OpenAI's newest model tier, Sol, costs one-fifth of the company's flagship tier, Astra, for the kind of work managers actually delegate: writing and checking code, digesting long documents, and running multi-step business workflows. That gap is not a rounding error on a promotional sheet. It is a number that quietly changes two decisions most teams made without knowing the current price – which rejected ideas deserve a second look, and where spending needs a limit before it grows.
A quick note on units before the math: AI models charge by tokens, the chunks of text a model reads and writes, roughly three-quarters of a word each on average. Some of that input can be cached, meaning the system re-reads text from earlier in the same conversation instead of processing it fresh, and cached text bills at a lower rate. Per million tokens, GPT-6 Astra lists input at $10.00 and output at $50.00, while GPT-6.1 Sol lists input at $2.00, output at $10.00, and cached input at $0.10. [1] Those are published list prices anyone can check, which is exactly why the decision below rests on them rather than on capability claims.
On capability, OpenAI reports that Sol "nearly matches" Astra on agentic coding, computer use, and professional work, measured on benchmarks the vendor chose to publish. [1] Treat that as the vendor's account of its own model, not independent certification. The price gap, by contrast, needs no trust in anyone's marketing. Build the decision on the number that is checkable.
Your blocked-idea list has stale numbers
Most teams carry a list of AI ideas that died in planning because the math did not work. Take a common example: a procurement manager wants every incoming supplier contract – often dozens of pages – fed into a model to flag non-standard terms. The work is repetitive by design. The same master-service clauses reappear in nearly every contract, and the model re-reads them in each new session.
That repetition is exactly what the new pricing rewards. Cached input costs ten cents per million tokens for Sol, one-tenth of Astra's cached rate, so work that loops over the same long documents gets cheaper the more it repeats. OpenAI says the cached rate exists to give room to run agents that reuse context across requests, which describes much of the work managers delegate. [1]
The consequence is a stale-numbers problem, not a strategy problem. A task the team rejected last quarter on cost may now clear the bar, because the price column in that old spreadsheet changed and the assumption did not. Before approving any new tool, pull the blocked-idea list from the last planning cycle and re-score each item at current prices. That is one afternoon of arithmetic, and it belongs ahead of any new subscription decision, because a subscription bought without checking current pricing is still a decision made on stale information.
Worth checking, too: Sol is open now to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex and through the API, and it is not yet in ordinary ChatGPT chat. [1] ChatGPT Work is the company's business workspace product; Codex is its coding product. If the team already pays for a Business plan, the access may already be there, which means the cheaper tier costs nothing extra to test.
Cheap removes its own brake
Price has always done two jobs. The obvious one is deciding whether a task is worth running at all. The quieter one is limiting how often people run tasks, and who can afford to run them at scale. When a capable tier arrives at one-fifth of the flagship's price, the second job weakens – for your team's budget, and for everyone else's.
Inside the team, the consequence is direct: usage that was self-limiting stops being so. A team member who ran a model check once a week because it felt costly will run it several times a week when the unit cost falls, so total spend can rise even as the price per unit drops. Reuters, reporting on the same market pressures, found that businesses have grown wary of ballooning AI costs even as labs cut prices. [2] The reliable control is a hard monthly spend cap per tool – a ceiling set in the billing dashboard, not a reminder in a chat channel. Set it before the cheaper model lands in the team's normal workflow, not after the first bill surprises you. Set in advance, a cap feels routine; set after a spike, it feels punitive.
The brake removal also sits outside the team entirely. Security researchers at the AI lab Anthropic, testing inside their own sandbox, recently showed that capable AI has become cheap enough that cost no longer meaningfully limits who can aim serious AI at an organization's internet-facing systems. [3] That finding does not ask a manager to run security evaluations. It asks for one line in the next security conversation: remove "cost controls access" from the list of assumptions about why your systems are safe, and hand the point to whoever owns that posture.
Two actions, before the next planning cycle
The practical shape is narrow. Neither action requires a new vendor relationship, a new subscription, or a meeting with IT.
First, pull the list of AI ideas the team rejected on cost, and re-run the numbers at current published prices – two dollars per million input tokens, ten cents per million cached. Mark any project that now clears the threshold. Those are this week's conversations.
Second, set the hard monthly spend cap per AI tool today, in the billing dashboard, before the cheaper tier becomes the default in normal workflow. Pick a ceiling the budget can absorb for genuine productivity. Deciding the number calmly is the whole point; the same number decided after a heavy month is a reaction, not a policy.
One caution travels with the prices themselves: OpenAI is cutting them under competitive pressure, and launch pricing is current, not guaranteed permanent. Both actions survive that uncertainty. Re-scoring the blocked list at today's prices is correct whether rates fall further or hold, and a spend cap protects the budget in both directions.
One-fifth the price, already inside plans many teams pay for. The question is not whether to use the cheaper tier – the vendor just made that easier. The question is whether the blocked-idea list is still blocked, and whether the spend cap gets set before the bill arrives or after it.