← All model releases

Model release · DeepSeek

DeepSeek V4.1-Flash 0910: $0.30 / $1.20 at peak, and a cache read at 2% of input

DeepSeek V4.1-Flash 0910 — the 2026-09-10 launch checkpoint — costs $0.30 per 1M input and $1.20 per 1M output at peak, $0.15 / $0.60 off-peak, and reads cached input at $0.006, with 19 verified benchmark scores. The API now serves the newer 1017 checkpoint, tracked separately.

Published 2026-10-19Released 2026-09-10Prices checked 2026-10-20

What it costs

USD per 1M tokens, read live from the provider pricing page — so this table can never go stale.

ModelInputOutputCachedContextOutput vs subject
DeepSeek V4.1-Flash 0910(Legacy)

Launch checkpoint of 2026-09-10 — the one the launch announcement's reference table was measured on. Superseded by the 1017 checkpoint on 2026-10-17; the API no longer serves it (`deepseek-flash` now routes to 1017). Rates are identical to the live checkpoint; kept for long-tail search and for the launch post's score table.

Source · verified 2026-10-20		$0.30$1.20$0.0061M—
DeepSeek V4-ProSource · verified 2026-10-20$1.32$3.96$0.0441M×0.3

What your usage would cost

1M input + 500K output tokens per month, at list price.

releases.costPeak

$0.90/ month

$0.30 input + $0.60 output

Off-peak hours

$0.45/ month

The same workload scheduled outside this provider’s peak window.

How it scores

Transcribed from the sources named on each leaderboard entry. Coverage differs per model; a missing cell means the source did not report that benchmark.

BenchmarkDeepSeek V4.1-Flash 0910GPT 5.6-SolClaude Opus 5Gemini 3.1 ProKimi K3GLM 5.3
GPQA Diamond90.9%94.1%93.4%94.3%92.9%88.1%
HLE(39.1 on text-only subset)36.8%44.5%56.3%44.4%43.5%42%
Codeforces (Rating)3471—————
MathArena Apex65.6%——60.9%65.6%—
Terminal-Bench 2.190.6%88.8%89.1%—88.3%88.2%
Terminal-Bench 3.030%34.4%43.3%—17.7%28.3%
Terminal-Bench 4.031.2%39.9%51.8%—12.6%37.9%
DeepSWE v1.174.2%73%74%—67.5%66.9%
ProgramBench20.3%23%37%—17.5%19%
NL2Repo-Bench65.4%56.8%75.3%—58%58%
CyberGym88.1%84.5%——80%84.5%
SEC-Bench Pro62.8%74.3%————
ExploitGym15.3%33.7%22.1%——15%
HLE (w/tools)63.9%—63.6%51.4%59.8%62.5%
Automation-Bench54.8%45.8%50.3%35.4%46.7%48.8%
Agents' Last Exam31.8%26.7%28.6%—27.6%28.5%
Chartography (w/tools)78.9%79.9%84%—68.1%—
BabyVision (w/tools)89.6%88.9%94.1%—85.7%—
ZeroBench-main (w/tools)49%53%52%—41%—

TL;DR

  • DeepSeek V4.1-Flash 0910 is $0.30 per 1M input and $1.20 per 1M output at peak, and $0.15 / $0.60 off-peak — exactly half. Released on 2026-09-10.
  • The cache read is the standout: $0.006 per 1M at peak, $0.003 off-peak — 2% of the peak input rate. A long, repeatedly re-read context costs almost nothing to re-read.
  • One of the cheapest models we track, with the widest score coverage we publish — 19 benchmarks. All 19 come from a single provenance lane: the launch announcement's reference table, transcribed by us. The API has since moved on to the 1017 checkpoint at the same price; that checkpoint is tracked separately, with its own table and its own post.

What shipped

  • A new Flash SKU with a two-tier daily clock. Peak rates are $0.30 input / $1.20 output per 1M; off-peak rates are $0.15 / $0.60, exactly half. Cached input is $0.006 at peak and $0.003 off-peak.
  • The published schedule, exactly as it stands. Mon–Fri (excluding Chinese public holidays), peak runs 01:00–04:00 and 06:00–10:00 UTC; off-peak is all other times. Weekends and Chinese public holidays are off-peak in full.
  • A 1,000,000-token context window with no long-context surcharge tier on file.
  • Released on 2026-09-10, from the launch announcement. This post covers that checkpoint — the 0910 one — because the scores below belong to it.
  • A sibling that costs several times more. DeepSeek V4-Pro is $1.32 / $3.96 at peak, and $0.66 / $1.98 off-peak. Both models are active.

