What are the strengths of this model?

Half of GPT-5.6 Sol's price at $2/$10, with context raised to 1.05M tokens 33.2% on AutomationBench at $0.27 per task - the launch's most honest cost disclosure Cached input at $0.20 per million (90% off) cuts the effective rate for long-running agents

What are the weaknesses of this model?

On DeepSWE v1.1 (68.8%) and OSWorld 2.0 it misses GPT-5.6 Sol's peak - what improved is cost, not the ceiling All figures are vendor-reported at xhigh/max effort, and no system card shipped with the launch Above 272K input tokens a 2x input / 1.5x output multiplier applies

What are the best use cases?

High-volume agentic workflows with well-scoped tasks Code review, test fixes and small features inside Codex Cost-sensitive multi-step tool calling at scale

OpenAIProprietary

GPT-6 Sol

Name: GPT-6 Sol
Author: OpenAI

Compare this model

OpenAI's mid-tier GPT-6 model, released September 22, 2026 at $2/$10 per million tokens - line-for-line Claude Sonnet 5's rate card, and half of GPT-5.6 Sol. It runs a 1.05M-token context window with 128K max output. The published numbers sell cost, not a new ceiling: 33.2% on AutomationBench at $0.27 per task, 56.4% on Agents' Last Exam - while DeepSWE v1.1 at 68.8% sits below GPT-5.6 Sol's best.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

1.05M

License

Proprietary

Release Date

2026-09-22

Benchmark Performance

AA Intelligence Index

47.5

LMArena Elo

—

HLE

—

ARC-AGI-2

—

SWE-bench Verified

—

GPQA Diamond

—

MMLU-Pro

—

LiveCodeBench

—

AIME 2025

—

MATH-500

—

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

$10

Billing Mode: standard

Strengths

•Half of GPT-5.6 Sol's price at $2/$10, with context raised to 1.05M tokens
•33.2% on AutomationBench at $0.27 per task - the launch's most honest cost disclosure
•Cached input at $0.20 per million (90% off) cuts the effective rate for long-running agents

Weaknesses

•On DeepSWE v1.1 (68.8%) and OSWorld 2.0 it misses GPT-5.6 Sol's peak - what improved is cost, not the ceiling
•All figures are vendor-reported at xhigh/max effort, and no system card shipped with the launch
•Above 272K input tokens a 2x input / 1.5x output multiplier applies

Use Cases

•High-volume agentic workflows with well-scoped tasks
•Code review, test fixes and small features inside Codex
•Cost-sensitive multi-step tool calling at scale

Deep Analysis

Artificial Analysis Intelligence Index

39.8 (medium) / 47.5 (max)

AA Index v4.3.2; Claude Opus 5.5 reaches 57.6 at max effort

Measured Cost per Task

$0.25 (medium) / $1.06 (max)

Roughly one fifth of Opus 5.5's cost at matched medium effort

Input / Output Price

$2 / $10 per 1M tokens

50% below GPT-5.6 Sol's promotional rate; cached read $0.20, cache write $2.50

Context Window

1.05M tokens

128K max output; above 272K input, rates double and output goes to 1.5x

Safety Movement

Refusal rollback 68% -> 64%

Unauthorized-instruction execution in a simulated message-board test fell 52% -> 11%

Release

Sep 22, 2026

Knowledge cutoff Apr 20, 2026; OpenAI API, ChatGPT Work and Codex at launch

Strengths

・Rates match Claude Sonnet 5 line for line while OpenAI claims roughly half the error rate of its predecessor
・On OpenAI's AutomationBench it scored 6.3% above Claude Opus 5 at about 9% of the per-task cost
・Refusal rollback and unauthorized-instruction numbers improved materially, and the 52% to 11% result is a large jump
・Prompt caching for the GPT-6 line improved the same day, cutting cached-prompt cost by up to 90% for long-running agents

Weaknesses

・About ten Artificial Analysis index points below Opus 5.5 at matched max effort, so the ceiling is genuinely lower
・No system card published at launch
・Runs only on the OpenAI API plus ChatGPT Work and Codex; not available in ChatGPT itself yet
・A 272K-token input threshold doubles input and cache rates and raises output 1.5x, which Anthropic does not charge

