Model comparison
DeepSeek V4.1 Flash vs DeepSeek V4 Pro
Two DeepSeek tiers compared on the figures that decide which one a workload actually needs.
DeepSeek V4.1 Flash: profile checked October 25, 2026DeepSeek V4 Pro: profile checked October 25, 2026Individual provider fields may changeEvidence confidence: Insufficient data — see receipts below
DeepSeek
DeepSeek V4.1 Flash
Balanced · Open weights
DeepSeek
DeepSeek V4 Pro
Frontier · Open weights
| Specification | DeepSeek V4.1 Flash | DeepSeek V4 Pro |
|---|---|---|
| Provider | ||
| Provider | DeepSeek | DeepSeek |
| Tier | ||
| Tier | Balanced | Frontier |
| Context window | ||
| Context window | 1M | 1M |
| Max output | ||
| Max output | 384K | 384K |
| Input / 1M tokens | ||
| Input / 1M tokens | Not verifiedUnverified | $1.32 |
| Output / 1M tokens | ||
| Output / 1M tokens | Not verifiedUnverified | $3.96 |
| Weights | ||
| Weights | Open | Open |
| Parameters | ||
| Parameters | 552B MoE backbone + 196B Engram memory (causal encoder-decoder; 8B active prefill, 16B decode) | 1.6T total / 49B active (MoE) |
| Reasoning levels | ||
| Reasoning levels | low, high, max | low, high, max |
| Modalities | ||
| Modalities | text, image | text |
| License | ||
| License | MIT | MIT |
| API model id | ||
| API model id | Not verifiedUnverified | deepseek-v4-pro |
| Released | ||
| Released | September 10, 2026 | April 24, 2026 |
| Artificial Analysis Intelligence Index [max] · source date DeepSeek V4.1 Flash: 2026-10-01 | ||
| Artificial Analysis Intelligence Index [max] · source date DeepSeek V4.1 Flash: 2026-10-01 | 39 | Not verifiedUnverified |
| Artificial Analysis Intelligence Index [max] · source date DeepSeek V4 Pro: 2026-10-01 | ||
| Artificial Analysis Intelligence Index [max] · source date DeepSeek V4 Pro: 2026-10-01 | Not verifiedUnverified | 36 |
| GPQA Diamond [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | ||
| GPQA Diamond [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | 90.9 | Not verifiedUnverified |
| GPQA Diamond [max] · source date DeepSeek V4 Pro: 2026-04-24 | ||
| GPQA Diamond [max] · source date DeepSeek V4 Pro: 2026-04-24 | Not verifiedUnverified | 90.1 |
| DeepSWE 1.1 [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | ||
| DeepSWE 1.1 [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | 74.2 | Not verifiedUnverified |
| DeepSWE 1.1 · source date DeepSeek V4 Pro: 2026-08-13 | ||
| DeepSWE 1.1 · source date DeepSeek V4 Pro: 2026-08-13 | Not verifiedUnverified | 62.7 |
| Terminal-Bench 2.1 [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | ||
| Terminal-Bench 2.1 [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | 90.6 | Not verifiedUnverified |
| Terminal-Bench 2.1 · source date DeepSeek V4 Pro: 2026-08-13 | ||
| Terminal-Bench 2.1 · source date DeepSeek V4 Pro: 2026-08-13 | Not verifiedUnverified | 87.9 |
| Terminal-Bench 3.0 [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | ||
| Terminal-Bench 3.0 [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | 30 | Not verifiedUnverified |
| Terminal-Bench 4.0 [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | ||
| Terminal-Bench 4.0 [max] · source date DeepSeek V4.1 Flash: 2026-10-02 | 31.2 | Not verifiedUnverified |
| SWE-bench Verified [max] · source date DeepSeek V4 Pro: 2026-04-24 | ||
| SWE-bench Verified [max] · source date DeepSeek V4 Pro: 2026-04-24 | Not verifiedUnverified | 80.6 |
| Humanity's Last Exam [max] · source date DeepSeek V4 Pro: 2026-04-24 | ||
| Humanity's Last Exam [max] · source date DeepSeek V4 Pro: 2026-04-24 | Not verifiedUnverified | 48.2 |
| Humanity's Last Exam · source date DeepSeek V4 Pro: 2026-08-13 | ||
| Humanity's Last Exam · source date DeepSeek V4 Pro: 2026-08-13 | Not verifiedUnverified | 42.7 |
| Humanity's Last Exam · source date DeepSeek V4 Pro: 2026-08-13 | ||
| Humanity's Last Exam · source date DeepSeek V4 Pro: 2026-08-13 | Not verifiedUnverified | 60 |
| MMLU-Pro [max] · source date DeepSeek V4 Pro: 2026-04-24 | ||
| MMLU-Pro [max] · source date DeepSeek V4 Pro: 2026-04-24 | Not verifiedUnverified | 87.5 |
Dates beside benchmark scores refer to the linked source’s publication or capture date unless that source explicitly states when it measured the result. A shared URL alone does not establish comparable runs; unverified setups are marked. Prices are USD per million tokens at standard rates, excluding batch and caching discounts. Values we could not confirm from the provider are shown as “Not verified” rather than estimated.
