GPT-6.1 Sol Ultrafast pricing: $12 in, $60 out, and more per token than GPT-6 Astra
OpenAI's new speed tier doesn't change what Sol is. It changes what Sol costs, by a flat 6x. So the only real question is what your time is worth.
Photo by Lucas Gallone on Unsplash
Six times everything
| Per 1M tokens | Input | Cached | Cache write | Output |
|---|---|---|---|---|
| Ultrafast, up to 272K | $12.00 | $0.60 | $15.00 | $60.00 |
| Ultrafast, over 272K | $24.00 | $1.20 | $30.00 | $90.00 |
| Fast, up to 272K | $4.00 | $0.20 | $5.00 | $20.00 |
| Standard, up to 272K | $2.00 | $0.10 | $2.50 | $10.00 |
| GPT-6 Astra Standard | $10.00 | $1.00 | $12.50 | $50.00 |
OpenAI pricing page and GPT-6.1 Sol model page, read October 15, 2026. Ultrafast went live in the API, Codex and ChatGPT Work on October 8.
Every line is Standard times six, the 272K long-context step included. That makes the maths easy. Whatever your token mix or cache hit rate, an Ultrafast bill is six Standard bills. No workload gets a better deal than another.
The odd part is the last row. For the same tokens, Ultrafast GPT-6.1 Sol costs 20% more than GPT-6 Astra on Standard. OpenAI says so itself: its rollout post calls it "1.2x the cost of Astra" and pitches it as near-Astra intelligence, faster. Only cached input comes out cheaper, $0.60 against $1.00.
What one coding loop costs
Our test case is a 12-turn agent loop. Each turn re-sends a 32K prefix that's already cached, adds 8K of new context (tool output, file reads) and writes 2K tokens. We bill the new 8K at the cache-write rate, since it becomes next turn's prefix. Same tokens on every row, which flatters Claude and Gemini a little if their tokenizers count your code differently.
- GPT-6.1 Sol Ultrafast$3.11
- GPT-6 Astra Standard$2.78
- Claude Opus 5.5 fast mode$2.07
- Claude Opus 5.5 Standard$1.04
- GPT-6.1 Sol Fast$1.04
- Gemini 4 Argon Priority$0.85
- GPT-6.1 Sol Standard$0.52
USD per loop: 384K cached input, 96K cache-write input, 24K output in total. Opus 5.5 rows use its 0.05x cache read and 1.25x five-minute write. Argon Priority has no write fee, so the 8K bills at its $3.60 input; Google's page doubles that card on December 1. Solid accent: Ultrafast. Faded: other OpenAI tiers. Grey: other vendors.
$0.52 becomes $3.11. Run 50 of those a day for 22 working days and it's $570 on Standard against $3,421 on Ultrafast. Opus 5.5 in fast mode is the closest rival on speed pricing and still comes in a third cheaper, though Anthropic only promises up to 2.5x the output speed and keeps it behind a waitlist.
What the extra minutes cost
OpenAI's rollout post says "up to 8x faster" and the API guide gives no tokens-per-second figure at all. The 300 tokens per second doing the rounds came from DevDay, where it described Astra Ultrafast. So we went to Artificial Analysis, which now benchmarks every OpenAI tier of Sol at max effort.
| Sol at max effort | Output tokens/s | vs Standard | Index cost per task |
|---|---|---|---|
| Ultrafast | 318.2 | 6.5x | $4.35 |
| Fast | 98.2 | 2.0x | $1.45 |
| Flex | 74.5 | 1.5x | $0.36 |
| Standard | 48.7 | 1x | $0.72 |
Artificial Analysis provider benchmarks, read October 15, 2026. Cost per task is across its Intelligence Index. Flex being faster than Standard is what they measured; we'd expect that to change with load.
So the 6x price buys about 6.5x the speed. That's a fairer trade than we expected. Run our loop through it: 24K output tokens take 8.2 minutes on Standard, 4.1 on Fast and 1.3 on Ultrafast.
| Switch | Extra cost per loop | Minutes saved | Cost per hour saved |
|---|---|---|---|
| Standard to Fast | $0.52 | 4.1 | $7.50 |
| Standard to Ultrafast | $2.59 | 7.0 | $22 |
| Fast to Ultrafast | $2.07 | 2.8 | $44 |
Output generation time only, at the measured speeds above. Tool calls and test runs don't get faster, so they don't change the minutes saved, just the share of the loop they make up.
Fast is the bargain nobody's talking about. It gets you more than half the time back for a fifth of Ultrafast's surcharge. The last few minutes are where it gets pricey: going from Fast to Ultrafast costs $44 for every hour it saves. And the input matters more than the speed. Our loop reads 20 tokens for every one it writes, and you pay 6x on all that reading for a speedup that's about output. Double the context per turn and those hourly figures climb, while the minutes saved stay where they were.
Codex plans pay 8x, not 6x
In Codex, Ultrafast eats your included plan usage at 8x the Standard rate. Purchased credits and Enterprise pay-as-you-go go at 6x. It's only on the $500 Pro plan and on eligible Enterprise and Edu workspaces, and Enterprise admins have to switch it on. Plus and the $100 and $200 Pro plans can't use it at all, even with credits.
So if you're on Pro $500 and burning through included usage, Ultrafast is the most expensive way to spend it. Bring your own API key and Codex bills straight API token prices instead, which is the 6x card above.
Limits, regions and a missing Astra
Ultrafast has its own rate limits, separate from Standard and Fast: 1M tokens per minute on the Build tier, 4M on Launch and 40M on Grow. Sol Ultrafast runs with US and EU data residency. Astra Ultrafast is US only. The docs don't say whether it combines with Batch or Flex, and since they're all service_tier values we'd assume not.
Ultrafast is priced as a multiple of Standard, so if OpenAI ever moves Sol's card, expect this one to move with it. And the model it's being compared to may not get a successor soon. Press reports say a GPT-6.1 Astra planned for October was dropped. OpenAI hasn't confirmed that, but it would make sense of why fast Sol is being sold as the near-Astra option.
Who should flip the switch
For anything in a pipeline nobody watches, skip it. Standard Sol at $0.52 a loop is one of the best deals on the frontier right now, and Batch halves it again.
For a developer sitting in front of Codex, try Fast first. At about $7.50 per hour of waiting saved it's an easy yes for anyone on a salary. Ultrafast at $22 an hour against Standard is still cheap next to a developer's time, but most of that value is already in Fast. Measure before you switch a team: pull the timestamps from ten of your own loops on each tier. One benchmark at max effort isn't your workload. Our GPT-6.1 Sol launch post covers the Standard card in full.
Pages we read on October 15
- OpenAI: API pricing and GPT-6.1 Sol model page - Ultrafast, Fast and Standard rates, long-context step
- OpenAI: Ultrafast guide and changelog - availability, rate limits, data residency
- OpenAI community: Ultrafast rollout for GPT-6.1 Sol - "up to 8x", "1.2x the cost of Astra"
- OpenAI: Codex speed modes - plan eligibility, 8x and 6x usage rates
- Artificial Analysis: GPT-6.1 Sol providers - measured speed and cost per task for each OpenAI tier
- Anthropic: fast mode and pricing - Opus 5.5 at $8 and $40, up to 2.5x output speed
- Google: Gemini API pricing - Gemini 4 Argon Priority rates and the December 1 change
- Scalevise: Ultrafast in the API, Codex and ChatGPT Work - how Ultrafast is billed across API, Codex and ChatGPT Work
- AI Weekly: GPT-6.1 Astra reportedly cancelled - press report, not confirmed by OpenAI