GPT-6.1 Sol launches at GPT-6 Sol prices, with cached input cut in half

[OpenAI released GPT‑6.1 Sol](https://openai.com/index/introducing‑gpt‑6‑1‑sol/) on September 29 as the new workhorse in its GPT‑6 family. The API model ID is `gpt‑6.1‑sol`, with standard prices of $2 per million input tokens and $10 per million output tokens—the same headline rates OpenTools records for GPT‑6 Sol. The direct price change is cached input: $0.10 per million tokens, half GPT‑6 Sol's $0.20 cached‑input rate and 95% below standard input. That difference matters most for agents that repeatedly reuse a long system prompt, repository context, policy corpus or other stable prefix. At Standard‑tier list price, one billion eligible cached tokens costs $100 with GPT‑6.1 Sol instead of $200 with GPT‑6 Sol. The [current GPT‑6.1 Sol model reference](https://developers.openai.com/api/docs/models/gpt‑6.1‑sol) adds an important boundary: a request with more than 272,000 input tokens is billed at twice the input and cache rates and 1.5 times the output rate for the entire request. The $100 example therefore applies only below that threshold; it does not cut the price of fresh input or output, and workloads with little cache reuse should not expect that saving. OpenAI markets the model as approaching GPT‑6 Astra across agentic coding, computer use and professional tasks at one‑fifth of Astra's standard input and output prices. That is a model‑maker comparison, not a universal cost‑per‑task guarantee. The practical change is that developers can now test Astra‑like workflows at the Sol price tier, then measure whether task completion, output length and retries improve enough to reduce the total bill. The maintained [GPT‑6.1 Sol model record](https://opentools.ai/llms/gpt‑6‑1‑sol) carries the stable pricing and model facts, while the [OpenAI organization profile](https://opentools.ai/organizations/openai) connects the launch to its maker.

The strongest launch evidence is not one aggregate score. [OpenAI reports](https://openai.com/index/introducing‑gpt‑6‑1‑sol/) that GPT‑6.1 Sol matches GPT‑6 Astra on DeepSWE v1.1 at roughly one‑fifth the cost and exceeds GPT‑6 Sol's best score by 6.4 percentage points at a lower effort and cost. On the document‑heavy GDP.pdf evaluation, OpenAI says it beats Opus 5.5 with fallbacks at less than half the cost across the tested reasoning settings and approaches Astra at about one‑fifth the cost per task. The company also reports a 2.2‑percentage‑point lead over Opus 5.5 on AutomationBench at medium reasoning effort at roughly one‑third the cost, and a 4.8‑percentage‑point improvement over GPT‑6 Sol at the same setting. For computer use, OpenAI reports GPT‑6.1 Sol seven percentage points above GPT‑6 Sol on OSWorld 2.0 partial reward, measured on the offline set release v2026.08.08 at maximum effort, and within 2.1 percentage points of Astra at roughly one‑seventh the cost per task. On Terminal‑Bench Science 0.1, OpenAI says Sol more than doubles GPT‑6 Sol's score and averages $5.47 per task, while Astra still leads the tested models with 68.1%. These are first‑party evaluation results. OpenAI notes that its research environment and API harnesses may differ from production ChatGPT, and several claims compare different reasoning settings or include competitor fallbacks. A production decision should preserve the exact benchmark, effort, tools and cost denominator rather than compressing the launch into “near Astra.” This article therefore labels the five launch comparisons as first‑party OpenAI results. Separately stored HealthBench observations in the OpenTools model record retain OpenAI first‑party provenance; neither set is presented as an independent rerun.

[OpenAI's factuality test](https://openai.com/index/introducing‑gpt‑6‑1‑sol/) uses de‑identified conversations where users had already flagged a prior model error. At low reasoning effort, the share of answers containing at least one factual error fell from 11.4% for GPT‑6 Sol to 7.7% for GPT‑6.1 Sol. That is a 3.7‑point absolute change and about a 32% relative reduction. Across the tested effort levels, OpenAI says the new model remains within 1.9 points of Astra. The test is deliberately difficult and is not representative of ordinary traffic. It measures whether an answer contains at least one factual error, not the number or severity of errors. Those limits belong beside the result because otherwise a reader could mistake a challenge‑set rate for the expected failure rate of a deployed support or research agent. The [system‑card addendum](https://cdn.openai.com/pdf/38e3efcf‑545e‑44cd‑99ec‑2b7eb395f4cc/oai_GPT_6_1_Sol.pdf) provides a second safety boundary. OpenAI treats GPT‑6.1 Sol as Critical for cybersecurity and High for biological and chemical capability under its Preparedness Framework, using the same safeguards stack as GPT‑6 Astra. It also reports broad improvements over GPT‑6 Sol on agentic safe‑completion tests, while some individual categories move less or regress. That makes safety behavior another migration surface to test rather than a reason to assume compatibility from the shared Sol name.

[GPT‑6.1 Sol is available now](https://openai.com/index/introducing‑gpt‑6‑1‑sol/) through the API and to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. OpenAI says it is not yet available in Chat. Developers choosing a model for a conversational product should therefore distinguish API access from the ChatGPT model picker. Ultrafast is also a separate launch state. OpenAI says GPT‑6.1 Sol Ultrafast will arrive in the coming days, with up to eight‑times faster token generation than standard speed in Codex. The standard model is live now; the faster tier should not be treated as current Sol availability until its own endpoint and pricing are documented. Teams moving from GPT‑6 Sol can start with an exact task set: pin the model and effort, confirm cache‑hit rates, measure strict completion and retries, and record cost per successful task. Agentic coding deserves repository‑level tests, computer use needs the actual applications and permission boundaries, and document work needs representative tables and charts. The launch evidence supports a serious evaluation. It does not remove the need to reproduce the result in the environment paying the bill. *Figure: GPT‑6.1 Sol launch pricing and verification checklist. Source: [OpenAI's GPT‑6.1 Sol launch](https://openai.com/index/introducing‑gpt‑6‑1‑sol/) and [system‑card addendum](https://cdn.openai.com/pdf/38e3efcf‑545e‑44cd‑99ec‑2b7eb395f4cc/oai_GPT_6_1_Sol.pdf), published September 29, 2026. Figure by OpenTools Team; no third‑party expressive material used.*