DeepSeek
DeepSeek V4.1 Flash
On 2026-10-17 DeepSeek switched the deepseek-flash endpoint to V4.1-Flash-1017. The pricing page, checked 2026-10-19, still lists peak $0.30/$1.20 and off-peak $0.15/$0.60. Artificial Analysis's 1017 page puts HLE without tools at 43.33 (max). DeepSWE 77.0 and Agents' Last Exam 35.4 are DeepSeek's own 1017 figures.
DeepSeek V4.1 Flash benchmarks and pricing, every number sourced: 9 tracked DeepSeek V4.1 Flash benchmark scores (5 independently run, 4 still resting on a vendor’s own claim), priced at $0.30 per million input tokens and $1.20 per million output.
DeepSeek V4.1 Flash architecture: Mixture-of-Experts (Causal Encoder-Decoder); 552B total parameters (8B prefill / 16B decode activated per token); 1M-token context window.
- Released
- 2026-09-10
- License
- open-weights
- Context window
- 1M tokens
- Knowledge cutoff
- Not disclosed
- Verified
- sources checked 2026-09-10–2026-10-19
- Succeeded
- DeepSeek V4 Flash (0731) (2026-07-31)
- Parameters
- 552B (8B prefill / 16B decode active)
- Architecture
- Mixture-of-Experts (Causal Encoder-Decoder)
DeepSeek V4.1 Flash’s verified record
DeepSeek V4.1 Flash’s most-compared rival is Gemini 3.1 Pro Preview: 1 trail, 1 lead, 1 tie, and 5 not callable — no overall winner across their 8 shared comparisons. DeepSeek V4.1 Flash is priced at $0.30/$1.20 per 1M tokens (in/out) vs Gemini 3.1 Pro Preview’s $2.00/$12.00.
Against the 243 head-to-head comparisons DeepSeek V4.1 Flash shares with other tracked models:
A gap counts for DeepSeek V4.1 Flash only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — DeepSeek V4.1 Flash trails on 30 of them.
HLE · no tools Reasoning
±2 is noiseAnalystAgent Spreadsheet & document analysis
±11.2 is noiseLiveBench Composite score across 7 domains
±2.7 is noiseTerminal-Bench 2.1 Terminal ops
±10.6 is noiseTerminal-Bench 4.0 · high Terminal ops
±12.4 is noiseNo verdict for DeepSeek V4.1 Flash anywhere on Agents' Last Exam, DeepSWE, HLE · with tools ( nothing independently confirmed on both sides); GPQA Diamond ( saturated).
- None of DeepSeek V4.1 Flash’s coding comparisons are independently confirmed on both sides yet.
What changed from DeepSeek V4 Flash (0731) to DeepSeek V4.1 Flash
The 7 benchmarks both models have been scored on, using the same variant each time. A raw DeepSeek V4.1 Flash gain is not a real gain until it clears that benchmark’s own noise band, so each row below carries the verdict and not just the arithmetic.
| Benchmark | DeepSeek V4 Flash (0731) | DeepSeek V4.1 Flash | Change | Verdict |
|---|---|---|---|---|
| HLE | 38.6 | 43.33 | +4.7 | Real gap |
| Terminal-Bench 2.1 | 78.65 | 74.53 | -4.1 | Setup-dependent |
| DeepSWE | 53 | 77 | +24.0 | Unverified |
| Agents' Last Exam | 25.2 | 35.4 | +10.2 | Unverified |
| GPQA Diamond | 89.9 | 90.9 | +1.0 | Tainted |
| LiveBench | 74.2 | 81.1 | +6.9 | Real gap |
| Terminal-Bench 4.0 | 18.69 | 19.7 | +1.0 | Tie |
DeepSeek V4.1 Flash API pricing
$0.30 in / $1.20 out per 1M tokens — official pricing source
What DeepSeek V4.1 Flash costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.042 |
| A codebase review | 1,000K / 100K | $0.420 |
| A day of agent work | 10,000K / 1,000K | $4.20 |
Computed from DeepSeek V4.1 Flash’s list rates above — cache discounts, its off-peak, and batch tiers are not applied.