What it costs

  • The hour you call at is worth a factor of two. Peak is $0.30 / $1.20, off-peak is $0.15 / $0.60. The published peak window is 7 hours a day (01:00–04:00 and 06:00–10:00 UTC) on weekdays only — about a fifth of the week. Everything else is half price.
  • So the schedule is part of the pricing model. Batch jobs, nightly summarisation, evaluation sweeps and any backfill: if they run outside those two UTC windows, they run at half price. On a Chinese public holiday the whole day is off-peak.
  • The cache read at $0.006 is 2% of the peak input rate — a 98% discount. In an agent loop that re-reads the same context every turn, this is the line that decides the bill.
  • Against the sibling model, the gap is 3–4×. V4-Pro is $1.32 / $3.96 at peak: 4.4× the input and 3.3× the output of V4.1-Flash. Both models halve off-peak ($0.66 / $1.98 for Pro), so the ratio holds in either period.
  • There is no long-context step to plan around. The 1,000,000-token window is billed at the same rate as a short prompt, according to the source we hold.

How it scores

  • This table describes the 0910 checkpoint, not the weights the API serves today. The API moved to the 1017 checkpoint on 2026-10-17, at the same price, and that checkpoint publishes its own, different score table — tracked separately here: /releases/deepseek-v4-1-flash-1017. Two checkpoints, two tables, never merged: the launch table measured the launch weights, and reading it as the current ones is exactly how a benchmark table gets misused.
  • Nineteen verified scores — the widest coverage of any model on our board. But they all come from one place: the reference table in the DeepSeek V4.1 Flash launch announcement, captured on 2026-10-19. That makes them internally consistent and traceable, and it makes them a vendor claim rather than an independent re-run. We print the caveat with the numbers.
  • The notes travel with the figures. HLE reads 36.8, and the source notes 39.1 on the text-only subset — so a model that only reports the text-only variant is not on the same footing. Codeforces is a rating, not a percentage, and should not be averaged with the rest.
  • Read the spread, not an average. The 19 rows split into two groups. Knowledge and math sit high: GPQA Diamond 90.9, a Codeforces rating of 3471, MathArena Apex 65.6. The agentic and coding rows range much more widely: terminal-bench 2.1 at 90.6, but terminal-bench 4.0 at 31.2. The same family of tasks gets harder between versions 2.1, 3.0 and 4.0, and the score follows it down — that is the shape of the card, not a weakness in one column.
  • What the comparison rows show. Against the flagship rows on the table — GPT-5.6 sol, Claude Opus 5, Gemini 3.1 Pro — V4.1-Flash leads the terminal-bench 2.1 column, but on terminal-bench 3.0 and 4.0 it is mid-pack rather than the leader. Against the Chinese-model rows it leads on terminal-bench 2.1 and 3.0, while GLM 5.3 is ahead of it on terminal-bench 4.0. Read the empty cells as "not reported", never as zero.
  • Coverage caveat: the reference table does not cover every benchmark we track, so a blank cell is missing data, not a bad result.

Should you switch?

  • If you are on a frontier flagship, the price gap is 10–20×, so the question is not "is it as good" but "does the quality gap matter on this traffic". A split usually beats a switch: keep the flagship for the calls where a wrong answer is expensive, and move the volume to V4.1-Flash.
  • If you can schedule work, schedule it off-peak. The off-peak rate halves everything. Moving a nightly job out of the two peak UTC windows is the cheapest optimisation available here, and it needs no code change beyond when the job starts.
  • If your workload re-reads the same context, $0.006 per 1M cached makes the repeat reads almost free next to the rest of the table. That changes the design: long, stable context becomes viable instead of a liability.
  • If you need V4-Pro's quality, note that Pro is 4.4× the input and 3.3× the output — the price gap is the smaller half of that decision, and the score table is the larger one.
  • If you are starting a new integration, note that the API now serves the 1017 checkpoint, not this one. Same price, better scores — read that post before you pin anything.
  • What would change our recommendation: an independent re-run of the 19 rows. Right now they are one announcement's table, transcribed by us. That is a legitimate source, and it is not the same thing as a third-party measurement.

Where to go next

Run the numbers yourself

The tables above are the same data the calculator and the comparison tool use.

Compare it head to headPrice your own usageSee the full model page

Source

DeepSeek V4.1 Flash 发布公告

Source re-checked on 2026-10-20. · DeepSeek V4.1-Flash 0910 pricing comes from the provider page linked in the table above.

More release analyses

DeepSeek V4.1-Flash 1017: the same $0.30 / $1.20 price, a much better checkpointGemini 4 Argon: $2 / $10 per 1M until November 30, then $4 / $20Claude Opus 5.5: Anthropic's flagship at $4 / $20, 20% below Opus 5