Benchmark receipts
Where each score comes from, and how far it can be compared across models.
- Source date2026-10-02DeepSeek V4.1-Flash native model card (vendor-reported; October 2 source capture, not experiment date) (DeepSeek Harness Minimal; Max effort; 1M context; temperature=1; top_p=0.95; sample count/task manifest and experiment date unverified)
Agentic terminal work: long multi-step tasks executed in a sandboxed shell environment.
Comparable with caveatDo not directly compare scores across Terminal-Bench versions: 4.0 revises and prunes the 3.0 task set and changes environment resources, requiring fresh trials. Within 4.0, harness, time limit, tools, sampling and resource settings must match; maintainer-agent and Artificial Analysis runs are not interchangeable.
- Source date2026-10-02DeepSeek V4.1-Flash native model card (vendor-reported; October 2 source capture, not experiment date) (DeepSeek Harness Minimal; Max effort; 1M context; temperature=1; top_p=0.95; sample count/task manifest and experiment date unverified)
Agentic terminal work across seven domains, including coding and non-coding deliverables.
UnknownDo not mix versions or harness configurations. The exact task manifest, sample count and experiment date of the retained DeepSeek observation are unverified, so it remains outside cross-model comparisons.
- Source date2026-10-02DeepSeek V4.1-Flash native model card (vendor-reported; October 2 source capture, not experiment date) (DeepSeek Harness Minimal; Max effort; N=3 samples/task; Linux containers; no network; 1M context; temperature=1; top_p=0.95; max_steps=500; experiment date unverified)
Agentic terminal work: multi-step tasks executed in a sandboxed shell environment.
Comparable with caveatNot comparable with Terminal-Bench 3.0 or 4.0 (different task sets) or v1; harness configuration (container, time limits) also shifts results.
- Source date2026-10-02DeepSeek V4.1-Flash native model card (vendor-reported; October 2 source capture, not experiment date) (mini-SWE; Max effort; N=8 samples/task; 1M context; temperature=1; top_p=0.95; max_steps=500; experiment date unverified)
Agentic software engineering: pass@1 on real repo tasks with reported cost per completed task.
Directly comparable
- Source date2026-10-02DeepSeek V4.1-Flash native model card (vendor-reported; October 2 source capture, not experiment date) (Max effort; reasoning evaluation; complete run settings and experiment date unverified)
Graduate-level science reasoning: multiple-choice questions written by domain PhDs.
Comparable with caveatDo not treat a fixed two-point band as universal measurement uncertainty; it depends on run count and evaluation setup. Chain-of-thought, prompt, and tool policy must match. The small pool limits fine-grained conclusions, and frontier scores can be near ceiling; Artificial Analysis dropped it from its Intelligence Index in v4.3.
- Source date2026-10-01Artificial Analysis (DeepSeek V4 Pro 0813 (max))
Composite index blending reasoning, knowledge, and coding evals into one 0–100 score.
Comparable with caveatOnly same-version scores are comparable. v4.3.2 was rebaselined (the top score fell from 66 on v4.1.1 to ~58), so v4.1.1 figures must not be compared with v4.3.2 figures — check the cited source and its date.
- Source date2026-10-01Artificial Analysis (DeepSeek V4.1 Flash (max))
Composite index blending reasoning, knowledge, and coding evals into one 0–100 score.
Comparable with caveatOnly same-version scores are comparable. v4.3.2 was rebaselined (the top score fell from 66 on v4.1.1 to ~58), so v4.1.1 figures must not be compared with v4.3.2 figures — check the cited source and its date.
