
GPT-6.2 Sol vs GPT-6 Sol: The Last Sol That Lets You Turn Reasoning Off
- 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
Three Sol models have now shipped at the same sticker price, and the oldest of them still does something the two newer ones refuse to do. GPT-6 Sol arrived on 22 September 2026 at $2 per million input tokens and $10 per million output, and it accepts none as a reasoning effort. GPT-6.2 Sol, released on 17 November 2026, charges the same $2 and $10 and rejects none outright — as does GPT-6.1 Sol, which landed seven days after GPT-6 Sol and quietly took the crown from it at an unchanged price. So anyone still running GPT-6 Sol in production is not running an obsolete model by accident. In several configurations, it is the only model in the family that will accept the request.
Same meter, one dial that actually moved
The launch posts make this look like a non-event: OpenAI described GPT-6.2 Sol as carrying forward GPT-6.1 Sol's rates, and GPT-6.1 Sol never raised what GPT-6 Sol charged. Read across all three rate cards, though, and exactly one line item has ever changed. Cached input.
• Input — $2.00 per million tokens on GPT-6 Sol, GPT-6.1 Sol and GPT-6.2 Sol. Unchanged across three generations.
• Output — $10.00 on all three. Unchanged.
• Cache writes — $2.50 per million on all three for 30 minutes of retention. GPT-6.2 Sol additionally documents a 12-hour cache write at $5 per million for short-context requests at Standard pricing; the GPT-6 Sol page describes no such option.
• Cached input — $0.20 on GPT-6 Sol, where it is billed as 10% of the uncached rate. $0.10 on GPT-6.2 Sol, where it is billed as 5%. The cache read got 50% cheaper while the headline number stayed put.
• Context — 1,050,000 tokens, 922,000 max input, 128,000 max output, text and image in, text out. Identical on all three.
• Knowledge cutoff — 20 April 2026 on GPT-6 Sol, 30 April on GPT-6.1 Sol, 1 June on GPT-6.2 Sol. Six weeks of extra world knowledge between the oldest and the newest.
• Reasoning effort — GPT-6 Sol accepts none, low, medium, high, xhigh and max. GPT-6.2 Sol accepts the same list minus none.
• Data residency — GPT-6 Sol supports EU residency with Standard, Flex and Batch, but not Fast. GPT-6.2 Sol supports US and EU residency with Standard, Fast and Batch. On the newer model, the fast tier is eligible for EU processing; on the older one it is not.
• Cloud — GPT-6 Sol is listed on Amazon Bedrock as openai.gpt-6-sol. GPT-6.2 Sol has no cloud-provider listing on its model page yet.
Cache-hit workloads are where that 50% shows up. A retrieval pipeline that sends 800,000 cached tokens and 20,000 fresh ones per request is paying $0.16 of cache read on GPT-6 Sol and $0.08 on GPT-6.2 Sol, before either model reasons about anything. Multiply by request volume and the unchanged headline price starts to look like the least interesting number in the comparison.
A generation that lasted seven days
It is worth being precise about how GPT-6 Sol became the old model, because it happened faster than any deprecation cycle would suggest. OpenAI announced GPT-6 Sol and GPT-6 Luna on 22 September 2026. On 29 September it announced GPT-6.1 Sol, described as "near-Astra intelligence for coding, computer use, and professional work at one-fifth of Astra's standard API input and output token prices" — which is also a fair description of what GPT-6 Sol was supposed to be. One week, same price, better model.
Nothing about GPT-6 Sol has been withdrawn. OpenAI's lifecycle table still carries it as an active model, and its documentation page simply points readers at the newer Sol. But Artificial Analysis has marked it deprecated and now benchmarks it every 18 hours rather than every 6. Its banner recommends GPT-6.1 Sol rather than GPT-6.2 Sol — a reminder that these recommendation banners point one hop forward, not to the current best model, and are not a substitute for reading the board.
Forty-four against fifty-seven, and what that gap is made of
On Artificial Analysis's Intelligence Index v5.0.1, both models measured at max effort, the two generations are further apart than the 6.1-to-6.2 step:
• GPT-6.2 Sol — index 57, eleventh of 87 models, $1.01 weighted cost per index task, 60.9M output tokens for the full suite, time per task not yet published.
• GPT-6 Sol — index 44, fortieth of 87, $2.47 per task, 65.1M output tokens, 7.4 minutes per task at roughly 121 output tokens per second with 37.34 seconds to the first answer token.
Thirteen index points, a 59% lower cost per finished task, and 6.4% fewer output tokens. The index gap is the headline; the cost gap is the one that changes budgets, because $2.47 against $1.01 on identical per-token rates means the old model is burning roughly two and a half times as many billable tokens to land a worse answer. That is the failure mode of model generation comparison in one line — the price list says the two models are the same product, and the invoice says they are not.

The reverse comparison is just as instructive. GPT-6 Sol is the faster of the two on the board's own measurement — 7.4 minutes per index task against GPT-6.1 Sol's 23.1 minutes — and it is the only one of the three with a published latency figure at all. GPT-6.2 Sol's performance panel reads N/A. So the scoreboard does not currently let you claim the new model is faster; it only lets you claim it is smarter and cheaper per task. If your application is interactive, that unanswered number should hold the migration until it is answered, and a sample of your own traffic is the only substitute.

