DeepSeek V4.1 Flash vs V4 Flash: Beats Opus 5 on DeepSWE, 4.3x the Output Price

DeepSeek V4.1 Flash 1017 came out on October 17, got more than 860 points on Hacker News, and is being called the cheap model that beats Opus 5 on DeepSWE. Before you move your agent loop off V4 Flash, look at the output price: $1.20 per million tokens at peak. V4 Flash 0731 charged $0.28 in July.

That's 4.3x on output. On per-token list price, Claude Haiku 5.5 and GPT-6 Luna are now both cheaper than DeepSeek's own API.

DeepSeek says 1017 finishes tasks in fewer tokens, so the bill comes out lower anyway. Sometimes that's true. For long-context agents it's the cheapest option I can find, and for short tasks it loses. Below are the benchmarks, the per-token prices, the cost per task and a sample long-context session, pulled from DeepSeek's docs, the Hugging Face card, OpenRouter, Artificial Analysis and the HN thread. I haven't run any benchmarks myself, so every number links to its source.

What V4.1 Flash 1017 actually is

It's the V4.1 Flash DeepSeek shipped in early September with refreshed weights. The model card calls it a "minor update" with gains in reasoning, coding, agentic and visual tasks, and says it "reduces task execution time and output length." DeepSeek's release note calls it "a Pareto improvement in performance and cost."

The specs, according to DeepSeek's pricing page and the card:

API model name deepseek-flash. The old deepseek-v4-flash and deepseek-v4-flash-vision-exp names now get served by 1017 and billed at the 1017 price. If you pinned the old name thinking you'd stay on old pricing, you didn't.
1M context, up to 384K output tokens
Thinking on by default, non-thinking available. Tool calls, JSON output, the Responses API, an Anthropic-format endpoint and vision are all supported.
MIT license for the weights. The HF card lists 763B parameters. VentureBeat reported the September V4.1 as 552B backbone plus 196B of Engram memory, with 16B active during decode. V4 Flash was 284B total and 13B active.

DeepSeek V4.1 Flash price per token vs V4 Flash

Official DeepSeek API rates, USD per 1M tokens:

| | V4 Flash 0731 | V4.1 Flash 1017 off-peak | V4.1 Flash 1017 peak |
|---|---|---|---|
| Input, cache hit | $0.0028 | $0.003 | $0.006 |
| Input, cache miss | $0.14 | $0.15 | $0.30 |
| Output | $0.28 | $0.60 | $1.20 |

The V4 Flash figures come from PureAI's 0731 writeup. Peak is 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. That covers the European morning and the late US evening, so plenty of teams will pay peak rates without noticing. Off-peak, the output hike is 2.1x instead of 4.3x.

Because the weights are open, third-party hosts undercut DeepSeek. OpenRouter's endpoint list shows 13 endpoints:

DeepInfra's fp8 endpoint is $0.098 input / $0.294 output, with a 131K max output. Its "flex" tier goes down to $0.078 / $0.235, below what V4 Flash cost on DeepSeek's own API.
Cloudflare charges $0.10 input, $0.60 output and $0.003 for cache reads, with no output cap worth worrying about.
Fireworks, Parasail, BaseTen, Wafer and PrimeIntellect charge DeepSeek's peak rate around the clock. Wafer serves an mxfp4 quant.

DeepSeek V4.1 Flash benchmarks: 1017 vs September, GPT-5.6 Sol and Opus 5

These are DeepSeek's own numbers from the model card. Treat them as marketing until someone independent reruns them.

| Benchmark | V4.1 Flash 1017 | V4.1 Flash (Sept) | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 45.8 | 50.3 |
| HLE w/o tools | 42.6 | 36.8 | 44.5 | 56.3 |

An 11-point jump on Terminal-Bench 4.0 from a "minor update" is big. DeepSeek ran it with the Claude Code harness and ran DeepSWE with Mini-SWE-V2, so your own harness may land somewhere else. Opus 5 still wins four of the six rows.

I couldn't line V4 Flash up directly. The 0731 release reported Terminal-Bench 2.1 and a DeepSWE score with no version attached, so those numbers don't compare. The independent figure that does work across versions is the Artificial Analysis Intelligence Index. It has V4 Flash 0731 at 34 and the September V4.1 Flash at 39. AA hasn't scored 1017 yet.

Cost per task vs Haiku 5.5, GPT-6 Luna and Gemini 3.8 Flash

For agents I trust this table more than the per-token one. Artificial Analysis publishes the cost of one Intelligence Index task for each model, max reasoning effort where available:

| Model | AA Index | Cost per index task | Output tokens for the index | $/1M in / out |
|---|---|---|---|---|
| DeepSeek V4 Flash 0731 | 34 | $0.17 | 240M | $0.14 / $0.28 |
| DeepSeek V4.1 Flash (Sept) | 39 | $0.49 | 250M | $0.30 / $1.20 |
| GPT-6 Luna (Max) | 38 | $0.07 | 140M | $0.10 / $0.50 |
| Claude Haiku 5.5 (Max) | 43 | $0.21 | 440M | $0.10 / $0.50 |
| Gemini 3.8 Flash (high) | 41 | $1.24 | 170M | $0.75 / $3.75 |

Sources: AA pages for Haiku 5.5, GPT-6 Luna and Gemini 3.8 Flash.

