GPT-6.1 Sol changed one number on GPT-6 Sol's rate card, cached input, from $0.20 to $0.10. The bigger cut is in the tokens: on medium it matches GPT-6 Sol's best score for a fifth of the cost per task.
OpenAI shipped GPT-6.1 Sol at DevDay yesterday, one week after GPT-6 Sol. The rate card moved by ten cents. The per-task bill moved a lot more than that, and almost none of it came from the price.
Photo by Chris Barbalis on Unsplash
One cell changed
GPT-6.1 Sol costs $2.00 in and $10.00 out, the same as GPT-6 Sol. Cached input is now $0.10 instead of $0.20. OpenAI moved the cache discount from 90% off to 95% off and left every other multiplier alone, so the halving shows up in Batch, Flex, Fast and the long-context rows too.
| Per 1M tokens | GPT-6.1 Sol | GPT-6 Sol | Sonnet 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Input | $2.00 | $2.00 | $2.00 | $10.00 |
| Cached input | $0.10 | $0.20 | $0.20 | $1.00 |
| Cache write | $2.50 | $2.50 | $2.50 (5 min) | $12.50 |
| Output | $10.00 | $10.00 | $10.00 | $50.00 |
| Batch / Flex in, cached, out | $1.00 / $0.05 / $5.00 | $1.00 / $0.10 / $5.00 | $1.00 / $0.10 / $5.00 | $5.00 / $0.50 / $25.00 |
| Fast in, cached, out | $4.00 / $0.20 / $20.00 | $4.00 / $0.40 / $20.00 | none | $20.00 / $2.00 / $100.00 |
| Over 272K input | $4.00 / $0.20 / $15.00 | $4.00 / $0.40 / $15.00 | no surcharge | $20.00 / $2.00 / $75.00 |
| Effort levels | low to max | none to max | low to max | low to max |
OpenAI and Anthropic pricing pages, read September 30, 2026. The over-272K row applies to the whole request, not just the tokens past the line. All three OpenAI models have a 1,050,000-token window (922,000 input plus 128,000 output) and add 10% for regional processing. GPT-6.1 Sol has no Ultrafast tier yet; DevDay coverage says one is coming.
What ten cents is worth on an agent loop
A cache discount only pays when most of your prompt is a cache hit. Agents are the case where it is: the system prompt, tool definitions and conversation so far get resent every turn. We priced 1,000 turns with 2,000 output tokens each and counted every uncached token at the plain input rate.
| Prompt, cache hit rate | GPT-6.1 Sol | GPT-6 Sol | Saving | Sonnet 5.5 |
|---|---|---|---|---|
| 100K, 50% | $125.00 | $130.00 | 4% | $130.00 |
| 100K, 90% | $49.00 | $58.00 | 16% | $58.00 |
| 200K, 95% | $59.00 | $78.00 | 24% | $78.00 |
| 400K, 90% | $262.00 | $334.00 | 22% | $172.00 |
Cost per 1,000 turns at Standard rates, same token counts for every model. Cache writes are ignored for all three.
Half of the cache price becomes 16% to 24% off the bill at the hit rates agents actually get. At 50% hits it is 4%, which you won't notice. Output is the other big line and it didn't move, so the saving shrinks as the replies get longer.
The last row is where Sonnet 5.5 wins. Anthropic bills its whole 1M window flat. OpenAI doubles input and cache and adds 50% to output once a prompt passes 272,000 tokens, and half-price cache doesn't close that gap: a 400K turn at 90% hits is $0.262 on GPT-6.1 Sol and $0.172 on Sonnet 5.5. Under 272K, GPT-6.1 Sol is never more per token than Sonnet 5.5 and is cheaper whenever there is a cache hit.
The bigger cut is in the tokens
Artificial Analysis had GPT-6.1 Sol on its index within a day. On max it scores 52 for $0.72 a task. GPT-6 Sol on max scored 48 for $1.05. That is four points more for 31% less, on a card where only the cache line changed. At max, 6.1 Sol wrote 67M output tokens across the index, against 77M for GPT-6 Sol.
- 58Opus 5.5 max$5.98
- 56Sonnet 5.5 max$7.60
- 53GPT-6 Astra max$3.26
- 52GPT-6.1 Sol max$0.72
- 48GPT-6.1 Sol medium$0.21
- 48GPT-6 Sol max$1.05
Score on the Artificial Analysis Intelligence Index v4.3.2 and cost per index task at list price, read September 30, 2026.
