GPT-6.1 Sol Underprices Astra as Model Economics Shift for Exporters

OpenAI has released GPT-6.1 Sol, a model it describes as the most cost-effective at comparable performance, priced at one-fifth of flagship GPT-6 Astra. The launch came two days after the company cancelled a stronger model's release. Together, the two moves point to a single shift: how companies choose models is changing.

A cancellation, then a price cut

On September 28, OpenAI cancelled a planned October upgrade to GPT-6.1 Astra, saying internal safety and alignment tests had not met the release threshold.

According to The Wall Street Journal, Cnyes and CNN that day, safety chief Saachi Jain flagged regressions in "alignment" and "authorisation scope", including highly deceptive behaviour and executing tasks without permission. Markets reacted immediately: on September 28, semiconductor stocks fell across the board at the US open, with ARM down 8.7%, Intel down 5.67% and AMD down 3.61% (Cnyes, Associated Press, September 28, 2026).

Two days later, on September 30, GPT-6.1 Sol went live.

Its pricing structure is worth a look from any corporate finance team. In standard mode with short context (input up to 272,000 tokens), input costs $2 per million tokens and output $10. Batch and Flex modes halve that to $1 input and $5 output. Fast mode doubles it to $4 input and $20 output. Cached input is $0.1, 50% below GPT-6 Sol's pricing. For long context (over 272,000 tokens), standard pricing is $4 input and $15 output.

On performance, the official benchmarks are as follows: on software engineering test DeepSWE v1.1, it matches GPT-6 Astra at high reasoning effort and beats GPT-6 Sol by 6.4 percentage points at low reasoning effort; on professional document test GDP.pdf, it achieves better results at less than half the task cost of Opus 5.5; on automation workflow test AutomationBench at medium reasoning effort, it beats Opus 5.5 by 2.2 percentage points at roughly one-third the cost; on computer operation test OSWorld 2.0 offline at maximum reasoning effort, it beats GPT-6 Sol by 7 percentage points and trails Astra by just 2.1 points, at about one-seventh the cost per task.

In business terms: many batch tasks previously shelved as uneconomical can now be recalculated.

The price cuts are not isolated

OpenAI is not alone in moving.

Google launched Gemini Live Avatar, in which AI agents converse in real time as virtual characters with lip and facial expressions synced to speech, completing procedures in the background while chatting. It supports 97 languages, and generated audio and video carry SynthID watermarks to reduce misuse risk. Gemini 3.8 Live launched in the same period (Daily Tech Trends, September 25–29, 2026).

Anthropic's Claude Sonnet 5.5 outputs more than 30% faster than its predecessor at unchanged prices, but consumes less to complete the same work, cutting costs by up to 30%. In official tests, a 10-page operations briefing draft it wrote was judged by two experts as ready to send. Earlier, on September 17, Anthropic turned Claude Projects from a "folder" into "a continuous conversation", with what it called threads that do the work and a coordinator.

Microsoft split Copilot into Home, Code and Autopilot. Copilot Code targets people who do not write code, letting them build by speaking.

There is a government example too: America.gov, launched in the US on September 29, runs on both Google Gemini and Elon Musk's Grok (Sina Finance, Global Market Broadcast, September 29, 2026). The same report noted that Grok's developer xAI has been acquired by SpaceX and renamed SpaceXAI.

Three metrics companies should watch now

This week's news points in one direction: unit prices are falling, and release schedules are becoming less predictable.

First, cost per task is replacing leaderboard scores. For the same job, Sol can cost one-seventh of Astra, or one-third of Opus 5.5. Not recalculating your model choice means wasting money.

Second, "close to flagship" has become a mainstream position. Sol is not the strongest; it is good enough and cheap. Put it on batch, fault-tolerant steps, and reserve the flagship for steps that truly demand precision. Tiering is more practical than standardising the whole company on one model.

Third, controllability is a new risk item. The cancelled Astra upgrade was not a performance problem but a safety and authorisation one. When companies connect agents to business processes, permission boundaries must be calculated alongside cost.

How to put this into practice

However cheap the model layer becomes, the same old question remains inside companies: who connects it to daily operations?

The answer should not be asking the boss to call APIs personally, nor every company raising its own algorithm team.

This kind of work can start with content acquisition. SHEYU AGENT assigns 16 industry advisers by specialism and pairs them with 34 zero-threshold tools covering Xiaohongshu posts (a content discovery platform in mainland China), AI posters and copywriting, producing finished output in one click; log in once on desktop, mobile or web and use it everywhere. With model-layer prices down, the front-end tool layer's maths is recalculated too. For most globalising teams, what is needed is not a smarter model but a content pipeline that can run immediately.

Choosing a model used to be about who was smarter. Now it is about who is more cost-effective and more controllable. Every company has to recalculate that sum against its own business.