GPT-6.1 Sol: What OpenAI’s Near-Astra AI Model Changes
GPT-6.1 Sol brings near-Astra performance to coding, computer use and professional work at lower cost. Here is what the benchmarks, pricing and early independent tests show.
On this page
OpenAI released GPT-6.1 Sol on September 29, positioning it as a cheaper model that comes close to GPT-6 Astra on coding, computer use and professional work. The interesting part is not simply another model number: GPT-6.1 Sol keeps the same $2 input and $10 output price as GPT-6 Sol while OpenAI reports substantially higher scores on several demanding tasks, including a seven-point gain on OSWorld 2.0 and a 6.4-point improvement over GPT-6 Sol on DeepSWE v1.1.
GPT-6.1 Sol is aimed at work that needs reasoning without Astra pricing
OpenAI built GPT-6.1 Sol for complex coding, computer interaction and professional workflows rather than positioning it as a general replacement for every model in the GPT-6 family. The model has a 1,050,000-token context window and supports reasoning efforts from low through maximum, giving developers a way to trade computation time and cost against task difficulty. It is also available through the Responses API, where it can use tools such as web search, file search, code execution, computer use and multi-agent workflows.
For developers building an AI agent, that combination matters because a long-running workflow can generate a large amount of context before it produces a useful result. OpenAI says GPT-6.1 Sol also benefits from improved caching, with cached input priced at $0.10 per million tokens. That is half the cached-input price of GPT-6 Sol and only a small fraction of the standard input rate, although actual application costs still depend on how much context a workload sends and reuses.
The price did not fall from GPT-6 Sol, but the amount of work per dollar can
There is an easy detail to miss in the launch. GPT-6.1 Sol did not receive a lower standard input or output price than GPT-6 Sol: both are listed at $2 per million input tokens and $10 per million output tokens. The meaningful cost change comes from how much work the newer model needs to reach a given result, plus the lower cached-input rate. OpenAI reports GPT-6.1 Sol at roughly one-fifth of GPT-6 Astra's standard input and output token prices, so the comparison is mainly about moving some workloads down from the flagship tier rather than making Sol itself cheaper than its immediate predecessor.
| Model | Input per 1M tokens | Cached input | Output per 1M tokens |
|---|---|---|---|
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
Those are standard API rates for prompts within the documented short-context pricing range. OpenAI also applies different rates to long-context requests and processing modes, so the headline $2/$10 figure should not be treated as a universal per-request price. For applications that repeatedly send the same large instructions, documents or conversation state, the cached-input reduction is more directly relevant than the unchanged output price.
The coding results show where the upgrade is supposed to matter
DeepSWE v1.1 is designed around long-horizon software-engineering tasks in real codebases, rather than short programming questions. OpenAI reports that GPT-6.1 Sol matches GPT-6 Astra on this evaluation while costing about one-fifth as much per task, and says it beats GPT-6 Sol's best score by 6.4 percentage points at a lower reasoning effort and cost. That is a more useful distinction than simply saying the new model is “smarter”: the claimed benefit is that a developer can run a capable coding agent for longer before model cost becomes the dominant constraint.
There is a reason to treat that result as an OpenAI measurement rather than a universal performance guarantee. The company says its own evaluations were performed in its research environment or through its API, while competitor results were taken from publicly available reports. Different prompts, tools, system instructions and inference settings can change the outcome, so a benchmark score describes a particular test configuration rather than every coding workload.
An independent September 30 benchmark published on GitHub gives a useful second perspective, although it has much smaller sample sizes. The author reports 3,527 API calls across seven tracks in a ChatGPT Pro Codex environment and found GPT-6.1 Sol reached full marks on the tested reasoning, coding, vision, instruction-following, function-calling and JSON-output tracks. The same test measured visible streaming at about 31 tokens per second, slower than GPT-6 Sol and GPT-6 Luna in that particular environment, and reported that most accuracy tracks hit a ceiling where differences between the top models were difficult to separate.
Computer use is another major part of the upgrade
GPT-6.1 Sol is not limited to producing text or source code. Computer-use models operate software by interacting with interfaces, which makes them useful for workflows where an agent must open applications, inspect information, make changes and continue through several steps. OpenAI reports that GPT-6.1 Sol scored seven percentage points higher than GPT-6 Sol on the offline portion of OSWorld 2.0 at maximum reasoning effort, while costing less than half as much per task in its evaluation. It came within 2.1 percentage points of GPT-6 Astra while costing roughly one-seventh as much per task in that test.
