most useful thing about it is the number that did not move: $2.00 per million input tokens and $10.00 per million output, the same sticker GPT-6 Sol carried on 22 September and GPT-6.1 Sol carried on 29 September. What moved is the measurement. Artificial Analysis reads GPT-6.2 Sol at 57 on its Intelligence Index at maximum effort — thirteen points above GPT-6 Sol at 44, two above GPT-6.1 Sol at 55 — at $1.01 per index task, against $2.47 and $1.27 for the two models it succeeds. GPT-6.2 Astra, the flagship that shipped the same afternoon at $8 / $40, reads 62. The release arrived as a news post time-stamped 17:45 GMT on 17 November, a changelog entry the same day, and two model pages listing a 1,050,000-token window identical to the September model's. This page started life in early September as a leak file on a Sol tier nobody had confirmed; three launches later it is the record of what each generation actually shipped, what the third one costs against the two before it, and which sentences on the earlier versions of this page are now simply wrong.
What shipped on 17 November, precisely
The changelog entry names two identifiers, gpt-6.2-astra and gpt-6.2-sol, both available on the Responses API and on Chat Completions, with tool calling requiring the Responses API; a second entry the same day adds both to the discounted background completion windows. The news post is titled "Introducing GPT-6.2 Astra and Sol" and describes better professional outputs, coding, computer use and efficiency "at lower cost" across ChatGPT Work, Codex and the API. Everything below comes from OpenAI's own model page and published rate card, which means it is vendor-stated rather than independently measured — except where a figure is attributed to the tracker.
• Identity and modalities — model ID gpt-6.2-sol, text and image in, text out, with audio and video unsupported
• Context envelope — 1,050,000-token window, 922,000 maximum input, 128,000 maximum output, unchanged from both earlier Sols
• Knowledge cutoff — 1 June 2026, a month later than GPT-6.1 Sol's 30 April and six weeks later than GPT-6 Sol's 20 April
• Reasoning ladder — low, medium, high, xhigh and max; the model page states that the API rejects requests setting effort to none, and the migration guide tells integrations currently passing none or minimal to start at low instead
• Endpoints — Chat Completions and Responses supported, Batch supported, and Realtime, Assistants, fine-tuning, image generation, video and every audio endpoint unsupported
• Tools — web search, file search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use, MCP and tool search, the same ten the September model lists
• Standard rates — $2.00 input, $0.10 cached input, $2.50 for a thirty-minute cache write, $5.00 for a twelve-hour one, $10.00 output, all per million tokens
• Long-context step — above 272,000 input tokens, input and cache rates double and output goes to 1.5× for the whole request, which bills $4.00 / $15.00 rather than the headline pair
• Service tiers — Batch at 50% of Standard, Fast mode at 2×, Ultrafast at 3× through 28 February 2027 and 6× from 1 March 2027, and 30% or 40% off Standard inside the two- and eight-hour background windows
One detail is more useful than the spec sheet: the vendor's own selection guidance changed with the release. The models index now opens by recommending GPT-6.2 Astra for demanding work and GPT-6.2 Sol "to balance intelligence and cost", with GPT-6 Luna for high-volume workloads — and GPT-6.1 Sol has dropped out of that sentence entirely. The older model's page does not carry a pointer to the newer one, while GPT-6 Sol's page still says "See GPT-6.1 Sol for the newer Sol model". OpenAI's own documentation chain now stops one generation short of the model it shipped this week.
Three generations, one rate card
Set the three Sol rate cards side by side and the pattern is unusual for this family. The headline pair never moved: $2.00 in, $10.00 out, on 22 September, 29 September and 17 November. The long-context cliff never moved either — the same 2× input and 1.5× output step above 272,000 tokens, which is where a million-token agent run stops being a cheap one. The cache line moved exactly once, at the first point release, from $0.20 to $0.10 per million cached tokens, taking the cache discount from 90% to 95%, and the third generation kept it there. The reasoning ladder lost one rung at GPT-6.1 Sol and has not regained it.
What that adds up to is a capability release at a flat price rather than a price move — and the contrast with the flagship is the interesting part. GPT-6.2 Astra did cut its numbers, from GPT-6 Astra's $10 / $50 to $8 / $40. OpenAI discounted the expensive model and left the workhorse alone. The Sol tier's position relative to the flagship therefore narrowed without anyone announcing it: at $2 against $8 the workhorse is now a quarter of the flagship's input rate rather than a fifth, and on the tracker's cost-per-task column it is $1.01 against $4.00, which is 25%.
The absolute numbers are worth stating plainly because they are the reason this tier exists. A million input tokens and a million output tokens on GPT-6.2 Sol costs $12 before caching. The same volume on GPT-6 Astra costs $60, and on Claude Opus 5.5 at $4 / $20 it costs $24. The September model already made that argument; the third generation makes it with four more index points behind it and the same invoice.
What the leak file got right, and what it got wrong
This page has now been overtaken three times, and the scorecard is worth keeping because it is the only part of the leak era that still has information in it. The September report named Sol and Luna, put them below Astra, and gave a price of $2.50 in / $15 out that turned out to be 20% too high in the reader's favour. The date was right to the day: Tuesday 22 September, the same Tuesday the skeptical reading of the story had written off. Seven days later the tier moved again with GPT-6.1 Sol at the same price and a halved cache rate.
The 6.2 cycle repeated the pattern with a worse calendar. On 4 November a paid newsletter reported that OpenAI could ship a next-generation Astra, a new Sol and cyber models as soon as Friday 6 November, citing people with knowledge of the models. Nothing shipped that Friday. Through the following ten days the 6.2 identifiers surfaced in Codex billing screenshots and in a user's claim that GPT-6.1 Sol had routed them to an unannounced 6.2 — evidence that something was being served, and explicitly not evidence that anything had been announced. Then, in the middle of that fortnight, a well-followed account argued the model might never ship at all and that OpenAI was building 6.5 instead. It shipped three days later, on a Tuesday, beside a new flagship.
• The names — right again. gpt-6.2-astra and gpt-6.2-sol are the exact strings, not a marketplace label for a reasoning mode
• The date — wrong by eleven days. The reported window was Friday 6 November; the release was Tuesday 17 November
• The price — right by inheritance. The 4 November report carried no figure, and the model that arrived kept the $2 / $10 the tier had held since September. The leak-era guesses of $2.50 / $15, $4 / $30 and $4 / $20 were all wrong and remain wrong
• The family — still incomplete. Reporting described a Sol / Terra / Luna trio arriving together in September; Terra shipped a month late with a different tier position, and there is no 6.2 Luna and no 6.2 Terra in the catalogue today. The 6.2 release is two models, not three
• The shelving story — dead. A model that ships three days after the claim that it may never ship is not a model anybody shelved. The reasoning behind the claim — that no explanation for the delay held up — was a statement about the author's information, not about OpenAI's roadmap, and this is the second time on this page that a confident reading of an absence has been contradicted by a Tuesday

The graphic above is the pre-launch price ladder as the leak reporting drew it, with the Sol row left blank and the vendor's statement at the bottom reading "unconfirmed". That blank has now been filled in three times at the same figure.
The independent reading, re-based
A correction first, because it affects every old number on this page. Artificial Analysis has moved to Intelligence Index v5.0.1, and the earlier versions of this article quoted v4.3.2. The two are not comparable: GPT-6 Sol, published here at 48, now reads 44; GPT-6.1 Sol, published at 52, now reads 55; GPT-6 Astra, published at 53, now reads 57; Claude Opus 5.5, published at 58 and about $5.98 a task, now reads 59 at $10.04. Every figure below was read on the tracker's pages on 17 and 18 November, on the new version, at maximum effort unless stated. That is the tracker's measurement, not the vendor's claim.