Medium, the default, is the row we would point most people at. It scores 48, the same as GPT-6 Sol flat out, for $0.21 a task. That is a fifth of the money for the same result. If you ran GPT-6 Sol on max last week, dropping to 6.1 Sol on medium is the biggest cost cut in this post.
Vals points the same way. Its index puts GPT-6.1 Sol at 61.15% for $3.24 a test, and GPT-6 Sol at 57.54% for $7.58. That is 57% cheaper per test on a different set of evals.
One point under Astra, at a fifth of the card
GPT-6 Astra on max scores 53 on the same index for $3.26 a task, 4.5x the cost of 6.1 Sol for one more point. OpenAI's own launch material leans on this: DeepSWE v1.1 about level with Astra at roughly a fifth of the cost per task, OSWorld 2.0 within 2.1 points at about a seventh. On Terminal-Bench Science 0.1 at max, OpenAI quotes $5.47 a task for 6.1 Sol, $23.21 for Opus 5.5 and $23.80 for Astra.
We couldn't load OpenAI's launch post, so the figures in that paragraph are OpenAI's as reported by The Next Web, VentureBeat and others, which all agree with each other. Vellum's reading of the charts shows the catch on AutomationBench: at medium, 6.1 Sol does edge Opus 5.5 there, 35.4 to roughly 33, but Sonnet 5.5 scores 44.7. OpenAI's comparison names Opus.
Sonnet 5.5 also leads both independent indexes: 56 against 52 on Artificial Analysis at max, and 67.04% against 61.15% on Vals. It pays for that. On max it costs $7.60 a task on AA, 10.5x what 6.1 Sol costs, and $21.34 a test on Vals against $3.24. For the same token price, the two models spend very different amounts of tokens.
Check this before you swap the model ID
GPT-6.1 Sol does not accept reasoning.effort: "none". GPT-6 Sol and Luna do. If you turned reasoning off on GPT-6 Sol to keep output short, your requests will error until you switch to low, and low will spend some reasoning tokens that none did not. Tool calling is still Responses API only.
GPT-6 Sol isn't deprecated. Its model page is live and points to 6.1, and it has no entry on the deprecations page, but it has dropped off the main table on OpenAI's pricing page. Codex already defaults to 6.1 Sol.
On Azure, GPT-6.1 Sol is generally available at OpenAI's card in the Global data zone, 10% more in the US zone and 20% more in EU and APAC, so an EU deployment pays $2.40 in, $0.12 cached and $12.00 out. It also launched on Amazon Bedrock the same day. AWS hasn't printed a rate on its own pages yet; LiteLLM carries the Global rows at list and the US in-region rows at 1.1x.
Where we land
Anyone on GPT-6 Sol should move. Same card or cheaper on every line, a higher score on both indexes, and fewer tokens per task. The one thing to fix first is the none effort setting.
Against Sonnet 5.5 it depends on prompt length. Under 272K tokens, 6.1 Sol is the cheaper bill per token and much cheaper per task, and Sonnet is the higher score. Past 272K, Sonnet is cheaper per token too.
Astra is now hard to justify for most work. One index point for 4.5x the per-task cost is a trade for a narrow set of tasks, and you should measure it on your own. The calculator prices any token mix against the new card.
Sources
- OpenAI: API pricing and GPT-6.1 Sol model page - Standard, Batch, Flex, Fast and long-context rows, context window, effort levels, regional uplift. Read September 30, 2026
- OpenAI: GPT-6 Sol model page and latest model guide - GPT-6 Sol rates, the pointer to 6.1, and the missing none effort
- Anthropic: Pricing - Sonnet 5.5 and Opus 5.5 rows
- Artificial Analysis: GPT-6.1 Sol - Intelligence Index v4.3.2 scores, cost per task and output tokens. Read September 30, 2026
- Vals AI: GPT-6.1 Sol - Vals Index score and cost per test, with the GPT-6 Sol, Astra, Sonnet 5.5 and Opus 5.5 pages
- The Next Web: GPT-6.1 Sol at DevDay - OpenAI's DeepSWE, OSWorld, AutomationBench and Terminal-Bench Science cost claims
- Vellum: GPT-6.1 Sol benchmarks explained - AutomationBench scores read from OpenAI's charts
- Microsoft: GPT-6.1 Sol in Foundry and AWS: GPT-6.1 Sol on Bedrock - availability and data-zone pricing