# GPT-6.1 Sol vs Astra: move the testable work, keep the hard calls

> Sol matches Astra in OpenAI's DeepSWE coding test at far lower task cost. Independent results are close but effort-dependent; documents, computer use and hard research set different limits.

- Published: Oct 2, 2026
- Author: Linden Kern, Chief Scientist
- Tags: models, coding-agents, cost
- Canonical: https://pirouter.ai/blog/gpt-6-1-sol-astra-replacement

---
**GPT-6.1 Sol matched Astra in OpenAI's DeepSWE coding evaluation at roughly one-fifth the cost per task. That makes it a serious candidate for work whose result you can test, not a reason to route every difficult decision away from Astra.** The launch's broader “one-fifth” headline is about **standard input and output token prices**; it is not the price of finishing every job ([OpenAI's September 29 announcement](https://openai.com/index/introducing-gpt-6-1-sol/)).

Consider a hypothetical team with three things due today: a small software fix with regression tests, an answer drawn from a densely footnoted PDF, and a scientific analysis whose error would be hard to spot. The cheap first attempt matters in all three. The cost of discovering a bad result matters differently in each. That is the real Sol-versus-Astra choice, and a combined model score cannot make it for you.

The official release, [API model specification](https://developers.openai.com/api/docs/models/gpt-6.1-sol) and [Artificial Analysis's independent report](https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence) were checked again on October 2. Published evaluation results and stated availability are evidence, not tests we ran with this team's imaginary workload.

## A patch with tests is not the same purchase as a result you must trust

Start with the software fix. Suppose the bug has a reproducible test and the proposed patch goes through the normal review queue. If Sol produces a working patch at lower cost, you can see it in a passing test, a readable diff and an accepted pull request. If it fails, the failure is comparatively cheap to detect. This is an *example of how to evaluate a switch*, not an account of our own deployment or a guarantee for every repository.

OpenAI's [DeepSWE v1.1 result](https://openai.com/index/introducing-gpt-6-1-sol/) gives this move a reason to try. The benchmark asks agents to solve long-horizon software tasks in real codebases. OpenAI says GPT-6.1 Sol **matches Astra at about one-fifth of the cost per task**, and beats the earlier Sol model's best score by **6.4 percentage points** at lower reasoning effort and cost. This is a vendor evaluation of a specific agent setup. It does not tell us how many changes a particular team's tests will catch, or whether those changes are the same kind of work.

The independent comparison is close but less tidy. At **max** reasoning effort, Artificial Analysis places Sol 6.1 **two points below Astra** on its Coding Agent Index. Set Sol to **xhigh**, and the report puts it **one point above Astra** and at **less than 15% of Astra's task cost on that index**. It also says Sol xhigh gains three points on Sol max. Leave the effort setting out of a recommendation and “behind” can become “ahead” without either source changing its data ([Artificial Analysis — coding results](https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence)). DeepSWE and this Coding Agent Index are different tests, not duplicate measurements of one universal success rate.

![Artificial Analysis Coding Agent Index showing GPT-6.1 Sol at max and xhigh effort alongside Astra, with corresponding task costs below.](/blog/images/gpt-6-1-sol-astra-replacement-coding-index.png "Effort changes Sol's position in this coding-agent evaluation; this is not a universal replacement claim. Source: [Artificial Analysis — GPT-6.1 Sol evaluation](https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence)")

| Coding Agent Index | Score against Astra | Cost against Astra |
|---|---:|---:|
| Sol 6.1 at max | 2 points lower | No verified figure quoted here |
| Sol 6.1 at xhigh | 1 point higher | Less than 15%, in this evaluation |

The table deliberately leaves Sol max's cost blank rather than reverse-engineering a number from a dense chart. Try the effort setting in the harness you intend to use, compare *accepted* changes, and keep a review path for ambiguous cross-file or security-sensitive patches. It is not yet a claim that Astra always wins that review.

## Documents, desktop actions and scientific results fail in different ways

Now take the PDF. OpenAI's **GDP.pdf** evaluation tests answers to professional questions from documents with tables, diagrams and fine print. Its report says Sol approaches Astra there at about **one-fifth of Astra's task cost** ([OpenAI — professional work](https://openai.com/index/introducing-gpt-6-1-sol/)). That supports trying Sol to prepare a sourced answer. It does not remove the need to check which row and note the answer came from before anyone signs a contract or relies on a clinical detail. The distinction is not “PDF good, science bad”; it is whether a wrong sentence will be caught before it matters.

The same question looks different when an agent operates a desktop. On **OSWorld 2.0's offline set**, which tests multi-step computer-use workflows, OpenAI reports Sol within **2.1 percentage points of Astra** at maximum effort, at roughly **one-seventh the task cost**. The reported measure is *partial reward* on an offline set released as v2026.08.08, not a production approval rate ([OpenAI — computer use](https://openai.com/index/introducing-gpt-6-1-sol/)). In our hypothetical workflow, letting an agent fill a form that a person will review is a different decision from letting it submit the form. A benchmark close to Astra cannot authorize the submission.