• GPT-6.2 Sol (max) — 57, ranked 11th of 91, $1.01 per index task, 7.9 minutes per task, 57.6 seconds to first answer token, 85 tokens per second
• GPT-6.1 Sol (max) — 55, 19th of 91, $1.27 per task, and 23.1 minutes per task, which is the number that makes the newer model feel different in a harness
• GPT-6 Sol (max) — 44, 43rd of 91, $2.47 per task, 7.4 minutes per task, 121 tokens per second
• GPT-6.2 Astra (max) — 62, third of 91, $4.00 per task, 10.4 minutes per task
• GPT-6 Astra (max) — 57, 13th of 91, $6.24 per task
• Claude Opus 5.5 (max) — 59, eighth of 91, $10.04 per task; Claude Fable 5.2 (max) — 65, first, $26.54 per task
• Gemini 4 Carbon (high) — 60, seventh, $5.48 per task; GPT-6 Luna (max) — 28, $0.15 per task, measured inside its own cheaper comparison class
The reading that matters is not the thirty-two-place climb. It is that GPT-6.2 Sol now ties GPT-6 Astra on the index — 57 against 57 — at a quarter of the successor flagship's cost per task and less than half its time to first answer token. The previous version of this page argued that the reason to keep paying for the flagship was the hard tasks where one index point was worth four and a half times the task cost. That argument has been retired by arithmetic rather than by opinion: the one-point gap has closed, and the sixfold gap in cost per task has not. The successor flagship, GPT-6.2 Astra, restores it at 62 for $4.00 a task — a real step, and about four times the Sol bill for five points.
The same run says something less flattering, and it belongs next to the score. On AA-Omniscience, the tracker's knowledge and hallucination evaluation, GPT-6.2 Sol scores 61.6% accuracy with a 55.4% hallucination rate, against GPT-6.1 Sol at 61.8% and 51.9%. The new model is more accurate on coding and agentic work and slightly worse behaved on recall: a 3.5-point rise in the rate at which it invents things, on a build whose aggregate score went up two points. Anyone treating "57 beats 55" as a uniform upgrade is reading one column of a table with nine evaluations behind it.

The screenshot above is the tracker's page for GPT-6 Astra as captured for the September version of this article, when that row read 53 and sat in the top ten of the models the tracker scored. It is here as a record of the re-basing rather than as a current reading: on v5.0.1 that model is at 57 and $6.24 a task, and the 6.2 flagship sits five points above it at $4.00 — cheaper and stronger, which is the reverse of the ordering the capture shows.
The Cyber Index still measures the guardrails
The second tracker column reorders the family in a way that has nothing to do with intelligence, and the 6.2 release does not change the pattern. On the Artificial Analysis Cyber Index at maximum effort, GPT-6.2 Sol reads 37.7 — a point above GPT-6 Sol's 36.8 and four above GPT-6 Astra's 33.4 — while GPT-6 Luna, the model that reads 28 on intelligence, reads 52.7. The row at the top is still GPT-6 Sol (Daybreak Blue, max) at 68.7, the trusted-access configuration that is not purchasable by the token. The highest publicly available row is Kimi K3.1 Pro at 64.1.
• GPT-6 Sol (Daybreak Blue) — 68.7, trusted-access only, not a product anyone can order
• Kimi K3.1 Pro — 64.1; Tencent Hy4 — 58.0; DeepSeek V4.1 Flash 1017 — 56.8; Grok 4.7 — 56.4; MiMo-V2.6-Pro — 56.1
• GPT-6 Luna (max) — 52.7, which is 29 points of index below the Sol tier and 15 points above it on cyber
• GPT-6.2 Sol — 37.7; GPT-6.2 Astra — 37.3; GPT-6 Sol — 36.8; GPT-6 Astra — 33.4; Claude Opus 5.5 — 28.7; Claude Fable 5.1 — 25.0
The index is an equally weighted average of three defensive evaluations — CWE-Bench-AA, adapted from Collinear AI's CWE-bench; DeepsecBench-AA, adapted from Vercel's DeepsecBench as a private held-out set; and CyberGym-E2E-AA, from Berkeley RDI — and the tracker reports refusals separately from the score rather than folding them in. When it opened the leaderboard to models with fewer guardrails on 8 October it published the number that explains this whole column: the public configuration of the September model refused 100% of its CyberGym-E2E tasks, and the Daybreak build refused none, which is the entire 32-point difference between two rows sharing a name, a date and a rate card. That is still the most useful thing on either leaderboard, and it is a statement about policy, not capability. If you are buying for vulnerability triage, the purchasing decision is the refusal rate, and on the 6.2 generation the Sol tier neither improves that materially nor claims to.
