GPT-6 Sol vs GPT-5.6 Sol: Price, Benchmarks & Writing

The short answer

GPT-6 Sol costs half as much as GPT-5.6 Sol at today's API rates: $2 versus $4 per million input tokens and $10 versus $20 per million output tokens. Tibo Sottiaux says the new cut is permanent; GPT-5.6 Sol's current rates are promotional through at least November 21, 2026. That distinction matters if you are planning a long-lived product or a heavy writing workflow. [1, 2, 4]

Our take: The clearest upgrade is value per dollar. Performance is more mixed than either “no improvement” or “a huge leap” suggests: some published comparisons are close, while OpenAI reports meaningful gains on coding and factuality tasks. Writing quality deserves its own trial.

The price comparison

API Standard, per 1M tokens	GPT-5.6 Sol	GPT-6 Sol	Change
Input	$4.00	$2.00	50% less
Cached input	$0.40	$0.20	50% less
Output	$20.00	$10.00	50% less
Context window	1.05M tokens	1.05M tokens	Same
Maximum output	128K tokens	128K tokens	Same

Sources: OpenAI model pages and launch announcement. Prices are for standard API processing, short context, excluding tool calls and regional premiums. [1–3]

For a writing workload using 1 million input tokens + 200,000 output tokens, the published token rates imply roughly $8 on GPT-5.6 Sol or $4 on GPT-6 Sol. This is a simple token-price example, not a prediction of your bill; reasoning tokens, caching, tools, and long-context rates can change the result.

“Half the price” refers to the API token rates and related credits. It does not mean every ChatGPT or Codex subscription itself is 50% cheaper. OpenAI says the new models bring higher usage limits, but subscription allowances still depend on the plan and workload. [1]

Did the benchmarks actually improve?

There is no single score that captures this release. The two Sol models have the same published context and output limits. On some evaluations, the gap is small. Yet OpenAI's launch report says GPT-6 Sol improves substantially over GPT-5.6 Sol on FrontierCode, a coding test that considers whether changes are ready to merge. Its internal factuality test reports about half as many mistakes on a selected set of conversations that previously triggered factual errors. The company cautions that this set is not typical everyday use. [1]

So “almost no improvement” is too broad. A better reading is that the biggest universally measurable change is price, while the performance gain depends on the task and the test. Published benchmark gains also do not guarantee that your own prompts will feel better.

What about writing?

Tibo Sottiaux described GPT-6 Sol and Luna as noticeably better in writing and general “you know when you try it” quality. That is a product leader's assessment, not a controlled writing benchmark. OpenAI's release post makes a narrower, concrete claim: Sol's communication style is clearer, less jargon-heavy, less prone to odd phrasing, and a little shorter without losing substance. It includes one side-by-side example, while acknowledging that style is subjective. [1, 4]

That matters for people who use Sol to draft articles, rewrite dense material, or polish emails. A model can score similarly on coding tests and still be more useful when its first draft needs less editing. The only reliable way to judge that part is to give both models the same brief and compare the drafts blind: voice, factual accuracy, structure, and how much you have to revise.

Which one should you choose?

Start with GPT-6 Sol for API writing and routine coding if the model is available to you. The published rate halves the cost of the same token mix, and OpenAI claims improvements in communication style, coding, and factuality. Keep GPT-5.6 Sol in your comparison set if your existing prompts consistently produce a voice you prefer; model taste is not settled by a benchmark chart.

For important factual work, review the output from either model against original sources. For demanding code, evaluate success on your actual repository and consider effort settings, tool use, latency, and total cost per completed task rather than token price alone.

AIPlanWise verdict

GPT-6 Sol is an unusually easy model to try: the API price is cut in half, and its advertised improvements target work that people do every day. The writing claim is promising, but still subjective. Put your own briefs through both models before switching a production workflow.

Sources and notes

OpenAI: Introducing GPT-6 Sol and Luna — pricing, performance claims, availability, and writing example.
OpenAI: GPT-5.6 Sol model page — current promotional rate and model limits.
OpenAI: GPT-6 Sol model page — current rate, limits, and long-context pricing.
Tibo Sottiaux on X — personal assessment of writing quality. The post is attributed as opinion.

AIPlanWise has not run an independent blind writing evaluation. This article separates published prices and limits from provider-reported benchmarks and subjective impressions.