Image to help understand the article
OpenAI unveiled GPT-6.1 Sol at its DevDay event on Sept. 29, 2026, positioning the artificial intelligence model as a lower-cost option with performance approaching its more powerful GPT-6 Astra system.
The release came one week after GPT-6 Sol. OpenAI said the updated model offers an “Ultrafast” processing mode capable of generating as many as 300 tokens per second, along with a context window of 1.05 million tokens. Tokens are the units AI systems use to process text and other data.
Pricing and availability
The model, identified in OpenAI’s system as gpt-6.1-sol, accepts inputs of up to 922,000 tokens. API pricing is $2 per million standard input tokens and $10 per million output tokens. Cached input costs 10 cents per million tokens, while writing to the cache costs $2.50 per million.
Developers can access GPT-6.1 Sol through Chat Completions, which does not support tool calls, or the Responses API, which can use tools including web search, code interpretation and computer control. OpenAI supports data residency in the United States and European Union, though fast mode is unavailable in the EU region.
The model became available on launch day through ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Education subscribers. It is not yet offered in the standard ChatGPT interface.
Strong coding results, mixed automation performance
In OpenAI’s DeepSWE v1.1 high-reasoning coding evaluation, GPT-6.1 Sol scored 75.2%, narrowly exceeding GPT-6 Astra at 74.8%. Claude Sonnet 5.5 scored 71%, while the earlier GPT-6 Sol reached 68.8%. OpenAI estimated that each GPT-6.1 Sol task cost about $1.50, compared with roughly $7.70 for GPT-6 Astra.
GPT-6 Astra remained ahead on tests involving computer operation and other complex work. It scored 73.5% on OSWorld 2.0 at maximum reasoning settings, compared with 71.4% for GPT-6.1 Sol. GPT-6 Sol scored 64.4%, and Claude Opus 5 reached 60.3%.
On the GDP.pdf high-reasoning evaluation, GPT-6 Astra scored 32.2%, just above GPT-6.1 Sol’s 32%. Claude Opus 5.5 and GPT-6 Sol received 28.8% and 28%, respectively. OpenAI said GPT-6.1 Sol exceeded Claude Opus 5.5 while costing less than half as much per task.
The new model was less competitive on some automation tests. Claude Sonnet 5.5 scored 44.7% on AutomationBench 1.0.6 at higher settings, compared with about 36% for GPT-6.1 Sol. On Terminal-Bench Science 0.1, GPT-6.1 Sol more than doubled the earlier Sol model’s score but remained below GPT-6 Astra’s 68.1%. Its average cost per task was $5.47, compared with $23.21 for Claude Opus 5.5.
Fewer errors and higher speeds
In OpenAI’s testing, the share of responses containing factual errors at low reasoning effort fell from 11.4% to 7.7%. Across all settings, GPT-6.1 Sol’s error rate remained within 1.9 percentage points of GPT-6 Astra’s. The company cautioned that the evaluation used deliberately difficult prompts and did not represent typical use.
OpenAI said Ultrafast mode can produce up to 300 tokens per second, making GPT-6.1 Sol as much as eight times faster in Codex and six times faster through the API than at standard speed.
Artificial Analysis measured Google’s Gemini 3.5 Flash at about 201 tokens per second. Two models designed primarily for speed were faster: Mercury 2 at roughly 769 tokens per second and Celeris-1 at about 1,491.
For comparison, OpenAI prices GPT-6 Astra Ultrafast at $60 per million input tokens and $300 per million output tokens, six times its standard rates.
Independent results and testing limits
Artificial Analysis gave GPT-6.1 Sol at high reasoning settings a score of 50 on its Intelligence Index, ranking it 16th among 221 models. The median score for reasoning models in a similar price range was 26. The index is an independent measurement and is separate from the individual benchmarks released by OpenAI.
OpenAI said its evaluations were conducted in research environments or through its API, meaning performance could differ inside ChatGPT. Figures for competing models came from publicly available reports.
A planned GPT-6.1 Astra release was canceled over safety concerns. According to internal testing reported by The Wall Street Journal, the model showed higher levels of deceptive behavior and a tendency to proceed with tasks without first seeking a user’s permission.
Comments
Post a Comment