- Source date2026-08-13DeepSeek August 13 GA changelog (vendor) (0813 GA refresh; vendor report, exact run configuration unverified)
Agentic software engineering: pass@1 on real repo tasks with reported cost per completed task.
Directly comparable
- Source date2026-08-13DeepSeek August 13 GA changelog (vendor) (0813 GA refresh; vendor report, exact run configuration unverified)
Agentic terminal work: multi-step tasks executed in a sandboxed shell environment.
Comparable with caveatNot comparable with Terminal-Bench 3.0 or 4.0 (different task sets) or v1; harness configuration (container, time limits) also shifts results.
- Source date2026-08-13DeepSeek August 13 GA changelog (vendor) (0813 GA refresh; with tools; vendor report, exact run configuration unverified)
Frontier-knowledge ceiling: extremely hard multi-domain questions written to be near-impossible without deep expertise.
Comparable with caveatSubset/tool configurations (text-only vs with-tools) differ between vendors and materially change scores; must match configuration to compare.
- Source date2026-08-13DeepSeek August 13 GA changelog (vendor) (0813 GA refresh; without tools; vendor report, exact run configuration unverified)
Frontier-knowledge ceiling: extremely hard multi-domain questions written to be near-impossible without deep expertise.
Comparable with caveatSubset/tool configurations (text-only vs with-tools) differ between vendors and materially change scores; must match configuration to compare.
- Source date2026-04-24DeepSeek V4 Pro HF model card (vendor, Think Max, exact match) (Original April 24 checkpoint; historical vendor evaluation, not the served 0813 refresh)
Broad knowledge across 14 domains with harder, reasoning-focused multiple-choice questions.
Comparable with caveatFrontier models cluster in the high 80s–90s; near-ceiling differences are weak evidence. CoT vs direct answering must match.
- Source date2026-04-24DeepSeek V4 Pro HF model card (vendor, Think Max, with tools; 37.7 no tools) (Original April 24 checkpoint; historical vendor evaluation, not the served 0813 refresh)
Frontier-knowledge ceiling: extremely hard multi-domain questions written to be near-impossible without deep expertise.
Comparable with caveatSubset/tool configurations (text-only vs with-tools) differ between vendors and materially change scores; must match configuration to compare.
- Source date2026-04-24DeepSeek V4 Pro HF model card (vendor, Think Max, pass@1) (Original April 24 checkpoint; historical vendor evaluation, not the served 0813 refresh)
Graduate-level science reasoning: multiple-choice questions written by domain PhDs.
Comparable with caveatDo not treat a fixed two-point band as universal measurement uncertainty; it depends on run count and evaluation setup. Chain-of-thought, prompt, and tool policy must match. The small pool limits fine-grained conclusions, and frontier scores can be near ceiling; Artificial Analysis dropped it from its Intelligence Index in v4.3.
- Source date2026-04-24DeepSeek original V4 Pro HF card (vendor, Think Max; verifier/harness comparability unverified) (Original April 24 checkpoint; historical vendor evaluation, not the served 0813 refresh)
Real GitHub issue resolution: does the model's patch pass the hidden tests.
Comparable with caveatPost-audit vendor claims and pre-audit scores sit on different task trust levels; scaffolding (agent harness, compute budget) also dominates results. Never aggregate across scaffolds.
Pricing tiers
DeepSeek V4.1 Flash: The original September checkpoint is no longer separately sold through the rolling first-party alias. Its MIT weights remain available for self-hosting; no current independent hosting tariff is verified here. Current rolling deepseek-flash pricing belongs to DeepSeek-V4.1-Flash-1017, not a new evaluation of these original weights.
DeepSeek V4 Pro: Current hosted version is DeepSeek-V4-Pro-0813. USD/MTok peak input/output $1.32/$3.96, off-peak $0.66/$1.98; cached input $0.044 peak or $0.022 off-peak. Peak hours: weekdays 01:00–04:00 and 06:00–10:00 UTC excluding Chinese public holidays; all remaining hours, weekends and holidays are off-peak. Search and Fetch each cost $0.75/1,000 successful requests. Server-side calls through DeepSeek Harness, Claude Code or Codex are free until October 31, 2026 at 16:00 UTC (November 1 at 00:00 Beijing time), then billed; standalone API prices are separate. The original 1.6T/49B MIT checkpoint and April benchmark observations are separate from the hosted 0813 revision.