The difficult scientific task is where keeping Astra in the comparison has direct support. OpenAI's **Terminal-Bench Science 0.1**, a research workflow evaluation involving code and terminal tools, reports average maximum-effort task costs of **$5.47 for Sol 6.1 and $23.80 for Astra**. Sol more than doubles the older Sol's score, yet Astra has the highest tested score at **68.1%**. OpenAI itself recommends Astra for the most difficult scientific research tasks ([OpenAI — scientific research](https://openai.com/index/introducing-gpt-6-1-sol/)). That does not forbid Sol from research or establish a fixed price for a real project. It says there is at least one important measured slice where the more expensive path still buys a better result.

Artificial Analysis's **AA-Briefcase**, an evaluation of agentic knowledge work, adds another useful warning: Sol 6.1 improves its rubric score and Analytical Quality Elo over the old Sol, while Presentation Elo slips slightly ([Artificial Analysis — knowledge-work results](https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence)). Even within “professional work,” analysis and presentation need not move together. What looks like a near-tie on a composite can be a different trade in the part a reader needs.

## One-fifth token pricing is not one-fifth work

The official [Sol 6.1 API model page](https://developers.openai.com/api/docs/models/gpt-6.1-sol) lists four **Standard** prices in USD per million tokens: **$2 ordinary input, $0.10 cached-input reads, $2.50 cache writes and $10 output**. The 95% cache-read discount is relative to ordinary input, not to the entire bill. OpenAI's comparison with Astra is specifically the standard **input and output token rates**. A long answer, a failed attempt, another tool call or a second human review can change the completed-task bill without changing any of those rates.

Artificial Analysis measured one such bill, for **its own Intelligence Index**: at max effort, **$0.72 per task for Sol 6.1 versus $3.26 for Astra**. Sol is **one index point below Astra**; its task cost is less than one-quarter of Astra's in that test, *not* exactly the one-fifth ratio of standard token prices ([Artificial Analysis — Intelligence Index](https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence)). This index combines different work from the Coding Agent Index above; the two charts are not one leaderboard. The same report says Sol 6.1 uses **about 10–30% more output tokens than the older Sol 6**, not than Astra. Cheap output tokens do not imply an equal reduction in output volume.

![Artificial Analysis Intelligence Index and logarithmic cost-per-task plot comparing Sol 6.1 effort levels with Astra.](/blog/images/gpt-6-1-sol-astra-replacement-index.png "The lower panel measures cost per Intelligence Index task, not token price. Source: [Artificial Analysis — GPT-6.1 Sol evaluation](https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence)")

| Intelligence Index at max | Score against Astra | Average task cost |
|---|---:|---:|
| Sol 6.1 | 1 point lower | $0.72 |
| Astra | Reference | $3.26 |

There is also a billing boundary, not an excuse to turn this into a second pricing tutorial. The vendor model page says prompts with **more than 272,000 input tokens** incur **twice the input and cache rates and 1.5 times the output rate for the whole request**, rather than just the portion beyond the threshold. **Fast** costs twice Standard; Batch and Flex cost half; regional processing adds a premium where available. These are OpenAI's direct API conditions, not a guarantee about every third-party settlement path. Our [older Sol 6 task-cost comparison](/blog/gpt-6-sol-vs-claude-opus-5-5-cost-per-task) goes into cache mechanics; its September 22 Sol 6 cache-read price is not Sol 6.1's new $0.10 rate.

## Can your team make the switch where it actually works?

Before testing quality, check the entrance. The September 29 announcement, as reread October 2, says Sol 6.1 is available in **ChatGPT Work and Codex** for **Plus, Pro, Business, Enterprise and Edu**, and directly via the API as `gpt-6.1-sol`; it still says **“not yet available in Chat.”** That is the release page's wording, not a live audit of every subscriber account. It also describes Sol Ultrafast as coming in the following days; don't silently substitute that planned service for today's Standard price ([OpenAI — pricing and availability](https://openai.com/index/introducing-gpt-6-1-sol/)).

The [model page](https://developers.openai.com/api/docs/models/gpt-6.1-sol) adds a practical API constraint: **Responses** is the route for tool calling; **Chat Completions** is supported without it. If a coding agent depends on tool calls, changing only the model name may not give it the same work surface. The model accepts text and images, emits text, and lists no native audio or video; a tool exposed through Responses is not the same as a native output modality.

For the hypothetical team, the first experiment is modest: run the tested code fix at the intended effort, inspect the patch and actual cost per accepted change; then try the PDF with its original source open and record correction time. Put the high-stakes research question through a separate comparison before moving its final decision. Count failed attempts, tool charges and review time. We have not run those experiments, and published benchmark dollars cannot stand in for their results.

Sol 6.1 has earned a place at the **start** of this selection process. Its lower price and close coding results make it a compelling candidate for verifiable work; Astra remains a justified comparator where mistakes are costly and the scientific result is difficult to check. Which tasks ultimately move is decided by the team's acceptance evidence, not by the name on a price sheet. For another example of why Astra's premium varies by task slice, see our earlier [Astra evaluation comparison](/blog/astra-eval-worth-the-price).

---

## Sources

### openai.com

- [OpenAI's September 29 announcement](https://openai.com/index/introducing-gpt-6-1-sol/)

### developers.openai.com

- [API model specification](https://developers.openai.com/api/docs/models/gpt-6.1-sol)

### artificialanalysis.ai

- [Artificial Analysis's independent report](https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence)