GPT-6.1 Sol specs, long-context prices, and published comparisons

Specs and prices below come from OpenAI’s model page and pricing table. The comparisons come from the announcement. They were checked on 2 October 2026. The announcement prints deltas and costs, not a table of absolute scores that can be copied out. Folkbench did not rerun these evaluations. “About one fifth of the cost” describes the vendor’s own test rig.

OpenAI describes GPT-6.1 Sol as an upgrade to gpt-6-sol. The API model ID is gpt-6.1-sol. The announcement says it comes close to gpt-6-astra on complex coding, computer use, and professional work, at one fifth of Astra’s standard input and output prices. Cached input is $0.10 per million tokens, 5% of standard input and half of gpt-6-sol’s cached-input price. For the hardest scientific research, the announcement still points to Astra. gpt-6-sol remains on the price table.

The GPT-6 Sol and Luna announcement points to this model in an update dated 29 September 2026. This article does not treat the date on a different news card in the footer as GPT-6.1 Sol’s own announcement date.

For the Claude list prices and benchmark tables, read [Opus 5.5](../claude-opus-5-5-review-pricing/) and [Sonnet 5.5](../claude-sonnet-5-5-review-pricing/).

Published specifications

The 922,000 maximum input matches the 1,050,000 context window minus the 128,000 maximum output: 1,050,000 − 128,000 = 922,000. The model page says reasoning tokens are supported. The passages cited here do not print a separate billing formula for them. Budget from the input, cache, and output on the usage record.

List prices

Short context and long context

Prices are US dollars per million tokens. Both tiers come from the API price table checked on 2 October 2026. That table says short context is input at or below 272K tokens, and long context is input above 272K tokens. The model page says that above 272K, input and cache are priced at 2× and output at 1.5×, for the whole request.

Input at or below 272K uses the short tier. Once a request goes above 272K, the earlier tokens are repriced at the long-tier rate as well. The extra tokens are not the only ones that change price. Cached input on 6.1 is half of Sol. Standard input and output are one fifth of Astra. Cached input is not one fifth of Astra: Astra’s short-context cached input is $1, and 6.1’s is $0.10.

A cache write is 1.25× uncached input, so the short tier is $2 × 1.25 = $2.50 and the long tier is $4 × 1.25 = $5. Long-tier output is 1.5× the short tier, $10 × 1.5 = $15.

Other multipliers on the price table

The model page lists three more official multipliers. Fast mode is 2× standard. Batch and Flex are 50% below standard. Regional processing, where it is available, adds 10%. On the same day’s price table, Batch, Flex, and Fast each have a row, and the numbers match those multipliers. Fast mode for 6.1 at short context, for example, is $4 input, $0.20 cached input, $5 cache write, and $20 output.

The announcement also says Ultrafast would arrive in Codex in the following days, with token generation up to 8× the standard speed. The flagship price table read on 2 October 2026 has no separate Ultrafast model row. Whether an official multiplier is already on a given bill is a question for the usage record.

Three examples that can be recomputed on the standard tier. The first two rows stay on the short tier. The third row is 280,000 input tokens, so the whole request uses the long tier. The cache row assumes the usage record counts those input tokens as cache hits.

Pricing those 280,000 tokens at the short-tier rate would give 0.28 × $2 + 0.01 × $10 = $0.66. That is not this tier. Fast, Batch, Flex, and the regional uplift are not applied in these three rows.

How to read the announcement comparisons

The announcement prints deltas and costs, not a table of absolute scores. The rows below keep only comparisons the page states. The cost sentences describe OpenAI’s test rig.

The factuality sample is de-identified conversations where a user had flagged an earlier error. The announcement says these prompts are chosen to be hard and do not represent ordinary traffic. The Terminal-Bench Science dollars are OpenAI’s published average cost per task, not a relay group price. The highest scientific score among the models tested is still Astra at 68.1%, and the announcement says the hardest scientific research should stay on Astra.

The same page also prints failure rates from alignment tests that are built to be hard. When the search tool is already broken, the share of max-effort runs that do not tell the user is 2.1% for 6.1, 4.9% for Sol, 1.5% for Astra, and 28.7% for Luna. The announcement says it observed no attempt to bypass an automated safety reviewer, matching Astra and Sol. The detail is in the GPT-6.1 Sol system card addendum. These failure rates are not failure rates on ordinary traffic.

The announcement also says the evaluations ran in a research environment or through the API, which can differ from production ChatGPT in system prompt and tools. Competitor figures are taken from public reports.

Two request paths to keep apart

Tool calls use Responses. The default effort is already medium. The model page says Chat Completions can carry text without tools.

# Responses
model: gpt-6.1-sol
reasoning.effort: medium

# Chat Completions, without tools
model: gpt-6.1-sol
reasoning_effort: medium

# Not accepted on the model page
reasoning.effort: none
reasoning.effort: minimal
tools on a Chat Completions request

When moving from gpt-6-sol or gpt-6-luna, which accept none, compare again at low. Do not leave none on the request. Tools, computer use, file search, and web search go through Responses. Chat Completions carries requests that have no tools. The Responses tools listed on the model page include web search, file search, the code interpreter, computer use, and MCP. Image input uses a content array on the user message.

Short-tier input and output match gpt-6-sol. Cached input falls from $0.20 to $0.10. Long-tier output is $15 on both, and long-tier cached input falls from $0.40 to $0.20. When switching to 6.1, pin the model name on the completed result. A completion that lands on another ID is not this switch.

Checks before moving traffic

1. Pin gpt-6.1-sol. The model name on the completed result has to be that ID as well.
2. Replay one set that contains no sensitive data at low, medium, high, xhigh, and max. none and minimal are not on this card.
3. Put tools on Responses. Leave tools off the Chat Completions request.
4. Price one prompt at or below 272K input tokens and one above 272K. The second uses the long-tier rate for the whole request.
5. Budget from the two standard rows first. Fast, Batch, Flex, and the regional uplift are whatever amount the usage record actually shows.
6. Read the cache tokens on the usage record before using the cache example above.
7. EU data residency cannot be combined with Fast mode. The model page says US and EU residency are supported. It does not describe residency as a parameter that can be added on an arbitrary gateway.
8. Keep the hardest scientific research on gpt-6-astra. gpt-6-sol is still on the price table.

After the model is chosen, whether a relay offers this ID, and that relay’s published availability and price, come from the directory record. The tables in this article are OpenAI list prices, not a station’s price after a group multiplier.