DeepSeek V4.1 Flash
The September 2026 MIT-licensed DeepSeek V4.1 Flash checkpoint, retained separately from the October 17 API and weight update.
Best for
- Open-weight price-performance
- High-volume hosted agents
- Self-hosting with MIT weights
Watch out
Vendor task scores and the Artificial Analysis composite index measure different things. API low/high/max effort is separate from the checkpoint's 1–100 control; non-thinking API mode is also supported. Peak/off-peak billing follows time and Chinese public holidays. Responses and Anthropic-compatible APIs are supported; FIM is non-thinking only. Native prompt encoding maps low/high/max to 50/75/100 and defaults to high=75 in thinking mode; this does not establish the hosted API default. The native 552B backbone excludes 196B Engram conditional memory. Published minimal inference is a readable reference, not verified production serving; its uninitialized self-test is not real-model validation. The 384K output cap is a hosted API limit, not a confirmed native generation ceiling. This is the September checkpoint profile: hosted limits are historical reference values, while current rolling-API calls and prices belong to Flash-1017. No successor benchmark result is inherited.
DeepSeek V4 Pro
DeepSeek V4 Pro is DeepSeek's large open-weight (MIT) MoE with a 1M-token context — DeepSeek says the cheaper V4.1-Flash now beats it.
Best for
- Cost-sensitive hosted agents
- Open-weight deployments
- High-volume coding
Watch out
DeepSeek reversed the earlier September 14 reroute plan and continues serving V4 Pro until further notice. The tariff names V4-Pro-0813; original April checkpoint benchmarks do not evaluate that hosted revision. The API supports thinking/non-thinking, low/high/max effort, Responses and Anthropic-compatible routes; FIM is non-thinking only. Self-hosting requirements depend on the exact weights, quantization and serving stack.
When the cheaper one wins
One of these rows is missing a verified output price, so this page will not name a cheaper winner. Run the model picker against the actual job instead of guessing.
Run the model pickerEvidence confidence: Insufficient data
How strong and complete the evidence behind this comparison is — not a prediction of which model is better. Hard limitations apply: missing pricing; evidence too sparse.
- Pricing incomplete — No verified token pricing for DeepSeek V4.1 Flash.
- 5/5 core specs verified on both sides — All core specifications verified for both models.
- No directly aligned benchmark setup — No benchmark has matching cited sources, source dates, units, effort and recorded run details on both models.
- Verified within the last 90 days — Older of the two catalog checks was 24 days ago.
- Both models carry source citations — Each side has at least two catalog sources on record.
Source receipts
Each catalog figure was checked against the provider or an independent second source on the date shown.
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4.1 Flash: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4 Pro: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4 Pro: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4 Pro: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4 Pro: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
- DeepSeek V4 Pro: Whole-record check: current publisher specifications, licence, access or tariff (accessed 2026-10-25)
Related comparisons
- Amazon Nova 2 Lite vs DeepSeek V4.1 Flash
- Amazon Nova 2 Pro vs DeepSeek V4.1 Flash
- Base 1 vs DeepSeek V4.1 Flash
- Claude Fable 5.1 vs DeepSeek V4 Pro
- Claude Fable 5 vs DeepSeek V4 Pro
- Claude Mythos 5.1 vs DeepSeek V4 Pro
Diving deeper on one model? DeepSeek V4.1 Flash · DeepSeek V4 Pro
Common questions
DeepSeek V4.1 Flash vs DeepSeek V4 Pro
Answered from the verified figures on this page rather than general guidance.
Which has the larger context window, DeepSeek V4.1 Flash or DeepSeek V4 Pro?
Do DeepSeek V4.1 Flash and DeepSeek V4 Pro support the same reasoning levels?
Should I use DeepSeek V4.1 Flash or DeepSeek V4 Pro?
Next step
Choosing between them
The verified figures that separate this pair, computed from the catalog rather than restated boilerplate.
- Positioning: DeepSeek V4.1 Flash sits in the balanced tier, DeepSeek V4 Pro in the frontier tier. Compare accuracy, reliability, latency, and cost on representative tasks before choosing.
Choosing a harness rather than a model — Cursor, Copilot, Claude Code, Windsurf, or Lovable? Compare agentic harnesses · Latest releases.