DeepSeek V4.1 Flash is one of 3 DeepSeek models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| DeepSeek V4.1 Flash | $0.30 | $1.20 |
| DeepSeek V4 Pro (0813) | $1.32 | $3.96 |
| DeepSeek V4 Flash (0731)(previous version) | $0.44 | $1.32 |
DeepSeek V4.1 Flash benchmark scores
| Benchmark | Score |
|---|---|
| GPQA Diamondsaturated[1] Expert science Q&A — not ranked at any gap size | |
| HLE(with tools)[2] Reasoning · ±2 is noise | |
| Terminal-Bench 2.1[3] Terminal ops · ±10.6 is noise | |
| DeepSWE[4] Long-horizon coding · ±9.5 is noise | |
| Agents' Last Exam[5] Professional work · ±3.2 is noise | |
| LiveBench[6] Composite score across 7 domains · ±2.7 is noise | |
| HLE(no tools)[7] Reasoning · ±2 is noise | |
| Terminal-Bench 4.0(high)[8] Terminal ops · ±12.4 is noise | |
| AnalystAgent[9] Spreadsheet & document analysis · ±11.2 is noise |
Who ran these numbers: 5 of 9 independent — vals.ai (2), artificialanalysis.ai (2), livebench.ai (1); vendor self-reported (4).
- GPQA Diamond: Vendor launch-chart claim (Max reasoning effort). GPQA Diamond is graded saturated on this site, so the number carries little weight either way; kept self-reported until an independent run exists — vals.ai's live board (checked 2026-09-10) does not yet list this model.
- HLE: Vendor launch-chart claim (Max reasoning effort, tool use enabled). No independent run exists yet.
- Terminal-Bench 2.1: vals.ai's own Terminus 2 run, board row 'DeepSeek V4.1 Flash' (74.53 ± 1.63, reasoning effort high, pass@1), from its archived Terminal-Bench 2.1 table, dated 2026-10-01, before the 1017 endpoint switch. AA's 1017 object has terminalBench21 null (checked 2026-10-19). Replaces DeepSeek's self-reported 90.6 on its own "DSH Minimal" harness (Hugging Face card), 16.1 points higher; DeepSeek's own scaffold table ranged 84.1 to 90.6 across eight scaffolds.
- DeepSWE: DeepSeek V4.1-Flash-1017 README table (checked 2026-10-19); Mini-SWE-V2 harness. Predecessor self-report on the 09-10 card was 74.2.
- Agents' Last Exam: DeepSeek V4.1-Flash-1017 README table (checked 2026-10-19). Predecessor self-report on the 09-10 card was 31.8.
- LiveBench: Board row 'DeepSeek V4.1 Flash Max Effort' on the 2026-06-25 LiveBench release: 81.1, the only tier populated for this model, observed 2026-10-03, before the 1017 endpoint switch. Not re-read as a 1017-specific row on 2026-10-19. Fourth on the raw board behind Claude Fable 5.1 (83.4), Claude Opus 5.5 (83.2) and GPT-6 Astra (82.2) — the highest LiveBench score among this site's DeepSeek records.
- HLE: AA model page deepseek-v4-1-flash-1017, effort max, checked 2026-10-19. Embedded hle 0.433271547729379 → 43.33. Replaces 39.2 (0.392493 on the pre-1017 slug deepseek-v4-1-flash), which still showed that older figure the same day. The 1017 README's own no-tools figure is 42.6, with 44.0 marked text-only; this cell is AA's number, not the vendor's. The with-tools 63.9 row stays the 09-10 vendor report.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 19.697 at effort high, dated 2026-10-08, before the 1017 endpoint switch. Artificial Analysis's pre-1017 slug still read terminalBench40 0.267677 (26.77) at max on 2026-10-19, a different effort tier from this high row. AA's separate 1017 page (max) reads 0.338384 → 33.84. Different harness from this vals row, so 33.84 stays in this note. The vendor's Claude Code figure is 42.3.
- AnalystAgent: AA's own run, board row 'DeepSeek V4.1 Flash (Max)': analystAgent 0.35 = 35 pass^5. pass@1 is 58.5. Read from the pre-1017 DeepSeek V4.1 Flash (Max) page on 2026-10-15. The 1017 object (deepseek-v4-1-flash-1017, checked 2026-10-19) has analystAgent null, so this 35 is not a 1017 re-run.
