SearchBook a call

BULLETIN / OpenAI

GPT-6.1 Sol: review

OpenAI's mid-tier model gives near-top results at a low cost per task, but it was one day old at review, with little independent testing.

Strong value among frontier models on early numbers, but wait for more independent results before switching critical work.

Maker
OpenAI
Released
29 Sept 2026
Licence
Proprietary
Access
OpenAI API, ChatGPT paid plans, Codex
Context
1,050,000 tokens
Price
$2 input / $10 output per million tokens (USD); higher above 272K input tokens

Good at

  • Low cost per task at high quality
  • Coding and computer use
  • Very long context
  • Efficient token use

Bad at

  • Very little independent data yet
  • Surcharge on prompts over 272K tokens
  • No audio, realtime or fine-tuning
  • Cannot switch reasoning off

Use it for

  • Coding agents on a budget
  • Production apps that need frontier quality at scale
  • Teams already using the OpenAI Responses API

Avoid it for

  • Voice and realtime apps
  • Teams that must fine-tune
  • Work that needs the very top score regardless of cost

Probably yes, for most production work. GPT-6.1 Sol scores close to the top of the Artificial Analysis Intelligence Index at a fraction of the cost per task of the leaders. The catch is age: it came out on 29 September 2026, one day before this review, so almost no independent evaluation had covered it. It also cannot do voice or be fine-tuned.

First principles

  • Value is score per dollar, measured per task. Price per token matters only through the tokens a task uses. On the Artificial Analysis index, Opus 5.5 at max effort scored 6 points more than Sol at max effort, for about eight times the cost per task ($5.98 against $0.72). Whether those 6 points are worth it depends on what an error costs you.
  • Evidence builds up over time. A score is one measurement on one test set. A model released the day before we read our sources has one independent index score, so our confidence in it is lower than in a model that several evaluations have measured. Results for its siblings, GPT-6 Sol and GPT-6 Astra, share a lineage but are not results for this model.
  • Long prompts change the price curve. Sol’s rates step up above 272K input tokens, so the cost of very long prompts grows faster than the headline price suggests. Model your real prompt lengths before you compare it with a flat-priced rival.
  • Safeguards trade some refusals for less misuse. OpenAI rates Sol Critical for cybersecurity, so it ships with Astra’s safeguards. Attack and defence often use the same techniques, so a filter strict enough to slow attackers will sometimes stop defenders too.

What it is

GPT-6.1 Sol is the middle tier of OpenAI’s current lineup, released on 29 September 2026. OpenAI describes it as close to GPT-6 Astra in quality at a lower cost, aimed at coding, computer use and professional work. It takes text and images and returns text. The context window is 1,050,000 tokens, with up to 922,000 tokens of input and 128,000 tokens of output. Its knowledge cutoff is 30 April 2026.

Reasoning effort runs from low to max. There is no setting that turns reasoning off.

The siblings:

  • GPT-6 Astra (released 3 September 2026) is the top tier at $10 input and $50 output per million tokens. OpenAI’s system card says advanced cyber access for Astra is phased in through its Trusted Access for Cyber program.
  • GPT-6 Luna (released 22 September 2026) is the small tier at $0.10 input and $0.50 output.

Where it is good

Value is the headline. On the Artificial Analysis leaderboard, read on 30 September 2026, GPT-6.1 Sol at max effort scored 52. Opus 5.5 at max effort scored 58. But Sol cost $0.72 to run one index task, against $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra. At high effort Sol scored 50 for $0.32 a task. Few models near the top come close to that ratio.

Its siblings give a sense of the family’s strength in independent tests:

  • Epoch AI ranks GPT-6 Astra first on its Capabilities Index at 166.6, and first on FrontierMath Tiers 1 to 3 at 93.7%.
  • Scale’s Humanity’s Last Exam leaderboard (updated 17 September 2026) puts GPT-6 Astra first at 54.80.
  • GPT-6 Sol, the earlier Sol release, placed sixth on the Arena web development board at 1692.

The model page lists structured outputs, function calling, web search, file search, code interpreter and computer use through OpenAI’s Responses application programming interface (API). Batch and Flex processing cost half the standard rate.

Where it is bad

There is little independent evidence on this exact version. The Vals Index lists GPT-6 Sol (57.54%) but not GPT-6.1 Sol. Epoch AI has no Capabilities Index score for any GPT-6 Sol or Luna model yet. Much of the case above rests on one leaderboard and on its older sibling.

The 57.54% Vals score for GPT-6 Sol is also well behind the Claude 5.5 models at about 67%.

Long prompts cost more. Above 272K input tokens, input and cache prices double and output rises by half.

It has gaps in features. The model page lists no support for realtime, audio, fine-tuning or predicted outputs.

OpenAI’s system card addendum rates it Critical for cybersecurity and High for biological and chemical risk under its Preparedness Framework. That means it ships with the same safeguards as Astra. Some legitimate security work may meet refusals.

Price and access

Tier Input per million Output per million Context
GPT-6 Astra $10 $50 1,050,000
GPT-6.1 Sol $2 $10 1,050,000
GPT-6 Luna $0.10 $0.50 1,050,000

Cached input for GPT-6.1 Sol is $0.10 per million tokens. The entry usage tier allows 500 requests per minute. TechCrunch reports it is in ChatGPT Work and Codex for paid plans, but not yet in standard ChatGPT chat. We could not confirm this on an OpenAI page, because openai.com returned errors to our fetches. There are no open weights; see our gpt-oss review for OpenAI’s open models.

Who should use it

On this evidence it is the model to test first if you want frontier-class coding or agent work at a sustainable price. Keep an evaluation harness running against it, because the independent picture will fill in over the coming weeks. For voice, use a realtime model instead. For the hardest maths and science problems, GPT-6 Astra has the stronger independent record.

Sources