That result also explains why the model is more relevant to agent developers than a conventional chatbot upgrade would be. A computer-use agent can spend many model calls on a single task, so a small improvement in the number of successful steps can matter more than a small difference in the quality of one answer. At the same time, computer-use performance depends on the surrounding tools and environment, meaning a benchmark improvement does not automatically translate into the same improvement inside a particular application.
OpenAI also reports a substantial factuality improvement
GPT-6.1 Sol's changes are not limited to task completion. OpenAI tested factuality using de-identified conversations in which users had previously flagged an answer from an earlier model as containing a factual error. At low reasoning effort, the reported error rate fell from 11.4% for GPT-6 Sol to 7.7% for GPT-6.1 Sol, a reduction of about 32%. OpenAI says the test deliberately selects difficult, error-inducing conversations and therefore should not be interpreted as the factual-error rate users will experience in ordinary conversations.
That qualification is significant. A benchmark designed around known mistakes is useful for measuring whether a model has improved on difficult failure cases, but it does not tell you that a model will make an error in exactly 7.7% of normal requests. For production software, the practical question is whether the model's lower measured error rate survives on the specific documents, databases, codebases and workflows that the application actually handles.
The model is powerful enough that OpenAI's safety classification matters
GPT-6.1 Sol is also covered by a more detailed safety evaluation than a simple product specification would suggest. OpenAI's deployment safety addendum classifies the model as Critical for cybersecurity capability and High for biological and chemical capability under its Preparedness Framework. The company says GPT-6.1 Sol therefore uses the same safeguards stack as GPT-6 Astra.
OpenAI also reports lower failure rates than GPT-6 Sol in several alignment tests, including tests involving broken search tools, explicit restrictions and unauthorized outcomes during agentic tasks. These are deliberately difficult evaluations, not measurements of ordinary user behavior. The system card makes the same distinction, so the results are best read as evidence about specific stress tests rather than a claim that the model cannot behave incorrectly in real applications.
The biggest limitation is that near-Astra does not mean equal to Astra
OpenAI's own results show why “near-Astra” needs to be read carefully. On Terminal-Bench Science 0.1, GPT-6.1 Sol reached a much lower score than Astra while costing considerably less per task; OpenAI explicitly says Astra remains the model for its most difficult scientific research workloads. That is consistent with the way OpenAI has positioned the two models: Sol is intended to bring much of the useful capability into a cheaper operating tier, while Astra remains the top-end option for workloads where the extra capability justifies the additional cost.
Independent testing also adds an important wrinkle: accuracy and speed do not necessarily move together. The September 30 benchmark found GPT-6.1 Sol highly accurate on its selected tasks but measured it as slower than several other GPT-6 models in visible token generation. The benchmark's author also cautioned that the sample sizes were small and that many tracks reached ceiling performance, so those numbers should be treated as an early data point rather than a definitive ranking.
For developers, the useful change is a different cost-performance boundary
The practical effect of GPT-6.1 Sol is therefore less dramatic than a new flagship model number suggests, but more consequential for developers running agents at scale. The model keeps GPT-6 Sol's standard token rates, lowers cached-input cost, improves on several demanding evaluations and is designed to handle coding, documents and computer-use workflows that previously pushed applications toward a more expensive model. OpenAI also supports a 1.05-million-token context window and up to 128,000 output tokens, although large-context requests have separate pricing rules.
There is already evidence outside OpenAI that the model behaves differently in real environments, but it is still early. A September 30 community benchmark measured thousands of calls, while smaller hands-on tests and developer reports are beginning to expose issues such as latency that official task-cost charts do not capture. For teams deciding whether to migrate, the sensible test is therefore not a single benchmark number: run the model against the actual codebase, tool chain and long-running workflows that determine your bill and failure rate.
GPT-6.1 Sol is now available in OpenAI's API and to eligible users in ChatGPT Work and Codex, while OpenAI says it is not yet available in regular ChatGPT. The company also says an Ultrafast mode is coming to Codex, with up to eight times faster token generation than its standard speed. That makes the next comparison worth watching less about raw model scores and more about how much useful work developers can complete per dollar and per minute once the model is running at production scale.
Share this post



