Verdict, October 28, 2026. OP on a Cursor Premium, Super, or Ultra seat when a mid-40s score on Cursor's own chart is enough and you are already spending prepaid credit on a third-party model. Mid on Free and Plus, where that model is not available. Mid on Teams and Enterprise, where a $0.25 per million token rate sits on a model the exclusion list leaves out. The scores and per-task costs below are Cursor's. OP or Mid has not run the tasks.
Pages below were read October 28, 2026. CursorBench figures are vendor claims. DeepSeek's API card is DeepSeek's.
What changed on October 26
NEW On October 26, 2026, Cursor posted that DeepSeek V4.1 Flash is available in Cursor, that every effort level sits on the CursorBench cost frontier from $0.60 to $0.98 per task, and that no model scores higher. Turn it on in Settings, then Models. Source: Cursor on X (opens in a new tab).
The model page names DeepSeek V4.1 Flash 1017, id deepseek-v4.1-flash-1017: a 1M-token context, reasoning levels None, Low, Medium, High, and Max, High by default, no image inputs. Cursor serves it through inference partners. It sits outside included usage. On The X Subscription the card below comes from prepaid credit. Legacy plans draw the Other Models pool. The price is flat across the window, with no long-context surcharge and no cache-write fee. DeepSeek's API lists the same version as deepseek-flash and marks Vision supported. Cursor's row does not take images. Sources: model page (opens in a new tab), DeepSeek pricing (opens in a new tab).
The token card
PRICE Per million tokens, the model page and the models table print input $0.3, cache write as a dash, cache read $0.006, output $1.2. The table marks the row hidden by default. Sources: model page (opens in a new tab), Models & Pricing (opens in a new tab).
PRICE DeepSeek's peak column matches that card, then halves it off-peak: peak cache-hit input $0.006, peak cache-miss input $0.3, peak output $1.2; off-peak $0.003, $0.15, and $0.6. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Cursor's card has one column. Source: DeepSeek pricing (opens in a new tab).
PRICE Free ($0) and Plus ($8/mo) list third-party models as not available. Plus includes Cursor's own models only: Grok 4.7, Grok 4.6, Grok 4.5, and Composer 2.5. Premium ($30/mo), Super ($100/mo), and Ultra ($200/mo) bill third-party models at list price from prepaid credit, from the first request. A monthly cap still limits paid usage when credit remains. Individual plans skip the Cursor Token Rate. Prices exclude tax. Source: Models & Pricing (opens in a new tab).
NERF Teams ($30, $100, or $200 per seat) and Enterprise add $0.25 per million input, output, and cached tokens on eligible third-party requests, including a Cursor Router Balance or Intelligence route. The exclusion list names Haiku 5.5. It leaves out DeepSeek V4.1 Flash 1017. Source: Models & Pricing (opens in a new tab).
Cursor's cost chart
VENDOR CLAIM CursorBench 4.0 scores agents on ambiguous, multi-file tasks from real Cursor sessions. Insights says cost per task applies published input, cache-read, cache-write, and output rates to the tokens each task used, that results vary, and that small score gaps may not be statistically meaningful. Its High-effort slice lists this model at 45.6% and $0.77. Sources: CursorBench (opens in a new tab), Insights (opens in a new tab).
The full table, read October 28, has four DeepSeek rows and no None row. Neighbors are the rows next to those prices. Every figure is Cursor's.
| Row | Score | Cost / task | Tokens / task | Steps / task |
|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 Low | 44.9% | $0.60 | 290,651 | 161 |
| DeepSeek V4.1 Flash 1017 High | 45.6% | $0.77 | 385,196 | 186 |
| DeepSeek V4.1 Flash 1017 Medium | 46.3% | $0.80 | 398,255 | 189 |
| DeepSeek V4.1 Flash 1017 Max | 47.7% | $0.98 | 537,047 | 187 |
| Haiku 5.5 Extra High | 44.3% | $0.56 | 143,813 | 93 |
| Haiku 5.5 High | 42.3% | $0.32 | 77,057 | 59 |
| Haiku 5.5 Max | 48.4% | $1.12 | 325,934 | 163 |
| Sonnet 5.5 High | 47.8% | $1.20 | 37,391 | 41 |
| Composer 2.5 | 27.7% | $0.68 | 17,347 | 41 |
Source: CursorBench (opens in a new tab).
PRICE On that table, a higher score than each DeepSeek row costs more than that row. Nothing above 47.7% costs $0.98 or less. Nothing above 46.3% costs $0.80 or less. Nothing above 45.6% costs $0.77 or less. Every other row at $0.60 or less scores under 44.9%. The next scores above Max are Sonnet 5.5 High at 47.8% and $1.20, and Haiku 5.5 Max at 48.4% and $1.12. The closest cheaper row is Haiku 5.5 Extra High, at 44.3% and $0.56. That is arithmetic on Cursor's table, not a second eval.
NERF High is the default. Medium costs $0.03 more, $0.80 against $0.77, for 46.3% against 45.6%. Insights says a gap that size may not be meaningful. Max takes 187 steps at $0.98. Sonnet 5.5 High takes 41 steps at $1.20. Haiku 5.5 High takes 59 steps at $0.32, in the Claude family. Opus 5.5 Max, the top row, is 57.8% at $13.43. The chart publishes no input, cache, and output split, so this article does not rebuild $0.98 from the token card. The October 18 price-per-task piece priced an illustrative API step. This meter is one finished task on Cursor's harness.
PRICE The weekly pool is a second meter. It covers Grok 4.7, Grok 4.6, Grok 4.5, and Composer 2.5, shared with Grok. DeepSeek on The X Subscription is prepaid credit from the first request. Composer 2.5's list-rate row is 27.7% at $0.68, and Plus can run it. Plus cannot run this DeepSeek row. Source: Models & Pricing (opens in a new tab).
Who pays the dollar
BUFF OP on Premium, Super, or Ultra when that prepaid credit is the budget and a mid-40s score is the job. Those plans skip the $0.25 rate. Low is the $0.60 row. The model stays hidden until you enable it.
NERF Mid on Free and Plus. Third-party models are not on those plans. Composer 2.5 is the included chart row, until the pool runs out.
NERF Mid as a Teams or Enterprise default. The $0.60 to $0.98 figures leave out $0.25 per million input, output, and cached tokens. Haiku 5.5 is excluded from that rate. This row is not.
NERF Mid for an image prompt, a short run, or a direct API call. Cursor's row does not take images. DeepSeek's off-peak API price is half the flat Cursor card, on the hours above. Sonnet 5.5 High is the 41-step row when the harness is the point.
The board stays on October 15
No letter moves. DeepSeek V4.1 Flash is absent from the October 15 Frontier Models list. A vendor chart is not a rank. S stays locked. Cursor stays in A on the October 15 coding-agent list. This is a price inside that harness.
What to do today
- On Premium, Super, or Ultra, turn the row on under Settings, then Models. It draws prepaid credit. Read the monthly cap.
- On Plus, stay on the included pool. This card is not on that plan.
- On Teams or Enterprise, add $0.25 per million tokens before it becomes the default. Haiku 5.5 is the cheap row that list excludes from the rate.
- When the run has to be short, put Sonnet 5.5 High (41 steps, $1.20) beside Max (187 steps, $0.98). Both figures are Cursor's.
The weekly patch is on beehiiv: subscribe free (opens in a new tab).
The OP or Mid Team
Sources
Checked October 28, 2026.