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Claude Opus 5.5	AA 57.6 (max)	N/A	N/A	$4/$20
GPT-5.6 Sol (promo)	AA 59	N/A	N/A	$4/$20
Claude Sonnet 5	N/A	N/A	N/A	$2/$10

Overview

GPT-6 Sol is the workhorse half of OpenAI's September 22 pair, and its argument is arithmetic rather than benchmarks. At $2 per million input tokens and $10 per million output it lands exactly on Claude Sonnet 5's rate card, 50% below GPT-5.6 Sol's promotional pricing, and it inherits training methods from the GPT-6 Astra flagship. Artificial Analysis measures it at 39.8 on the Intelligence Index at default medium effort ($0.25 per task) and 47.5 at max ($1.06 per task) — roughly ten points under Claude Opus 5.5, at about a fifth of the cost. OpenAI also published concrete safety movement: refusals rolled back 68% to 64%, and unauthorized-instruction execution in a simulated message-board test dropped from 52% to 11%.

Benchmarks & Performance

OpenAI's published comparisons are all cost-normalized. On its AutomationBench professional-work evaluation, Sol scored 6.3% above Claude Opus 5 while costing about 9% as much per task. On Artificial Analysis's independent harness it reaches 39.8 at default medium effort and 47.5 at max, both below Opus 5.5's 51.2 and 57.6 — but at $0.25 and $1.06 per task against $1.34 and $5.98. The safety deltas are the more interesting numbers: refusal rollback improved four points and unauthorized-instruction execution fell by a factor of nearly five. No system card accompanied the launch, and no SWE-bench Verified score has been published for Sol specifically.

Detailed Comparison

**GPT-6 Sol vs Claude Opus 5.5** - Same day, different bets: Sol gives up about ten AA index points to charge half the rate per token and roughly a fifth per task. - Opus 5.5 ships with external evaluation and multi-cloud availability; Sol ships with cheaper caching and a lower ceiling. **GPT-6 Sol vs GPT-5.6 Sol** - The rate card halves, and OpenAI claims roughly half the error rate in the same move. - GPT-5.6 Sol's promotional pricing was always scheduled to expire in November; the GPT-6 launch rates are not given the same expiry. **GPT-6 Sol vs Claude Sonnet 5** - Identical pricing at $2/$10 and $0.20 cached reads. OpenAI is fighting Anthropic's mid-tier with Anthropic's own number.

Community Feedback

Developers read the launch as OpenAI conceding the top of the benchmark table and competing on volume instead. The $2/$10 figure was widely noted for being line-for-line Claude Sonnet 5's rate card, and the $0.10 Luna tier prompted immediate comparisons with self-hosted open weights. Agents teams welcomed the prompt-caching improvement more than the price cut, since cached prompts up to 90% cheaper changes the economics of long-running loops. The absence of a system card drew the same criticism Anthropic faced, in reverse.

Use Cases

**1. High-volume coding and CI** At $2/$10 with $0.20 cached reads, Sol is positioned for large codebases where the same instructions repeat. Enable caching before anything else; it moves the bill more than the headline discount. **2. Professional document automation** OpenAI's AutomationBench framing targets multi-step professional work rather than one-shot answers. Sol suits pipelines that must complete reliably at volume and can tolerate a lower reasoning ceiling. **3. Codex agent loops** Sol and Luna are the models behind Codex's paid tiers, so agent workloads running there get the price cut automatically. Check the 272K threshold in your harness before assuming the cheap rate applies end to end.

Latest News

**Sep 22, 2026**: OpenAI released GPT-6 Sol at $2/$10 per million tokens, 50% below GPT-5.6 Sol's promotional rates, alongside the cheaper GPT-6 Luna. Sol scores 39.8 at default and 47.5 at max effort on the Artificial Analysis Intelligence Index, with unauthorized-instruction execution falling from 52% to 11% in simulated testing.

Sources

Analysis generated: 2026-09-23