OpenAI released GPT-6.1 Sol on September 29 as a lower-cost model for complex coding, computer use and professional work. Its headline API rates are $2 per million input tokens and $10 per million output tokens, one-fifth of GPT-6 Astra’s listed base rates, while retaining the same 1.05-million-token context window and 128,000-token maximum output.
That makes Sol the obvious first model to test for teams that want much of Astra’s capability without paying Astra prices. It is not, however, a drop-in budget switch. Long prompts trigger higher rates for the whole request, tool use requires the Responses API, the model cannot run with reasoning disabled, and the fastest processing options conflict with some data-residency requirements.
GPT-6.1 Sol pricing at a glance
| API usage | Standard rate per 1M tokens | Long-context rate |
|---|---|---|
| Input | $2.00 | $4.00 |
| Cached input | $0.10 | $0.20 |
| Cache writes | $2.50 | $5.00 |
| Output | $10.00 | $15.00 |
The long-context rates apply when input exceeds 272,000 tokens. Crucially, crossing that line reprices the full request, not only the tokens above the threshold. A request with 270,000 input tokens and 20,000 output tokens costs about $0.74 before tool charges. At 300,000 input tokens and the same output length, the cost rises to about $1.50.
Prompt caching can change that calculation substantially for applications that repeatedly send the same large instructions, reference material or tool definitions. Cached input is billed at five percent of the normal input rate, but writing a cache costs 1.25 times the uncached input rate. Teams should therefore measure how often a prompt prefix is actually reused before assuming caching will save money.
OpenAI also lists Fast processing at twice the Standard price, while Batch and Flex processing cost half as much as Standard. Regional processing adds 10 percent where available. Tool-specific charges, such as web search or computer use, sit on top of token costs. The current rates and limits are detailed on the GPT-6.1 Sol model page.
Why the Astra comparison needs real workload tests
OpenAI describes GPT-6.1 Sol as delivering “near-Astra performance,” but does not claim that the two models are interchangeable. Its own model-selection guidance tells developers to compare them on representative tasks. That matters because a fivefold reduction in token price does not guarantee a fivefold reduction in cost per completed job.
A model that needs more retries, longer reasoning traces or additional tool calls can erase part of the per-token advantage. Conversely, a cheaper model that clears a team’s quality threshold on the first pass can make Astra unnecessary for most requests. The useful unit is not cost per million tokens but cost per accepted result.
A sensible evaluation set should include the hardest common cases, known failure cases, long conversations, tool errors and tasks that require the model to stop or ask for approval. Record success rate, human review time, end-to-end latency, generated tokens and tool-call expense. Astra can then be reserved for jobs where the measured quality gain justifies the higher price.
Migration changes developers need to make
The model identifier is gpt-6.1-sol. OpenAI’s GPT-6 migration guide calls out several compatibility details:
- Use the Responses API for tools. Chat Completions accepts text-only requests, but GPT-6.1 Sol tool calling runs through Responses.
- Replace disabled reasoning. The model supports
low,medium,high,xhighandmax. It does not supportnoneorminimal; OpenAI recommends starting withlowwhen migrating those requests. - Remove incompatible sampling controls. Requests using reasoning should omit
temperature,top_pandtop_logprobs, along with the documented log-probability fields. - Review prompt-cache settings. Applications coming from GPT-5.5 or earlier should replace
prompt_cache_retentionwithprompt_cache_options.ttl. - Check residency before buying speed. Fast mode is unavailable with EU data residency. Ultrafast supports US residency and global processing, but not EU or other non-US regional endpoints.
GPT-6.1 Sol accepts text and images as input but does not accept audio or video. It supports streaming, structured outputs and function calling, and OpenAI lists web search, file search, code interpreter, hosted shell, computer use, MCP and other tools through Responses. Fine-tuning is not currently supported.
The lower price comes with Astra-level safeguards
OpenAI is treating GPT-6.1 Sol as having “Critical” cybersecurity capability under its Preparedness Framework and applies the same safeguard stack used for GPT-6 Astra. In the company’s internal ExploitBench evaluation, Sol achieved arbitrary code execution on 21.5 percent of recently disclosed vulnerability tasks, compared with 31.5 percent for Astra and 5.5 percent for the earlier GPT-6 Sol.
The system card addendum also reports 28 severity-three-or-higher misalignment flags across 49,650 simulated internal Codex tasks, or 0.056 percent. That was close to Astra’s 27 flags and below the earlier GPT-6 Sol’s 42. OpenAI separately found that monitoring the model’s full actions was more effective than inspecting its reasoning trace alone.
Those results are vendor-run evaluations, not a guarantee for any individual deployment. They do support a practical control choice: log and evaluate the model’s actions, tool calls and side effects, not only its final answer. Applications that can execute code, browse authenticated systems or modify production data still need scoped credentials, approval gates and rollback paths.
Who should move first
GPT-6.1 Sol is a strong candidate for coding agents, document-heavy professional work and computer-use workflows that currently rely on a premium frontier model. It is less compelling for simple classification, extraction or high-volume routing, where GPT-6 Luna or a specialized smaller model may remain cheaper.
Start with a shadow evaluation rather than changing a production alias immediately. Test Sol at low and medium reasoning, separate short- and long-context traffic, and calculate cost after retries and tool calls. Promote it for the workload segments where it clears the quality bar, then keep Astra as an escalation route for the small fraction of jobs that truly need it.
GPT-6.1 Sol’s launch is important because it moves advanced agent and coding capability into a much lower price band. The teams that benefit most will be the ones that treat the advertised rate as the beginning of the calculation, not the end.
