DeepSeek Flash update claims gains on 11 benchmarks, same price

DeepSeek on Saturday, Oct. 17, released DeepSeek-V4.1-Flash-1017, an update to the Flash model it launched on Sept. 10, with MIT-licensed weights on Hugging Face and no change to its API prices. DeepSeek’s own table shows higher scores on all 11 benchmarks it reports, from 2.0 points on ZeroBench to 11.1 on Terminal-Bench 4.0, and the company says the model now finishes tasks faster, with shorter outputs.

Every score here is DeepSeek’s, produced on test harnesses it chose, and none has been independently reproduced. What can be checked against the primary documents is the cost: a million uncached input tokens is $0.15 off-peak, a full-length session needs under a gigabyte of cache in GPU memory, and the weights come to 510 GB. TECHi worked through those numbers.

Article Brief

Key Takeaways

5 Points30s Read

• 01What shipped–An update to V4.1-Flash with open MIT-licensed weights. DeepSeek says its web app and the deepseek-flash API now run it, and the price is unchanged.
• 02The claimed gains–All 11 reported benchmarks rose. Terminal-Bench 4.0 gained 11.1 points (36%) and ProgramBench 8.8 points (43%). The scores are DeepSeek’s own and unverified.
• 03Against rivals–In DeepSeek’s table the new model beats Kimi K3 on 10 of 11 tests and GPT-5.6 Sol on nine, but trails Opus 5 on seven.
• 04The cost–A full 1M-token prompt costs $0.15 off-peak and $0.30 at peak. Flash input costs about one-fourth of DeepSeek V4 Pro, and output about one-third.
• 05The memory–At DeepSeek’s stated 890 bytes per token, a 1M-token context takes about 0.9 GB of cache in GPU memory. The weights are the bigger bill, at 510 GB.

What did DeepSeek release on Oct. 17?

The update is a post-training refresh of V4.1-Flash, not a new architecture. DeepSeek’s API change log says the web and app versions moved to the new model, and the deepseek-flash API name now serves it with pricing unchanged. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the models behind them are retired, and requests to them are billed at the Flash price. TECHi has covered an earlier round of DeepSeek model-name retirements.

The model card lists the weights under the MIT license in 48 files that total 510.3 GB. The architecture comes from the technical report in the base model’s repository, published with the September release: a multimodal mixture-of-experts model with 552 billion backbone parameters that activates 16 billion per token when generating text and 8 billion when reading a prompt, with a one-million-token context window.

How much did DeepSeek’s benchmark scores improve?

DeepSeek’s card compares the new model with the September version across 11 tests, and every score went up. The table gives both scores and the change.

DeepSeek V4.1-Flash, September version against the Oct. 17 update

Benchmark	V4.1-Flash (Sept. 10)	V4.1-Flash-1017 (Oct. 17)	Change
Terminal-Bench 4.0	31.2	42.3	+11.1 (+36%)
ProgramBench	20.3	29.1	+8.8 (+43%)
Apex Bench v1.1	64.6	72.1	+7.5 (+12%)
HLE without tools (text-only subset in brackets)	36.8 (39.1)	42.6 (44.0)	+5.8 (+16%)
Automation-Bench v1.0.6	54.8	60.5	+5.7 (+10%)
Chartography (with tools)	78.9	83.0	+4.1 (+5%)
Agents’ Last Exam	31.8	35.4	+3.6 (+11%)
BabyVision (with tools)	89.6	92.5	+2.9 (+3%)
DeepSWE v1.1	74.2	77.0	+2.8 (+4%)
NL2Repo-Bench	65.4	68.2	+2.8 (+4%)
ZeroBench-main (with tools)	49.0	51.0	+2.0 (+4%)

The largest gains are in agentic coding and automation: Terminal-Bench 4.0 rose 11.1 points from 31.2, and ProgramBench 8.8 points from 20.3. The smallest are on the image tests and on DeepSWE. DeepSeek describes the update as improving reasoning, coding, agentic and visual results while cutting task time and output length, which would lower a bill even at an unchanged price per token. The card does not publish the task-time or output-length figures, so TECHi cannot size that saving.

