GPT‑6.1 Sol Brings Near-Astra Intelligence at a Much Lower Price
OpenAI positions GPT‑6.1 Sol as a lower-cost model approaching Astra on selected coding, computer-use and professional evaluations. The important question is what that means in real work.
For much of the artificial intelligence race, attention has centered on the most powerful model. Each generation has been judged by its ability to solve harder problems, write better code or reach higher benchmark scores. GPT‑6.1 Sol introduces a different question: what if the most important advance is not only making AI more intelligent, but making near-frontier capability affordable enough for everyday work?
OpenAI introduced GPT‑6.1 Sol on September 29, 2026 as an upgrade to GPT‑6 Sol. The company says the model approaches GPT‑6 Astra in agentic coding, computer use and professional work while costing substantially less.
That comparison requires care. “Near-Astra intelligence” is OpenAI's description, based on evaluations selected and reported by the company. It does not mean Sol equals Astra across every task. Astra remains OpenAI's most capable model for problems where obtaining the best possible result justifies a higher cost.
The lower price may nevertheless matter more than a small improvement on a benchmark. A slightly less capable model that is significantly cheaper can run more tasks, serve more users and allow agents to work for longer.
What Exactly Is GPT‑6.1 Sol?
GPT‑6.1 Sol belongs to the GPT‑6 family, which offers different balances of capability, speed and cost:
- GPT‑6 Astra: OpenAI's most advanced model for the hardest work;
- GPT‑6.1 Sol: a balance of near-Astra capability and substantially lower cost;
- GPT‑6 Luna: a fast, economical option for everyday tasks and high-volume applications.
The “6.1” label represents a meaningful upgrade to the original Sol. OpenAI reports improvements in coding, complex documents, computer use, scientific work, factual accuracy and multi-step business workflows.
This position may be more important than it first appears. Many organizations do not need the strongest available model for every interaction. They need a system capable enough to finish work reliably while keeping costs predictable.
How Much Does GPT‑6.1 Sol Cost?
OpenAI's standard API prices are listed per one million tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT‑6 Astra | US$10 | US$1 | US$50 |
| GPT‑6.1 Sol | US$2 | US$0.10 | US$10 |
| GPT‑6 Luna | US$0.10 | US$0.01 | US$0.50 |
At the announced prices, Sol costs one-fifth as much as Astra for both standard input and output. The difference is even larger for reused context: Sol's cached input costs US$0.10 per million tokens, 90 percent below Astra's cached-input price.
Caching is especially important for agents and long-running conversations. When an application reuses instructions, documents or context across requests, it does not need to pay the full processing price every time.
This does not guarantee that every completed task will be exactly five times cheaper. A model may consume more tokens, repeat steps or need another attempt. Real cost depends on reasoning effort, tools, duration and success rate.
Coding and AI Agents
OpenAI highlights GPT‑6.1 Sol on long-horizon software-engineering work. On DeepSWE 1.1, an evaluation where agents operate in real codebases, the company says Sol matched Astra's result at approximately one-fifth of the cost in the compared configuration.
Agentic coding involves more than generating a short code sample. The system must explore files, understand dependencies, modify components, run tests and correct failures.
One benchmark cannot represent every software project. Real repositories contain incomplete documentation, historical decisions, conflicting requirements and unreliable environments. Quality must also be measured through passed tests, regressions, human review time and maintainability.
Complex Documents and Professional Work
On GDP.pdf, which tests professional questions based on documents containing tables, charts, diagrams and fine-print details, OpenAI says GPT‑6.1 Sol approached Astra and outperformed Claude Opus 5.5 under the reported conditions.
This capability could be useful in finance, healthcare, law, operations and research. However, reading a PDF should not be confused with professional authority. A model may interpret most of a document correctly and still fail on the single detail that determines the decision.
For sensitive work, users should require references, inspect the original page and retain qualified human review.
Computer Use
GPT‑6.1 Sol also improves on tasks performed inside applications. On OSWorld 2.0, OpenAI reports a seven-percentage-point improvement over GPT‑6 Sol at maximum reasoning effort. Its score came within 2.1 points of Astra at approximately one-seventh of the cost per task in the comparison.
Computer use is central to the future of agents. A model that can operate applications may fill forms, organize files, consult platforms and execute multi-step processes.
It also increases risk. An error can move beyond incorrect text and alter data, send a message or perform an unauthorized action. Permissions, approval steps, logs and access boundaries therefore matter as much as model intelligence.
