GPT-6.1 Sol OpenAI AI models

GPT-6.1 Sol: OpenAI's Near-Astra Model at a Fifth of the Price

OpenAI's GPT-6.1 Sol costs $2/$10 per million tokens with a 1.05M-token context window, lands one point below GPT-6 Astra on independent benchmarks, and cuts cost per task to under a quarter.

m
· 6 min read · — views · — reads
GPT-6.1 Sol: OpenAI's Near-Astra Model at a Fifth of the Price
Table of contents

GPT-6.1 Sol: OpenAI's Near-Astra Model at a Fifth of the Price

OpenAI used its DevDay 2026 stage on September 29 to launch GPT-6.1 Sol, a model the company sums up in one line: near-Astra intelligence for a fifth of the price. Priced at $2 per million input tokens and $10 per million output tokens, with a context window that stretches past one million tokens, it lands a single point below the flagship GPT-6 Astra on independent intelligence benchmarks — while a typical task costs roughly a quarter of what Astra charges.

The timing is as aggressive as the pricing. GPT-6.1 Sol replaces GPT-6 Sol only seven days after that model shipped, the fastest swap OpenAI has ever executed. The same stage introduced OpenAI Dots, always-on agents we cover separately. The message to developers is blunt: stop paying flagship premiums for everyday agentic work, because the gap just collapsed to a rounding error.

OpenAI DevDay 2026 keynote where the model was announced

Source: OpenAI official YouTube channel

Full specifications: what $2 per million tokens buys

The official model page on OpenAI's developer documentation lays out the numbers precisely. The context window is 1,050,000 tokens — up to 922,000 for input and 128,000 for output — among the widest commercially available today. The knowledge cutoff runs to April 30, 2026, recent enough for most ongoing workloads.

The detailed price sheet:

  • Input: $2 per million tokens (nominally identical to GPT-6 Sol).
  • Output: $10 per million tokens.
  • Cached input reads: $0.10 — a 95% discount, up from 90%, the deepest cache price OpenAI has ever published.
  • Cache writes: $2.50.
  • Long prompts: any request above 272,000 input tokens is billed at 2× for input and cache, and 1.5× for output.
  • Batch and Flex: 50% off. Fast mode doubles the price. Regional processing adds 10%.

Reasoning effort spans five levels from low to max, with medium as the default. The model accepts text and image input and returns text only, does not support fine-tuning, and runs with US and EU data residency. Rate limits climb to 15,000 requests per minute and 40 million tokens per minute at Tier 5 — enterprise-grade ceilings, not individual-developer numbers.

The official GPT-6.1 Sol safety evaluation page from OpenAI

Source: OpenAI Deployment Safety Hub

Quick comparison: Sol versus Astra versus the previous generation

Metric GPT-6.1 Sol GPT-6 Astra GPT-6 Sol
Input price (per 1M tokens) $2 Multiple times higher $2
Output price (per 1M tokens) $10 Multiple times higher $10
Cost per task (Artificial Analysis) $0.72 $3.26 $1.05
Intelligence Index 1 point below Astra Top reference 4 points lower
Terminal-Bench 4.0 +12 points vs Sol Highest Reference
Context window 1.05M tokens Narrower Narrower

The table's bottom line: for general intelligence work, Astra now costs more than four times as much as Sol for one extra benchmark point. Independent testing by Artificial Analysis found something even sharper — at xhigh reasoning effort, GPT-6.1 Sol scored one point above Astra while consuming under 15% of its cost. Even though the new model emits 10–30% more output tokens to finish the same job, the final arithmetic stays firmly on Sol's side.

What this means for you

If you build products on the API anywhere from Lagos to Lahore, this release changes your unit economics directly. The 95% cache discount matters most for agentic applications that re-send context on every step — cached reads are typically the single largest line in a real agent's bill, and cutting that line by a further five percentage points compounds across thousands of calls. A $0.72 cost per task puts daily cognitive automation within reach of startups that previously postponed it.