The switch that turns the meter down instead of up
Reasoning effort none is the most consequential non-obvious difference between these two models, and it gets almost no attention because every comparison is run at max. On GPT-6 Sol it is available, and on Chat Completions it is the only setting under which function calling works at all — OpenAI's page for the model says Chat Completions "supports function calling only with reasoning_effort set to none." On GPT-6.2 Sol the parameter is rejected, and tool calling requires the Responses API regardless.
Why that matters commercially: a reasoning model at effort none is a different cost object from the same model at max. The index numbers above are all max-effort — Artificial Analysis publishes its scores as the highest-effort configuration. A classification, extraction or routing step that needs a chat-shaped call with tools and no deliberation is exactly the workload where GPT-6 Sol at none can cost a fraction of what the index comparison implies, because most of the token burn on that board is reasoning tokens you would not be buying.
None of that makes GPT-6 Sol the better model. It makes it a model with a different shape, and shape is what decides whether a migration is a swap or a rewrite. If your integration depends on none, the move to GPT-6.2 Sol is not a cost optimisation — it is a redesign of how the request is built, plus a decision about whether you want the reasoning tokens you were previously declining to pay for.
What a migration from GPT-6 Sol actually involves
Going from GPT-6 Sol straight to GPT-6.2 Sol skips GPT-6.1 Sol, and the migration notes are cumulative rather than incremental. Four things change:
The reasoning-state behaviour is new in 6.2 and it is silent. From GPT-6.2 onward, the API drops reasoning items when it detects mid-conversation changes to the history — prefix edits, model downgrade, an end-user change, an organisation change. OpenAI warns that when the model loses those items, "answer quality can decline and the model may use more tokens to repeat earlier work." For an application that rewrites its system prompt between turns, this is a cost regression with no error attached to it. Keep the prefix stable and append new instructions instead.
Sampling parameters are gone. Omit temperature, top_p, logprobs and top_logprobs from Chat Completions requests; for Responses, omit temperature, top_p and top_logprobs and drop message.output_text.logprobs from include. This is the change most likely to fail loudly in a smoke test, which is a mercy.
Prompt caching is renamed and re-billed. If you are coming from GPT-5.5 or earlier in the same pass, prompt_cache_retention becomes prompt_cache_options.ttl set to "30m". Cache writes are a billed line item at $2.50 per million tokens for 30 minutes of retention, and $5 for 12 hours. A pipeline that assumed caching was free on the write side will see a new line on the invoice.
Effort levels have to be re-chosen. none is rejected, so anything that passed it — including every Chat Completions tool-calling path on GPT-6 Sol — needs a new effort level. OpenAI's guidance for new integrations is to start at medium and drop to low where an older integration used none or minimal.
Two of the three are callable on one key today
The awkward part of this comparison is availability, and it splits three ways. GPT-6 Sol and GPT-6.1 Sol are both in OrcaRouter's catalogue at pass-through pricing — $2 in, $0.20 cached and $10 out for GPT-6 Sol, the vendor's own list with 0% markup on top. That makes the honest first experiment cheap: run your own traffic against GPT-6 Sol and GPT-6.1 Sol on one key, no second contract, and find out what the thirteen-index-point gap is worth on your tasks rather than on a benchmark suite. Where the two differ in price, we differ too, because a vendor price change passes through here the same day it is announced rather than whenever a markup table is rebuilt.
GPT-6.2 Sol is not in the catalogue yet. It is on OpenAI's API, in ChatGPT Work and Codex since launch day, and that is the honest list — any article telling you otherwise, including one of ours, would be wrong. The routing pattern that covers the gap is to keep the two catalogue models on one policy, point the new model's traffic at the vendor API, and let failover handle the case where the newer model refuses a request the older one would answer. That last part is not hypothetical: OpenAI shipped both 6.2 models with tighter safeguards for higher-risk cybersecurity assistance and states in the launch post that the protections "restrict some legitimate vulnerability research and security testing workflows," with Daybreak access for 6.2 still to come.

Who should move, and who is already correct
Move to GPT-6.2 Sol if your workload is batch or background, if you were already paying for reasoning tokens, and if you can afford to re-measure latency on your own traffic rather than trusting a board that has not published the figure. You gain thirteen index points over GPT-6 Sol, a 59% lower cost per finished task on the same token rates, a halved cache-read price, EU residency on the fast tier, and a 12-hour cache option.
Stay on GPT-6 Sol if the request shape depends on effort none — particularly Chat Completions with function calling, which the newer models do not support at all. That is not nostalgia; it is the one capability the successor removed, and rebuilding around it is a project with its own budget.
Consider GPT-6.1 Sol as the middle rung if you want most of the intelligence gain without the 6.2 migration. It scores 55 against GPT-6 Sol's 44, costs $1.27 per index task against $2.47, keeps the 5% cached-input rate, and its launch material was explicitly about doing near-Astra work at a fraction of Astra's price. What it does not give you is speed: 23.1 minutes per index task is the slowest figure in the tier, against 7.4 for the model it replaced. If latency is the constraint you actually feel, GPT-6.1 Sol is the wrong upgrade and GPT-6 Sol is doing more for you than its index score suggests.
The useful way to hold all three in your head is that the Sol line has been optimised for token efficiency in every release, and the rate card has been held flat while that happened. GPT-6 Sol spends 65.1M output tokens to score 44. GPT-6.1 Sol spends 63.7M to score 55. GPT-6.2 Sol spends 60.9M to score 57. Nothing in that progression was bought with a price increase — which is precisely why a team that compares models by dollars per million tokens, rather than dollars per finished task, keeps concluding that nothing has changed since September.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