Harness choices matter. DeepSeek says it ran Terminal-Bench 4.0 in the Claude Code harness and DeepSWE in the Mini-SWE-V2 harness, and that ZeroBench allowed five attempts (Pass@5). A different harness or attempt limit can move a score, and several of these gains are only two to three points.

The card also lines the new model up against four rivals. In that table it scores higher than Kimi K3 on 10 of 11 tests, higher than GLM-5.3 on all eight where GLM reports a score, and higher than GPT-5.6 Sol on nine of 11. It trails Opus 5 on seven of 11, with the widest gaps on HLE (42.6 against 56.3) and Terminal-Bench 4.0 (42.3 against 51.8). These are DeepSeek’s chosen comparisons and settings, so they are claims to test, not a ranking.

What does DeepSeek Flash cost to run?

DeepSeek’s pricing page lists the deepseek-flash model at $0.15 per million input tokens that miss its cache and $0.003 per million that hit it, plus $0.60 per million output tokens, all at the off-peak rate. Peak rates are double. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays, so weekends are off-peak all day. Maximum output is 384,000 tokens.

What deepseek-flash costs, at DeepSeek’s published rates

Scenario	Calculation	Off-peak	Peak
One 1M-token prompt, no cache hit	1M × $0.15 off-peak or $0.30 peak	$0.15	$0.30
The same prompt re-read from cache	1M × $0.003 or $0.006	$0.003	$0.006
100,000 tokens of output	0.1M × $0.60 or $1.20	$0.06	$0.12
Hypothetical agent run: 2M input tokens with 80% cache hits, 100,000 output tokens	1.6M × cache-hit rate + 0.4M × cache-miss rate + 0.1M × output rate	$0.12	$0.25

Against DeepSeek’s own larger model, deepseek-v4-pro (version 0813), Flash is cheaper by a wide margin. The Pro off-peak rates are $0.66 for uncached input, $0.022 for cached input and $1.98 for output, so Pro charges 4.4 times as much as Flash for uncached input, 7.3 times as much for cached input and 3.3 times as much for output. The two models differ in capability, and Pro is a text-only model on DeepSeek’s feature table while Flash accepts images.

Why is a one-million-token window cheap to serve?

When a model reads a long prompt it stores intermediate results for every token, called the KV cache, and at long lengths that cache rather than the weights can limit how many sessions one machine can serve. DeepSeek’s report says V4.1-Flash cuts the part that must sit in GPU memory to 890 bytes per token, about a quarter of DeepSeek-V4-Flash and, by the report’s figure caption, 437 times less than DeepSeek-V1.

TECHi’s arithmetic from those figures: at the model’s full 1,048,576-token window, 890 bytes per token is about 0.93 GB. DeepSeek-V4-Flash, at about four times the footprint, would take about 3.7 GB, and DeepSeek-V1, at 437 times, roughly 408 GB. The ratios are DeepSeek’s approximate figures, so these are orders of magnitude, not measurements. The weights are the larger cost: 510.3 GB on disk, which would fit with about 130 GB to spare on eight 80 GB accelerators, though the report states no hardware requirement and real deployments need room for activations and a persistent cache.

What would change the read

The strongest case for taking the numbers seriously is that the weights are public, so outside labs can rerun the tests. The case for caution is DeepSeek’s own. Its report says the new design creates robustness limits that are not yet fully characterized, and that selection errors in its sparse attention and approximate state reconstruction in its cache-replay system may degrade capability in untested edge cases. It also says internal tests found no systematic degradation.

TECHi would update on three results: independent runs of Terminal-Bench 4.0 and ProgramBench on the Oct. 17 weights that land near DeepSeek’s scores, long-context retrieval that holds up at the full million tokens, and the Opus 5 gaps narrowing under shared harness settings. Until they appear, the checkable facts are the license, the file sizes and the price list, and the benchmark gains remain DeepSeek’s claims.