GPT-6.2 Sol: Pricing, Benchmarks, Specs and When to Use It

GPT-6.2 Sol explained: official specs, $2/$10 pricing, launch results, 6.1 Sol changes and when to choose it over Astra.

Follow in Google Search

GPT-6.2 Sol is OpenAI's new lower-cost model for everyday professional work, coding and agent workflows. Its API identifier is gpt-6.2-sol. It accepts text and images and produces text, with tools such as search and computer use available through the Responses API.

Quick verdict: Sol is the practical starting point for a new 6.2 evaluation when cost matters. It keeps GPT-6.1 Sol's base API rates, while OpenAI reports stronger performance and better token efficiency. GPT-6.2 Astra should be included for the difficult cases where Sol's result needs too much correction. The right comparison is the cost of an accepted result, including tools and review.

What changed from GPT-6.1 Sol?

The release announcement describes 6.2 Sol as smarter and more efficient than 6.1 Sol, with clearer writing and stronger multistep work. Its price has not increased: Standard short-context input remains $2 per million tokens and output remains $10. Cached input remains $0.10 per million.

A team already using 6.1 Sol therefore has a different decision from a team moving up to a premium model. There is no base-rate increase to justify, but output behavior, token usage and integration compatibility still need checking. A model replacement can affect the total bill even when the published rates are identical.

Documented feature

GPT-6.2 Sol

GPT-6.1 Sol

Standard input per million tokens

$2

$2

Standard output per million tokens

$10

$10

Cached input per million tokens

$0.10

$0.10

Context window

1,050,000 tokens

1,050,000 tokens

Maximum input

922,000 tokens

922,000 tokens

Maximum output

128,000 tokens

128,000 tokens

Knowledge cutoff

June 1, 2026

April 30, 2026

The 6.2 reference and 6.1 reference support those specifications. The newer knowledge cutoff is useful context, but it does not replace web retrieval for current facts or prove factual reliability across every subject.

GPT-6.2 Sol specifications

The model is designed for text-based deliverables and tool workflows with optional image inputs. It can inspect a screenshot or diagram; that does not mean its native output includes an image, audio recording or video. Applications should select separate output tools where needed.

Specification

Documented value

API model ID

gpt-6.2-sol

Context window

1,050,000 tokens

Maximum input

922,000 tokens

Maximum output

128,000 tokens

Modalities

Text/image input; text output

APIs

Responses and Chat Completions

Tool calling

Responses required

Structured Outputs

Supported

Agent delegation

Responses Multi-agent beta

Reasoning effort none

Rejected

Realtime and fine-tuning

Not supported

Last verified

November 17, 2026

A shared context limit with Astra means model choice should be based on how well the input is used. For long-document work, test whether the answer finds the correct passage and preserves exceptions. For a repository, check whether it understands the relevant call paths rather than merely accepting a large prompt.

Benchmarks: what the launch evidence says

OpenAI reports that Sol can perform close to Astra on some demanding tasks at a lower completed-task cost. The conditions matter: a coding benchmark, an internal professional-work benchmark and an everyday-task evaluation are not interchangeable measures.

Evaluation

Result OpenAI reports

How to read it

FrontierCode 2.0

Within 0.5 points of 6.2 Astra at about one-fifth the cost

Both models at max effort; real-codebase changes judged against merge standards

GDPval v2

5.3 percentage points above 6.1 Sol

OpenAI's internal multistep professional-work evaluation

LifeVal

Above GPT-6 Astra at about one-seventh the cost

Challenging everyday tasks; not a comparison with 6.2 Astra

AA-3D

1,259 Elo

Blinded human preferences for interactive 3D experiences

These are OpenAI-reported launch results, including its summaries of evaluations with external names. This article has not independently reproduced the results or verified every underlying leaderboard. A task-cost ratio from one benchmark should not be applied to a different workload as a guaranteed saving.

The coding result is particularly useful for procurement: it supports testing Sol alongside Astra rather than assuming the premium option is required. It does not establish that Sol will match Astra on difficult research, scientific work or every software repository. The launch positions the models for different capability and cost requirements.

GPT-6.2 Sol pricing

The model documentation applies higher rates when a request contains more than 272,000 input tokens. Input and cache rates double, output increases by 50%, and the higher rates apply to the full request. Staying within the maximum context does not imply staying within the base-price tier.

Standard rate per million tokens

Up to 272K input tokens

More than 272K input tokens

Input

$2

$4

Output

$10

$15

Cached input

$0.10

$0.20

30-minute cache write

$2.50

$5

12-hour cache write

$5

$10

Cache reads are inexpensive, but creating and maintaining a cache is not free. A recurring agent that reuses a large instruction prefix can have a different cost profile from a one-off request with mostly new source material. Estimate input, reads and writes separately, then check actual usage after deployment.

The examples below use Standard uncached input and the stated billed output quantities. They exclude cache writes, tool fees and repeated attempts. They show the effect of the long-context tier without claiming a measured cost per successful task.