The migration details that will actually bite
Three things in the 6.2 documentation change the behaviour of an existing integration, and only one of them is in the changelog. The first is the effort ladder. GPT-6.1 Sol already rejected none and minimal; GPT-6.2 Sol keeps that, and any prompt that leaned on none as a cheap setting will fail validation rather than quietly cost less. The vendor's own guidance is to start at medium for a new integration and to substitute low where none or minimal was used.
The second is prefix handling, and it is new with this generation. The guide states that starting with GPT-6.2 the API drops reasoning items when it detects mid-conversation changes to the conversation history — prefix edits, a model downgrade, an end-user change or an organisation change. When the model loses those items, answer quality can decline and it may spend tokens re-deriving work it had already done. Keep the prefix byte-identical between turns and you keep both the reasoning and the prompt-cache eligibility; edit an instruction mid-run and you pay twice, once in tokens and once in accuracy.
The third is tool calling. Chat Completions is supported, but tools are not: every built-in tool in the list above, and function calling itself in the documented path, goes through the Responses API. Zero Data Retention is supported for approved organisations on eligible endpoints, with store treated as false. Fine-tuning is not supported and neither is any audio or realtime endpoint, so a voice front end still needs a different model behind it.
What has not happened is a retirement. The deprecations page lists no shutdown date for GPT-6 Sol, GPT-6.1 Sol or GPT-6.2 Sol — the Sol identifiers are not on the schedule at all, unlike the GPT-5 generation, which retires on 11 December with gpt-6.1-sol named as the replacement. That is worth reading carefully: the vendor is pointing GPT-5 migrations at the second Sol at the moment the third one became the recommended default, which means the 6.1 identifier is going to keep accumulating integrations after it stopped being the answer to "which Sol".
What to route, and what not to promise
The honest position on availability is the one that costs us the easy sentence. As of 18 November the two 6.2 identifiers are not on OrcaRouter's routes: the catalogue lists openai/gpt-6-sol, openai/gpt-6.1-sol, openai/gpt-6-astra and openai/gpt-6-luna, and nothing named 6.2. We are not going to imply otherwise, because the only thing worse than a slow route is a route a reader believes exists.
What is there is priced the way the vendor prices it. GPT-6.1 Sol is still on our routes at OpenAI's list price, $2.00 / $10.00 with the $4.00 / $15.00 long-context step passed through and no markup layered on top, which is also the rate GPT-6.2 Sol lists — so the day a 6.2 route lands, the arithmetic does not change, only the identifier does. GPT-6 Sol is on the same key at $2 / $10, still callable, still pointed at by OpenAI's own deprecation table for GPT-5.3-codex and GPT-5.1 migrations. GPT-6 Astra sits beside both at $10 / $50 with the $20 / $75 long-context tier, and the full model catalogue carries 215 identifiers behind one key.

The capture above is our own model page for GPT-5.6 Sol — the tier the September piece used as the price comparison, at $4.00 / $20.00, still on sale for twice the Sol rate card and still carrying a promotional price with a stated end date. It is the row that makes the case for the cheap tier easiest to state, and the row that OpenAI's own guidance stopped naming once the Sol tier existed.
That combination is what makes a release week cheaper to survive than it used to be. A same-price successor is the easiest kind of model to test badly: nothing about the bill warns you that the answers changed, so the only defence is to send a slice of real traffic at the new string while the validated model stays primary, and let the error rate and the disagreement rate argue. Automatic failover is what keeps that experiment from being a bet on the production path, and a routing rule is what keeps it from turning into a second integration. When GPT-6.2 Sol does appear on our side, a canary is a model-ID change and a weight, not a contract.
The one thing worth watching is not the index. It is whether the vendor adds a Daybreak configuration for the 6.2 generation — the September model's gated build leads the cyber leaderboard at 68.7 while the 6.2 rows sit near 37, and if that gap closes on the newer weights the two leaderboards will finally agree about which model is the strongest OpenAI ships. Until then, the summary of this page is short. OpenAI put a third Sol on sale at the same $2 / $10, thirteen index points above the model this article was originally written about, tied with its own previous flagship and six times cheaper per task than it. The leak file got the name, the tier and the price right, and the date wrong by eleven days. The model that shipped is the best argument yet that in this family the version number is a marketing surface and the rate card is the contract.
Compared in this article4
Detected from this article · Benchmarks: Artificial Analysis · updated daily
