OpenAI cuts GPT-5.6 Sol API prices by up to 33%

• Sol's promotional API price is $4 per million input tokens and $20 per million output tokens, per OpenAI's announcement card, a 20% input and 33% output reduction.
• The lower rates run at least through November 21, 2026, according to Wes Roth's pricing post.
• Independent DeepSWE figures in reach_vb's chart put Sol Max at 72.7% for $6.47 per task, versus Claude Fable 5 Max at 69.7% and $21.63.
• Terra, not Sol, produced the roughly 82% lower cost per successful Terminal-Bench 2.1 task that OpenAI Developers reported from a Kiro test.
• Sol xHigh landed at $4.08 per Code task and $2.59 per Work task in Agent Arena's update.

A $4/$10 line in MatthewBerman's post conflicts with the $4/$20 rate shown in its attached OpenAI price card and in Wes Roth's later post. OpenAI's Kiro announcement also puts Sol alongside Terra and Luna inside AWS's spec-driven coding agent.

The official card resolves the mismatch: Sol is $4 per million input tokens and $20 per million output tokens. The stated percentage cuts imply previous list prices of $5 input and $30 output per million tokens.

The $10 output figure in MatthewBerman's post is therefore inconsistent with the visual announcement and the later pricing post. The promotion applies to Sol, following earlier Terra and Luna reductions.

The chart compares Max-effort runs on DeepSWE v1.1, a set of 113 original long-horizon software-engineering tasks. It places Sol at 72.7% pass@1 for a mean $6.47 per task, while Fable 5 Max reaches 69.7% at $21.63.

That is a 3.0-point pass-rate lead and roughly 70.1% lower mean task cost for Sol in the comparison. The chart identifies itself as independent analysis, rather than an OpenAI benchmark.

Agent Arena uses net improvement and median task cost over the preceding 14 days, rather than a pass-rate leaderboard. Its repricing update placed Sol xHigh on the Pareto frontier in two categories:

• Code: +13.1% net improvement at $4.08 per task.
• Work: +11.2% net improvement at $2.59 per task.

The same update says Sol's 20% price reduction moved its Code and Work position between Claude Fable 5 High and Kimi K3.

Kiro now offers all three GPT-5.6 models for planning, building, testing, and reviewing software. Its spec-driven flow turns a product idea into requirements, technical designs, and executable tasks, as OpenAI's Kiro post describes.

OpenAI and AWS reported an approximately 82% reduction in cost per successful Terminal-Bench 2.1 task for Terra running in that environment. The figure measures a model-harness combination and does not describe Sol's standalone API price or its DeepSWE result.

OpenAI's developer-community announcement describes the Sol change as a three-month reduction in API and credit pricing. It says API access is live, eligible ChatGPT Work plans and Codex credits are rolling out, and Pro, Plus, and Business subscription usage is unchanged.

The public end-date language is "at least through November 21, 2026," as Wes Roth's post records, leaving open whether OpenAI extends the promotional rate.