Notes on the record
DeepSeek's news post dated 2026/10/17 (api-docs.deepseek.com/news/news261017, checked 2026-10-19) says V4.1-Flash-1017 is live on the website, app, and API. That post does not mention price. The current Your First API Call page says the calling method is unchanged and that deepseek-flash is served by the 1017 weights; it carries no date. The USD pricing page the same day lists model version DeepSeek-V4.1-Flash-1017 at peak $0.30/$1.20 and off-peak $0.15/$0.60 per million tokens, the same schedule this page already stored. Weights are on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash-1017 under MIT. The 1017 card restates a 1M context window in its sampling table. It does not restate the 552B backbone or the 8B/16B activated figures, so those fields stay as previously recorded.
The 1017 README table self-reports HLE without tools at 42.6, with 44.0 marked as the text-only subset, Terminal-Bench 4.0 at 42.3 on a Claude Code harness, DeepSWE v1.1 at 77.0 on Mini-SWE-V2, and Agents' Last Exam at 35.4. This site updates the self-reported DeepSWE and Agents' Last Exam cells to those figures. It does not store 42.6 or 42.3.
Artificial Analysis's deepseek-v4-1-flash-1017 page (max, checked 2026-10-19) embeds hle 0.433272 → 43.33. That replaces the 39.2 still showing on the older deepseek-v4-1-flash slug the same day. The same 1017 object embeds terminalBench40 0.338384 → 33.84. The stored Terminal-Bench 4.0 cell stays vals.ai's 19.7, a different harness. Terminal-Bench 2.1, GPQA, and AA-AnalystAgent are null on the 1017 object.
GPQA Diamond 90.9 and HLE with tools 63.9 remain DeepSeek's 09-10 self-reports. The 1017 table does not republish them. LiveBench 81.1 (2026-10-03), vals.ai Terminal-Bench 2.1 at 74.53 (2026-10-01), vals.ai Terminal-Bench 4.0 at 19.7 (2026-10-08), and AA-AnalystAgent 35.0 (2026-10-15) are earlier measurements, not confirmed 1017 re-runs.
Compare with
FAQ
Did DeepSeek V4.1 Flash change on October 17, 2026?
Yes. DeepSeek's 2026/10/17 news post says V4.1-Flash-1017 is live on the API. This site keeps one model record. The independent HLE cell is now Artificial Analysis's 1017 figure, 43.33. DeepSWE and Agents' Last Exam move to the vendor's 1017 figures. It does not open a second page.
Has DeepSeek V4.1 Flash been independently benchmarked yet?
Five of the nine scores here are independent. Humanity's Last Exam without tools, 43.33, is Artificial Analysis's 1017 max row, checked 2026-10-19. The other four are earlier: vals.ai Terminal-Bench 2.1 at 74.53 (2026-10-01), LiveBench 81.1 (2026-10-03), vals.ai Terminal-Bench 4.0 at 19.7 (2026-10-08), and AA-AnalystAgent 35.0 (2026-10-15). AnalystAgent is null on the 1017 object.
What does DeepSeek V4.1 Flash cost after the 1017 update?
The USD pricing page, checked 2026-10-19, lists DeepSeek-V4.1-Flash-1017 at peak $0.30/$1.20 and off-peak $0.15/$0.60 per million tokens, the same schedule as before. The 2026/10/17 news post does not mention price.
Is DeepSeek V4.1 Flash open source?
Yes. Both the 09-10 and 1017 weight drops on Hugging Face use the MIT license. Hosted API pricing applies only if you call DeepSeek's endpoint rather than running the weights yourself.
Should I trust DeepSeek's 1017 Terminal-Bench 4.0 score of 42.3?
42.3 is DeepSeek's own Claude Code figure. This site's stored independent cell is still vals.ai's 19.7. Artificial Analysis's 1017 max row reads 33.84 on a different harness, so that figure stays in the note rather than replacing 19.7. compareScores() treats a cross-harness pair as setup-dependent.
Which independent DeepSeek V4.1 Flash scores are from the 1017 build?
Only Humanity's Last Exam at 43.33, from Artificial Analysis's 1017 page. LiveBench 81.1, Terminal-Bench 2.1 at 74.53, AA-AnalystAgent 35.0, and vals.ai Terminal-Bench 4.0 at 19.7 were measured earlier. Those five numbers are not one scale, so this page does not rank them against each other.
Further reading
- Models with 10M token context windows 2026 — DeepSeek V4.1 Flash is one of the 49 models it compares.
Benchmark guides
- Humanity's Last Exam — what DeepSeek V4.1 Flash’s reasoning score on it does and doesn’t prove.
- Terminal-Bench 2.1 — what DeepSeek V4.1 Flash’s agentic score on it does and doesn’t prove.