Business Automation
AutomationBench evaluates end-to-end workflows using 47 tools across sales, marketing, operations, support, finance and human resources. OpenAI reports that GPT‑6.1 Sol scored 2.2 points above Opus 5.5 at medium reasoning effort and 4.8 points above GPT‑6 Sol.
The result suggests that lower-cost models are becoming capable of coordinating complete sequences rather than merely answering isolated questions.
Enterprise adoption still depends on much more than a score. Integration, privacy, uptime, auditing, exception handling and responsibility when something goes wrong all affect the real outcome.
Scientific Research
On Terminal-Bench Science 0.1, OpenAI says GPT‑6.1 Sol more than doubled GPT‑6 Sol's score at maximum reasoning effort. The reported average cost was US5.47pertask, comparedwithUS23.80 for Astra. Astra retained the highest score at 68.1 percent.
This difference illustrates the intended roles. Astra may remain preferable when the strongest result justifies the cost. Sol may be more practical for running many preliminary analyses, simulations or experiments.
Neither model replaces scientific validation. Code-generated results must be reproduced, assumptions examined and conclusions tested against evidence.
Does GPT‑6.1 Sol Make Fewer Factual Errors?
OpenAI evaluated difficult conversations in which users had flagged a factual error from an earlier model. At low reasoning effort, the proportion of answers containing at least one factual error fell from 11.4 percent with GPT‑6 Sol to 7.7 percent with GPT‑6.1 Sol, an approximate reduction of 32 percent.
That is meaningful, but it does not establish a universal 7.7 percent error rate. OpenAI explains that the dataset was selected to elicit failures and is not representative of normal usage.
The responsible interpretation is straightforward: the new Sol performed better on this test, but it can still produce incorrect information.
Safety and Respecting Instructions
OpenAI reports improvements in transparency when a search tool is unavailable, compliance with explicit restrictions and avoidance of unauthorized outcomes.
In a difficult broken-search evaluation, GPT‑6.1 Sol failed to disclose the problem in 2.1 percent of cases, compared with 4.9 percent for GPT‑6 Sol and 1.5 percent for Astra. OpenAI also reports observing no attempts to bypass an automated safety reviewer.
These results are encouraging, but they do not remove the need for external safeguards. An agent should receive only the permissions it needs, request approval before consequential actions and keep records that make its behavior auditable.
Where Is GPT‑6.1 Sol Available?
At launch, GPT‑6.1 Sol became available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. OpenAI stated that it was not yet available in the conventional Chat experience.
Developers can access it through the API using the model identifier gpt-6.1-sol. OpenAI also announced a Sol Ultrafast option for Codex with token generation up to eight times faster than standard speed.
Availability may vary by product, plan, region and workspace administration. A model's presence in the GPT family does not automatically mean it appears in every interface.
Who Benefits Most?
GPT‑6.1 Sol appears particularly relevant to:
- developers running agents through long tasks;
- organizations processing large volumes of documents;
- teams automating workflows across multiple tools;
- researchers conducting many preliminary analyses;
- products that reuse cached context;
- businesses that find Astra too expensive for constant use.
Luna may remain more economical for simple, high-volume operations. Astra remains OpenAI's recommended choice where the strongest possible result matters most.
What “Near-Astra” Does Not Mean
The phrase does not mean general equality. Similar scores on one benchmark can conceal important differences elsewhere. A small percentage gap may also produce many additional failures when a system operates millions of times.
It does not mean every customer will achieve the advertised cost relationship. Attempts, context length, tools and reasoning settings all change the final bill.
Sol's real value will be determined through independent use: how many tasks it completes correctly, how much human time it saves and the total cost of each approved result.
NTS View
In practical testing by the NextTechSearch editorial team, GPT‑6.1 Sol does not feel like a complete break from the previous generation. Its strongest value appears when requests are written clearly, directly and with the necessary context. Good prompting remains essential; a more capable model cannot reliably compensate for unclear goals or missing information.
Our current editorial assessment is 8.5/10. Sol provides an excellent balance of capability, speed and cost, but Astra remains clearly ahead when the most demanding work requires maximum intelligence and deeper verification. This score is an editorial judgment based on practical use, not a laboratory benchmark.
GPT‑6.1 Sol may represent a more important shift than it first appears. The frontier advances not only when the strongest model becomes smarter, but also when previously expensive capability becomes affordable enough for daily use.
If near-Astra performance can be accessed at a much lower price, companies no longer need to reserve advanced intelligence for a handful of critical tasks. They can apply it more broadly to coding, documents, research and operations.
Lower prices can also allow errors to multiply faster. Automating thousands of decisions requires better supervision, not less. The real victory is not performing more actions; it is completing more correct, verifiable and authorized work for every euro spent.