On that data, the September V4.1 Flash scored lower than Haiku 5.5 and cost more than twice as much per task. AA flagged it as "very verbose," with 250M output tokens against a 130M median.

1017 goes straight at that weakness. On HN, apejcic reported "20% less tokens at each reasoning effort" and, in one setup, that it "produces 2.5x less tokens than old weights, without intelligence lost." A 20% cut takes $0.49 to roughly $0.39. That still doesn't catch Haiku 5.5 on AA's mix. Running off-peak or on a third-party host closes most of the gap.

Where DeepSeek wins anyway: long agent contexts

Real agent loops drag a lot of context through every call. Two pricing details flip the result once your context gets big:

Haiku 5.5 bills prompts over 100,000 tokens at $0.50 in, $2.50 out and $0.05 cache hits, five times its base rate (Anthropic pricing). GPT-6 Luna doubles its input price above 272K (OpenAI pricing). DeepSeek charges a flat rate up to 1M.

A DeepSeek cache hit costs $0.003-$0.006. Haiku and Luna charge $0.01 below their thresholds.

Here's a back-of-envelope session with list prices only: 20M cache-hit input, 1M uncached input, 400K output, every call carrying about 150K of context. For Anthropic and OpenAI I billed the uncached input as cache writes ($0.625 and $0.125 per 1M).

| | Session cost |
|---|---|
| V4.1 Flash 1017, DeepInfra fp8 via OpenRouter | $0.27 |
| V4.1 Flash 1017, Cloudflare via OpenRouter | $0.40 |
| V4.1 Flash 1017, DeepSeek off-peak | $0.45 |
| GPT-6 Luna | $0.53 |
| V4.1 Flash 1017, DeepSeek peak | $0.90 |
| Claude Haiku 5.5 (over-100K rate) | $2.63 |
| Gemini 3.8 Flash | $3.75 |

The token counts are held equal across models, which flatters the verbose ones, and Gemini's number ignores cache storage fees. Google's page also says Gemini 3.8 Flash doubles to $1.50 / $7.50 on January 1, 2027. The ordering still holds. Under 100K of context, Haiku and Luna win. Past that, Haiku falls off a price cliff, and 1017 off-peak comes in a bit under Luna, which scores 38 on AA to the September V4.1's 39.

What people on HN are reporting

The main thread has over 650 comments, and most of the praise is about speed and steerability:

pimeys says it "blasts through 10x faster" than GPT Sol/Terra/Luna, and that the tokens for two games came to "around 3 euros." Their advice: "don't use the max thinking for implementation, go with high or low."
bel8 pointed to Simon Willison's pelican SVG test, which took 11 minutes on max and 1 minute on high.
Leynos: "4.1 Flash on medium is a solid builder agent."
AmazingTurtle tested 1017 against GPT 6.1 Sol and said DeepSeek took "~4x longer in a review, found less and planned worse."
pmoriarty said V4.1 Flash "hallucinates like crazy," linked a video, and asked whether 1017 is any better.
peri-cl and eli point out that the official DeepSeek API has no zero-data-retention option and may train on your inputs. If that's a blocker, use one of the third-party hosts.

How I'd set it up today

Go through OpenRouter, sort providers by price, and use high effort instead of max:

json

{
  "model": "deepseek/deepseek-v4.1-flash-1017",
  "messages": [{ "role": "user", "content": "..." }],
  "reasoning": { "effort": "high" },
  "provider": { "sort": "price" }
}

Sorting by price pushes you onto the cheapest endpoints (DeepInfra, Relace, Morph, InferenceNet), so check two things. DeepInfra, InferenceNet and PrimeIntellect cap output at 131K, which matters if you lean on long thinking at max effort. And the cheap endpoints are mostly fp8, while Wafer serves mxfp4, so if your evals regress, pin a provider with "provider": { "order": [...] } before you blame the model.

If you call DeepSeek directly, move batch jobs and evals out of the 01:00-04:00 and 06:00-10:00 UTC weekday windows. Off-peak cuts the bill in half.

Verdict: is V4.1 Flash worth the price over V4 Flash?

For short-context agent loops, 1017 isn't the cheapest option per task. Haiku 5.5 scores higher on independent evals, and Luna costs a fraction as much per task. For long-context coding agents with big cached prefixes, 1017 off-peak or on a cheap host is the best cost-per-task deal I can find right now, and it's MIT-licensed, so you aren't locked to one vendor. Just remember it costs up to 4.3x more per output token than the V4 Flash it replaced. DeepSeek is betting you won't notice because the tasks finish sooner.

What's your default cheap model for agent loops right now, and would V4.1 Flash replace it? I'd especially like to hear from anyone who's watched Haiku 5.5's 100K cliff land on a real invoice.