OpenAI’s DevDay on September 29, 2026 wasn’t a flagship launch — it was something arguably more useful: GPT-6.1 Sol, a cost-efficient model priced at $2 per million input tokens and $10 per million output tokens. That’s about one-fifth the price of the GPT-6 Astra flagship, while staying strong at coding, agents, and computer use.
Then on October 8, OpenAI made it faster too: Ultrafast mode for GPT-6.1 Sol, reaching up to 300 tokens per second — up to 8x faster generation than standard, now available to all API users (subject to rate limits) with US and EU data residency.
The full picture
Sol ships with a 1,050,000-token context window and 128,000 max output tokens. Three service tiers let you choose speed per request:
| Tier | Input / 1M | Output / 1M |
|—|—|—|
| Standard | $2.00 | $10.00 |
| Fast | $4.00 | $20.00 |
| Ultrafast | $12.00 | $60.00 |
Ultrafast costs six times the standard rate on every line. And here’s the detail that keeps pricing honest: at $60, Sol’s Ultrafast output still costs more than Astra’s standard $50. You’re paying for speed, not a bigger model. OpenAI’s own guidance is refreshingly blunt: use Ultrafast when speed justifies the higher cost.
The rollout keeps moving, too. As part of a “28 days of quality of life” push, GPT-6 Astra and GPT-6.1 Sol became roughly 50% faster by default across OpenAI products in early October — no code changes required.
Why “Sol” matters more than it sounds
Model launches usually mean bigger and pricier. Sol is the opposite bet: near-flagship intelligence at a price where high-volume use finally makes business sense. For agents that make hundreds of calls per workflow, cost and latency are the product — and Sol gives you the quality where it counts (coding, tool use, agents) at a fraction of flagship cost.
This is also OpenAI unbundling intelligence from latency. You no longer buy “a smart model” — you buy smart-cheap, smart-fast, or smart-maximum, per request, via the `service_tier` parameter. That granularity is new, and it’s the real story: infrastructure thinking applied to AI pricing.
What this means for you
If you build AI products: Sol is the new default for agent workflows. Reserve Astra for the 5% of tasks where maximum reasoning actually changes the outcome; route everything else to Sol Standard. For live, customer-facing features where a human is waiting — support chat, interactive coding, real-time troubleshooting — test Ultrafast and measure whether the conversion or satisfaction lift covers the 6x premium. Often it will; sometimes it won’t. Now you can choose per request instead of per project.
If you’re a freelancer using AI tools: your existing subscription quietly got better — models feel roughly 50% faster by default. And if you resell AI-powered services (chatbots, automation, content pipelines), your cost basis just dropped. Reprice your margins or pocket the difference.
If you run agents at scale: the tier system rewards routing discipline. Batch jobs, overnight research, background summarization — Standard tier, every time. Interactive sessions — consider Fast or Ultrafast. The teams that’ll waste money are the ones who leave Ultrafast on for everything; the teams that’ll win are the ones who route deliberately.
The strategic read
Frontier-class AI is becoming a commodity input, and Sol is the clearest evidence yet. The competitive edge is shifting from “who has the smartest model” to “who routes each task to the cheapest tier that can handle it.” Start auditing which of your AI workloads genuinely need maximum intelligence. Most don’t — and every one you downgrade is margin earned.
One caution from the developer community: higher throughput in high-effort modes can dramatically increase token consumption, so watch usage dashboards after enabling Ultrafast. Speed is a lever, not a default.
The bottom line
GPT-6.1 Sol plus Ultrafast gives builders something new: a dial, not a choice. Near-flagship smarts at commodity prices, with speed available on demand at a premium. Learn the dial, and you’ll outcompete anyone still paying flagship prices for commodity tasks.
Source: OpenAI’s official pricing and DevDay 2026 recap — GPT-6.1 Sol launched September 29, 2026; Ultrafast added October 8, 2026.
A Worked Example: What Sol Does to an Agent Bill
Prices per million tokens are abstract until you run a workload through them. Take a realistic agent workload for a small product: 10 million input tokens and 2 million output tokens a month — roughly, a busy support or research agent making hundreds of calls a day:
| Route | Input cost (10M tokens) | Output cost (2M tokens) | Monthly total |
|---|---|---|---|
| GPT-6.1 Sol, Standard ($2 / $10) | $20 | $20 | $40 |
| GPT-6 flagship tier (about 5× Sol, per the launch pricing) | $100 | $100 | $200 |
| GPT-6.1 Sol, Ultrafast ($12 / $60) | $120 | $120 | $240 |
The three rows are this article’s whole argument in one table. Sol Standard makes the workload five times cheaper than the flagship for the calls where near-flagship quality is enough — which, for most agent traffic, it is. And Ultrafast, the premium tier, costs more than the flagship route for the same tokens: you buy it for the up-to-300-tokens-per-second speed when a human is waiting, never as a default.
Frequently Asked Questions
What is GPT-6.1 Sol in one sentence?
OpenAI’s cost-efficient GPT-6.1 tier: near-flagship quality for coding, agents and computer use at $2 per million input tokens and $10 per million output — about one-fifth of the flagship’s price — with a 1,050,000-token context window and service tiers that let you pay for speed per request.
When is Ultrafast mode worth six times the price?
When a person is staring at a screen waiting for the answer: live support chat, interactive coding, real-time troubleshooting. At up to 300 tokens per second — up to 8x standard generation speed — the wait mostly disappears. For background jobs, queues and batch work, Ultrafast is money burned; Standard exists for those.
Should I route everything to Sol and skip the flagship?
No — route by consequence. Sol is the sensible default for the high-volume 95%: agent steps, drafts, classification, support. Keep the flagship for the small share of calls where maximum reasoning changes the outcome: the final decision, the hard proof, the high-stakes summary. That split is where the five-times saving actually comes from.
How do I choose the tier per request?
Through the service_tier parameter on each API call — the unbundling this launch is really about. The same application can send a background summarization job to Standard and the customer’s live question to Ultrafast seconds later. Design your routing once, then measure: cost per completed task, not cost per token, is the number that tells you if it works.
Is Sol available outside the API?
The speed push reached the apps too: as part of the same quality-of-life rollout, GPT-6 Astra and GPT-6.1 Sol became roughly 50% faster by default across OpenAI products in early October, with no changes needed on your side. The per-request tier choice, though, is an API feature — that granularity is the developer story.
Keep Reading
Keep exploring with these related HustlePrompt guides: