AI
OpenAI's GPT-6.1 Sol matches premium Astra on performance for one-fifth the cost
A week after launching GPT-6 Sol, OpenAI upgraded the model to nearly rival its pricier flagship while keeping costs the same.
The New Stack
OpenAI released GPT-6.1 Sol on Tuesday, just seven days after its predecessor. The updated model delivers Astra-level performance on coding, agent work, and professional tasks while costing just one-fifth the price of the company's flagship offering.
Unchanged pricing, notably stronger results#
GPT-6.1 Sol maintains the same pricing structure as its predecessor: $2 per million input tokens and $10 per million output tokens, with discounted cached input at $0.10 per million tokens. It's now available through the OpenAI API and all ChatGPT Work and Codex tiers (Plus, Pro, Business, Enterprise, and Edu), though Chat access isn't yet enabled.
The model includes a new Ultrafast variant in Codex that generates tokens up to 8 times faster than standard speed.
Despite only a week of development time, benchmark comparisons reveal substantial gains. On nearly every metric OpenAI tested, GPT-6.1 Sol performs almost identically to the flagship GPT-6 Astra model—making it hard to justify spending five times more for Astra on most tasks.
Strong performance on key technical work#
On coding tasks, GPT-6.1 Sol improves 6.4 percentage points over GPT-6 Sol on the DeepSWE 1.1 benchmark while essentially matching Astra. The model also beats Anthropic's newly launched Opus 5.5 on complex document-handling tests: it scores around 32% on the GDP.pdf benchmark compared to Opus 5.5's 29%, again at significantly lower per-task cost.
Computer use represents a standout strength. GPT-6.1 Sol outperforms GPT-6 Sol by seven percentage points at maximum reasoning while cutting costs in half—and still matches Astra's performance level.
Mixed competition with Sonnet 5.5#
Comparisons with Anthropic's Sonnet 5.5—released Monday at matching prices to GPT-6.1 Sol—show less decisive results. On DeepSWE, GPT-6.1 Sol leads at 75% versus Sonnet 5.5's 71%, but on AutomationBench the positions reverse: Sonnet 5.5 achieves 44.7% compared to GPT-6.1 Sol's 36%. However, GPT-6.1 Sol maintains a substantial cost advantage per task on the latter benchmark ($0.30 versus $1.14).
