Playbooks · Coding tasks
GPT-6.1 Sol, Ultrafast, and the new Pro tiers
Written November 4, 2026
OpenAI changed three things at once on September 29, 2026, at its DevDay developer conference: a new mid-tier model, a faster service tier, and a new top subscription plan. The model, the speed tier, and the plan bill separately, and confusing them is how a bill surprises you.
What changed
GPT-6.1 Sol is a mid-tier model in the GPT-6 family. OpenAI describes it as close to the flagship GPT-6 Astra on agentic coding, computer use, and professional work, at a fifth of Astra's standard token prices, and that description is OpenAI's own: it rests on OpenAI's benchmarks, not on an independent measurement. In the API it is model id gpt-6.1-sol. The pricing page lists it at $2.00 per million input tokens, $0.10 per million cached input tokens, and $10.00 per million output tokens, for prompts up to 272K input tokens. Above that length, the full request is priced higher: $4.00 input, $0.20 cached input, and $15.00 output. Cached input is the cheap part of the bill by design, at 95% below the fresh-input rate, which is why agents that reuse the same context across requests feel the difference most.
Ultrafast is not a model. It is a service tier — a faster way of serving a model — listed on the same pricing page for two models only, GPT-6 Astra and GPT-6.1 Sol. For Sol, Ultrafast costs $12.00 per million input tokens, $0.60 per million cached input tokens, and $60.00 per million output tokens at short context: six times the Standard rates, for the same model. Above 272K input tokens the Ultrafast rates are $24.00, $1.20, and $90.00. Speed claims for the tier are OpenAI's: up to eight times faster token generation in Codex, at about 300 tokens per second, and up to six times faster in the API. No independent latency measurement is claimed here, and the pricing page's multipliers describe price, not measured speed on any particular task.
The subscription change is separate from both. The Pro tiers help page lists three monthly Pro plans: Pro 100 at $100, Pro 200 at $200, and Pro 500 at $500. Pro 500 offers the highest included usage of the three, and Ultrafast is included on Pro 500 only: on Pro 100 and Pro 200 it is marked not included, and buying credits on those tiers does not unlock it. In ChatGPT Work and Codex, Ultrafast for Astra and Sol uses the plan's included usage first at eight times the Standard rate for the selected model, then draws from a credit balance at six times the Standard credit rate. Fast mode, the tier below Ultrafast, uses included allowance at 2.5 times the Standard rate. These are allowance consumption rates, OpenAI says, not guarantees of faster performance.
The allowance history, in brief
Allowances moved twice in about a month, which is why a multiplier quoted from the launch week no longer describes the current plans. At DevDay, OpenAI announced Pro 500 with its highest usage allowance and, for existing Pro 200 subscribers, a lower allowance from October 30. That reduction was then canceled: the help page now says Pro 100 and Pro 500 include twice their previous Work and Codex allowance, Pro 200 subscriptions all include the same Work and Codex allowance, and subscribers who already had the higher allowance keep it. The help page does not print allowance sizes as multiples of Plus, so no multiple is stated here. Check the help page itself before buying: the allowance that matters is the one in force on the day you subscribe, and it has already changed once after launch.
What each choice actually changes
- Choosing Sol over Astra in the API changes the token rate, not the task. At Standard rates, Sol costs a fifth of Astra's $10.00 input and $50.00 output. Whether a task finishes in the same number of steps and tokens on Sol is a property of the task, and the cheaper rate is only a saving if the step count holds.
- Choosing Ultrafast over Standard in the API changes price per token by a fixed multiple: six times, at either context length. It does not change the model, the context window, or the task. On a long agent run, the multiple applies to every token, including retries, so a run that retries often costs six times its Standard retry bill too.
- Choosing Pro 500 over Pro 200 changes two things: the size of the included Work and Codex allowance, and access to Ultrafast in Work and Codex. The $300 monthly difference buys the highest allowance plus the speed tier; it buys no additional model, since Sol and Astra are available on the lower tiers in Work and Codex as well.
- Crossing 272K input tokens changes the rate for the whole request. A request just over the line is priced at the long-context rates on every token, not only on the tokens past the line. Long-context agent work should be budgeted at the long rates, not the headline rates.
Steps
- Price the job at the tier you will actually use. Take one real task, count its input, cached input, and output tokens from a Standard run, and multiply by the pricing page rates for Standard, then for Ultrafast. The difference between the two totals is the price of speed on that task, and nothing else.
- Separate the model decision from the speed decision. Run the same task on Sol at Standard and on Astra at Standard before trying either at Ultrafast. If Sol finishes the task, Ultrafast on Sol is the cheaper speed upgrade; if it does not, paying six times for the same failure is the dearest way to learn that.
- Check the subscription terms on the day you decide. Read the Pro tiers help page for the current allowance wording and the Ultrafast column, and read Settings → My Plan for the tier and dates on your own account. Do not decide from a launch-week summary: the Pro 200 reduction announced at launch was canceled, and the doubled allowances arrived after it.
- Watch the first Ultrafast bill, not the first Ultrafast run. In Work and Codex, Ultrafast draws included usage at eight times the Standard rate. A first run that feels fast and a first bill that empties the allowance are the same event viewed twice. Set the allowance warning in your account before the first long Ultrafast session, and compare tokens used, not minutes saved.
What can go wrong
- Paying the Ultrafast multiple for a task that waits on tools. A task that spends its time waiting for shell commands, file reads, or web calls gains little from faster token generation and still pays six times per token in the API, or eight times the allowance in a subscription.
- Budgeting at short-context rates for long-context work. Agent sessions accumulate context. A session that starts under 272K tokens and grows past it reprices the whole request at the long rates, in every tier.
- Assuming the allowance from a launch-week summary. The help page now states the doubled allowances without a multiple, and the Pro 200 cut announced alongside it was canceled. An old summary is the wrong document for a current decision.
- Treating a vendor benchmark as a task result. OpenAI's claim that Sol nearly matches Astra comes from OpenAI's own evaluations. A builder's own task, with its own tools and failure modes, is the measurement that decides between them.
Limits
- Prices in this page are the API list prices on OpenAI's pricing page, in US dollars per million tokens, before any regional processing uplift. Regional endpoints add 10% for models released on or after March 5, 2026.
- Ultrafast availability differs by surface: in the API it is a service tier on the request, while in ChatGPT Work and Codex it is tied to the Pro 500 plan. A rate that applies on one surface does not imply the same access on another.
- No speed, quality, or cost-per-task measurement is reported here. The only figures stated are OpenAI's published prices and allowance rules, with vendor capability claims labelled as OpenAI's.
Related playbooks
- GPT-6.2 Astra and Sol in GitHub Copilot
- GPT-5 snapshots and o3 shut down Dec 11
- GPT-5.5-Cyber preview ends February 1, 2027
- GPT-6 Cyber: OpenAI's restricted security model
Sources
- https://developers.openai.com/api/docs/pricing?latest-pricing=ultrafast
- https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers
Back to: Coding playbooks hub · Also: Use your own Anthropic API key with an agent: what it costs, and the billing trap to check first · What an agent costs, in plain figures · Agent cost estimator · Stay inside the Muse Code subscription allowance