If you are a ChatGPT Pro subscriber, the news cuts both ways. OpenAI's official subscriber email describes the model as near-Astra capability "in a model you can use every day," but the same email cuts the GPT-6 Pro chat allowance from 200 to 100 messages per week starting October 30, 2026. Smarter model, smaller quota — a genuinely good deal for heavy users and a mildly annoying one at the same time. Our earlier coverage of OpenAI pausing new Pro sign-ups explains the capacity pressure behind moves like this.

The safety picture: what to know before you deploy

The official system card addendum, published September 29, 2026 on OpenAI's Deployment Safety Hub, does not soft-pedal the classification. GPT-6.1 Sol's cybersecurity capabilities are rated Critical — meaning it meets the company's threshold for developing functional zero-day exploits end to end. The biological and chemical rating is High without crossing Critical, and self-improvement attempts sit below High.

Is that a reason to stay away? The official framing is: Astra-level capabilities shipped with Astra's own safeguards stack. The model beats GPT-6 Sol in five of eight production safety categories, recorded zero attempts to bypass internal monitors (versus 7.3% for GPT-5.6 Sol), and misrepresents its coding work less often. On HealthBench it lands within half a point of Astra. If you have followed the frontier-safety debate since Astra arrived, our piece on GPT-6 Astra's critical cyber capabilities provides the full backdrop.

Limitations you will not find in the announcement

Weigh these before a full migration. There is no audio or video output, and no fine-tuning on your data. The "none" and "minimal" reasoning settings are gone — every request runs through some reasoning path. Requests above 272,000 tokens double your input cost, so the advertised savings are conditional on your context length. Independent measurements also show 10–30% higher output token consumption than the previous generation, so calculate savings on full task cost rather than the headline per-token price.

One more honest note: even OpenAI's own documentation concedes that its ExploitBench results may be artificially inflated by contamination from historically public vulnerabilities. That level of candor is useful — build your decisions on relative numbers, not absolute ones.

Should you switch now?

Coming from GPT-6 Sol, yes without hesitation: 31% cheaper tasks and four more intelligence points at the same nominal price. Coming from GPT-5.6 Sol, the jump is bigger still — a 64% cost-per-task reduction. Only if your workload demands the absolute intelligence ceiling regardless of price does Astra still earn its premium, and even that gap is now one point.

Migration itself is trivial: the model id is gpt-6.1-sol across Responses, Chat Completions, and Batch. For the wider arc of how fast OpenAI's pricing floor has moved, our launch coverage of GPT-6 Sol and Luna shows how much changed in a matter of weeks.

And while you benchmark the new model against your stack, the same value logic runs through Truescho's free student platform — maximum practical output at minimum cost, from fully-funded scholarships to a smart AI study assistant.

Frequently asked questions

What is GPT-6.1 Sol and how is it different from GPT-6 Sol?

It is OpenAI's new model announced September 29, 2026 at DevDay. Independent testing puts it four intelligence points above GPT-6 Sol with a 31% lower cost per task ($0.72 versus $1.05) at the same nominal token prices, plus a deeper 95% input-cache discount. It replaces GPT-6 Sol just seven days after that model launched.

How much does GPT-6.1 Sol cost?

$2 per million input tokens and $10 per million output tokens. Cached reads cost $0.10 and cache writes $2.50. Batch and Flex modes are half price, while requests above 272,000 input tokens are billed at double the input rate and 1.5× the output rate.

Is GPT-6.1 Sol better than GPT-6 Astra?

On the overall Intelligence Index it sits one point below Astra while a typical task costs less than a quarter as much. At xhigh reasoning effort it even scored one point above Astra in Artificial Analysis testing. Astra retains the edge on the most open-ended, complex work, but the gap has narrowed to a point.

Is GPT-6.1 Sol available in ChatGPT or only via the API?

It is available through the API as gpt-6.1-sol on Responses, Chat Completions, and Batch, with US and EU data residency. For ChatGPT Pro subscribers, OpenAI's email positions it as the daily-use near-flagship model, while cutting the GPT-6 chat allowance to 100 weekly messages from October 30, 2026.

What is the context window of GPT-6.1 Sol?

1,050,000 tokens in total: up to 922,000 input tokens and 128,000 output tokens, with reasoning tokens supported. Note that prompts exceeding 272,000 input tokens are billed at 2× the input rate and 1.5× the output rate.

Sources