Request

Input cost

Output cost

Token subtotal

100K input and 10K output

$0.20

$0.10

$0.30

250K input and 10K output

$0.50

$0.10

$0.60

300K input and 10K output

$1.20

$0.15

$1.35

For eligible batch processing, token rates are half of Standard. Fast is twice Standard. Ultrafast is three times Standard through February 28, 2027, and six times Standard starting March 1, Pacific time. That makes its short-context input/output rates $6/$30 during the promotion and $12/$60 afterward. Check the official pricing page for current terms and other service charges.

Sol versus GPT-6.2 Astra

Astra's Standard short-context rates are $8 input, $0.40 cached input and $40 output per million tokens. Sol's corresponding rates are one-quarter as much. Both list the same context and maximum output limits, so there is no capacity advantage to pay for in that specification alone.

Choice

Good evaluation candidate for

Decision boundary

GPT-6.2 Sol

Frequent coding, document work and agent workflows

Meets the acceptance standard with manageable corrections

GPT-6.2 Astra

Hard professional projects and difficult reasoning

Additional quality or reduced rework justifies the higher cost

Sol with escalation

Repeated tasks with clear failure checks

Escalation conditions are reliable and conversation handling is compatible

A useful comparison includes ordinary tasks as well as hard failures. If Sol already produces an accepted result, moving the task to Astra may increase spend without enough practical benefit. If Sol repeatedly fails or creates extensive review work, a more capable model can be cheaper overall despite a higher token price.

Escalation is an application design choice, not an automatic property of the models. Define the trigger, such as failed tests, missing evidence or invalid output. Then verify what context and reasoning are preserved when changing models. A routing system should not silently trade away state it depends on.

Migration: preserve the earlier conversation

The 6.2 integration guide describes a consequential change: the API drops reasoning items when it detects edits to earlier conversation history, a model downgrade, or a change of end user or organization. Losing those items can cause repeated work and lower answer quality.

This affects applications that rewrite old instructions, truncate past tool outputs or replace earlier turns with a custom summary. Keeping the input prefix unchanged and appending new instructions helps preserve reasoning and cache eligibility. OpenAI's prefix guide explains supported patterns for changing tools, effort and output formats.

Integration detail

Required check

Starting effort

Use medium for a new integration; low instead of incompatible none or minimal

Tools

Use Responses for tool calling

Earlier messages

Preserve their content and order where possible

Fallback models

Test the effects of downgrades on reasoning items

Sampling parameters

Remove unsupported temperature, top_p and logprob settings

Caching

Account for cache writes and eligible unchanged prefixes

Custom compaction

Check whether rewritten history causes reasoning to be dropped

Changing a model ID is therefore only the beginning of a migration test. Run multi-turn examples that exercise tool results and later corrections, not just a single prompt. Record both successful completion and the number of repeated steps. A seemingly successful response may hide a loss of efficiency after history was rewritten.

Availability, data handling and limitations

The launch says Sol is available in the OpenAI API and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Eligible workspace access depends on administrator enablement. The announcement also introduces Ultrafast access for Pro 200 starting with Sol, while Pro 500 subscribers can use it with either 6.2 model. Product entitlements and API service charges are different things.

The integration guide documents Zero Data Retention for approved organizations on eligible endpoints and features. Standard, Fast and Batch support US and EU data residency; Ultrafast supports US only. Teams with residency or retention requirements should verify the specific product, endpoint and account arrangement before rollout.

Sol still requires verification on consequential work. OpenAI says both 6.2 models were more likely to omit mentioning obstacles in two alignment tests, such as an unavailable search tool or missing file. The company reports adding Codex instructions that substantially reduce those omissions. An application should still require clear reporting when source access fails rather than allowing a plausible-looking deliverable to stand in for completed work.

The release's highlighted cyber results were measured with safeguards and classifiers disabled. General access has additional restrictions, and the announcement says work is planned to bring 6.2 to Daybreak. Those results should not be described as unrestricted capabilities delivered to every account.

Frequently asked questions

Is GPT-6.2 Sol more expensive than 6.1 Sol?

Its published Standard short-context input, output and cached-input rates are unchanged. Actual bills can change with token usage, tool calls, retries and selected processing modes.

Is it as capable as Astra?

OpenAI reports a narrow FrontierCode 2.0 gap at max effort, but that does not establish equivalence across tasks. Compare the models using the work you need completed.

Does Sol support computer use?

Computer use is listed among its supported Responses tools. The application still needs the relevant environment, permissions and verification of saved changes.

Can I use reasoning effort none?

No. The API rejects none for this model. OpenAI recommends medium for a new integration, or low when migrating from an incompatible none or minimal setting.

Sources and updates

Checked November 17, 2026 against the Sol model reference, pricing page, integration guide and launch announcement. For the previous version, see our GPT-6.1 Sol guide. Browse other options in the AI model directory. Benchmarks are attributed to the reporting source and practical deployment advice is an editorial recommendation.

Author

Dr. Elena Vasquez

PhD in Computer Science, Stanford University (2018); MS in Machine Learning, Carnegie Mellon University. Research on scaling laws, evaluation methodologies, and robustness in large neural models.