GPT-6.1 Sol’s 80% Discount Comes With a 272,000-Token Trap

GPT-6.1 Sol’s million-token context window has a pricing cliff: exceed 272,000 input tokens, and the entire request gets billed at double the input/cache rates and 1.5 times the output rate.

Not just the excess. The whole request.

That is the number I would put in an engineering review before the shiny “near-Astra intelligence” headline. OpenAI’s September 29 launch delivers a substantial price cut. It also gives developers another reason to stop treating context windows like free storage.

$12 versus $60 is worth paying attention to

At standard API rates, Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Astra costs $10 and $50 respectively. An identical workload consuming one million uncached input tokens and one million output tokens therefore costs $12 on Sol versus $60 on Astra, before tools and other charges.

Cached input is an even bigger discount: $0.10 versus $1 per million tokens.

This is meaningful. Repeated repository analysis, document processing, and multi-step automation become easier to justify when the model bill loses a zero in the right places.

But token prices are only the opening bid.

AWS’s Bedrock launch post makes the same economic argument: mistakes produce more interactions, tool calls, and human intervention. A cheap agent that needs an engineer to rescue it is an expensive agent wearing a discount sticker.

“Near-Astra” needs a workload attached

OpenAI reports that Sol matches Astra on DeepSWE v1.1 at approximately one-fifth the task cost, improving over GPT-6 Sol by 6.4 percentage points. On OSWorld 2.0, maximum-effort Sol finishes within 2.1 points of Astra at roughly one-seventh the task cost.

Those are compelling launch results, not independently established production equivalence. OpenAI still recommends Astra for the hardest scientific research.

Salesforce’s Jayesh Govindarajan offers a more concrete example: Sol identified accessibility and language-support problems and worked around limitations in the company’s test setup. That sounds useful. It also sounds like exactly the behavior your own evaluation suite should test rather than borrow from a customer endorsement.

My judgment: Sol deserves a serious trial as the default for complex routine work. Astra should earn its premium on failures and genuinely difficult tasks, not inherit every request because it tops the product ladder.

The integration bill doesn’t disappear

Sol supports a 1,050,000-token context window, 128,000 maximum output tokens, image input, structured outputs, and function calling. There is no native audio or video support and no fine-tuning.

A few details belong in the migration ticket:

• Tool calling requires the Responses API; Chat Completions does not support it for this model.
• Reasoning effort defaults to medium, with options through max; there is no none or minimal.
• Fast mode doubles Standard pricing. Batch and Flex halve it.
• EU residency rules exclude Fast mode.

Choose reasoning effort through measurement. Sending everything at maximum effort is not an architecture; it is a spending preference.

What Nobody Is Talking About

Sol 6.1 arrived seven days after GPT-6 Sol. Your regression suite barely had time to cool down.

Rapid releases make model procurement look less like buying software and more like operating a dependency that keeps changing beneath your application. Evaluation, security approval, routing rules, and cost forecasts all need maintenance. The savings are real; so is the labor needed to capture them.

There is also the permissions problem. OpenAI classifies Sol as having Critical cybersecurity capability and High biological and chemical capability. Its coding deployment simulation recorded 28 severity-3-or-higher misalignment flags across 49,650 matched tasks, versus Astra’s 27. Those are simulation findings, not production incident rates.

Cheaper capability is still capability. Keep credentials restricted, execution sandboxed, actions logged, and consequential operations behind approval gates.

Sol’s strongest pitch is not that every developer can afford a smarter chatbot. It is that more companies can afford agents that do real work. Give those agents a budget—and considerably less authority than an enthusiastic new hire.