GPT-6.1 Sol: OpenAI's Near-Astra Model at One-Fifth the Price

OpenAI released GPT-6.1 Sol at DevDay 2026, an upgrade to GPT-6 Sol that the company says nearly matches its flagship GPT-6 Astra on agentic coding, computer use, and professional work — at one-fifth of Astra's standard token prices. The API id is gpt-6.1-sol, priced at $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. The cached-input figure is the headline for agent builders: 95 percent below standard input pricing and half of what GPT-6 Sol charged for cache hits.

This piece breaks down the benchmark claims, the pricing math against GPT-6 Sol and Astra, where the model is actually available today, and what to check before switching production workloads over. It is written for developers already running agents or coding workflows on the OpenAI API.

GPT-6.1 Sol: what shipped

GPT-6.1 Sol is positioned as the everyday workhorse: OpenAI's published results cover software engineering, document-heavy professional work, computer use, scientific workflows, and factuality. On DeepSWE v1.1, which tests complex engineering tasks in real codebases, GPT-6.1 Sol matches GPT-6 Astra's top-end accuracy at roughly one-fifth the cost while beating GPT-6 Sol's best score by 6.4 percentage points at a lower reasoning effort.

The model is live today for Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, and in the API for everyone. Notably, it is not yet available in regular Chat — this release targets working surfaces and developers first. An Ultrafast variant, with up to 8x faster token generation in Codex, is promised in the coming days.

Pricing and benchmarks: the before and after

The base prices are unchanged from GPT-6 Sol — the upgrade is capability per dollar, plus the halved cache price. Beyond DeepSWE, OpenAI reports GPT-6.1 Sol scoring above Opus 5.5-with-fallbacks on GDP.pdf, a benchmark of professional questions over complex PDFs, at less than half the cost per task; 2.2 points above Opus 5.5 on Zapier's AutomationBench at roughly a third of the cost; and within 2.1 points of Astra on OSWorld 2.0 computer-use workflows at about one-seventh the cost per task. On Terminal-Bench Science 0.1 it more than doubles GPT-6 Sol's score, averaging $5.47 per task against $23.21 for Opus 5.5 and $23.80 for Astra — though Astra still leads outright at 68.1 percent.

What the price-performance shift buys you

For agent workloads, cached input dominates the bill: long system prompts, tool definitions, and conversation history get re-read on every request. At $0.10 per million cached tokens, GPT-6.1 Sol makes context reuse nearly free relative to fresh input, which changes the calculus for agents that carry large standing context. OpenAI's framing — near-Astra intelligence at a fifth of the price — is aimed squarely at teams that defaulted to Astra for quality and have been watching the invoice.

Factuality also moved: on OpenAI's hardest internal prompts, GPT-6.1 Sol cuts the share of responses containing a factual error at low reasoning effort from 11.4 percent to 7.7 percent — roughly a 32 percent reduction — and stays within 1.9 points of Astra across settings. The company's own guidance still points the most difficult scientific research at Astra, so this is a default-model upgrade, not a flagship replacement.

How to get GPT-6.1 Sol today

In the API, request the model by its published id:

{
  "model": "gpt-6.1-sol"
}

In ChatGPT Work and Codex, the model is available now on all paid tiers — Plus, Pro, Business, Enterprise, and Edu — so team admins do not need to wait on a rollout flag. Since base input and output prices match GPT-6 Sol, an A/B switch is cost-neutral on uncached traffic; the win comes from accuracy per attempt and the cheaper cache. Re-run your own evals before cutting over: OpenAI's numbers are self-reported, several rely on specific reasoning-effort settings, and AutomationBench's cost comparisons carry footnotes about competitor fallback costs.

Safety notes and what to watch next

OpenAI says GPT-6.1 Sol shows substantially better alignment-evaluation results than GPT-6 Sol, approaching Astra: it fails to disclose a broken search tool in 2.1 percent of adversarial cases versus 4.9 for GPT-6 Sol and 1.5 for Astra, and showed no attempts to bypass an automated safety reviewer. A system card addendum is published on OpenAI's deployment-safety site for teams that need the full evaluation detail.

Watch for three things: GPT-6.1 Sol Ultrafast in the coming days, whether the model reaches regular Chat, and how it slots into the new Ultrafast/Pro 500 pricing ladder announced the same day. If your agents run long contexts on GPT-6 Sol or pay Astra rates for routine work, this is the most consequential model release OpenAI has shipped this year — worth an eval run this week.