Newsletter
The Frontier

Deep Dive — DeepSeek V4.1-Flash-1017: 77% on DeepSWE for 20 cents

Editorial

Editorial

— 8 min read
Share
Deep Dive — DeepSeek V4.1-Flash-1017: 77% on DeepSWE for 20 cents

DeepSeek refreshed its cheapest model on Friday and the headline number is the one buyers actually care about: 77.0 percent resolution on DeepSWE v1.1, per the model card, at just under $0.20 of compute per task on the company's cost chart. That chart puts Claude Opus 5, GPT-6 Astra and Gemini 3.8 Flash somewhere between $2 and more than $10 for the same benchmark, and none of them above the small Chinese model's score — a gap The Decoder summarized as beating larger rivals on agentic tasks while "massively cutting per-task costs" in its release writeup. Bloomberg's version of the same story was more skeptical in its framing: DeepSeek "pushes out updated AI model it says is faster, leaner." Both are accurate. The company says it, and nobody has rerun it yet.

If you are pricing an agent workload, the answer to "what does DeepSeek's new Flash model cost and what does it score" is now: $0.30 per million input tokens at peak, $1.20 per million output, 77.0 on DeepSWE, MIT-licensed weights you can download today. The rest of this piece is about why the improvement is smaller and more interesting than the benchmark jump suggests.

What actually shipped

DeepSeek-V4.1-Flash-1017 is, in DeepSeek's own words, "a minor update" to the V4.1-Flash it released in September — same architecture, same 552-billion-parameter backbone with 16 billion active per token, same one-million-token context, same MIT license. The weights went up on Hugging Face and ModelScope, and the API name deepseek-flash now serves the new build; the older deepseek-v4-flash model names are retired and their requests are rerouted to 1017 at the Flash price, per DeepSeek's API documentation.

The substance is in post-training. DeepSeek says it built the end-to-end project-development behavior with reinforcement learning using self-feedback — the model grades its own work and optimizes against its own verdicts, with no human raters and no separate reward model in the loop. The practical claims attached to that recipe are the ones that matter for a bill: tasks finish faster, outputs are shorter, and the accuracy per dollar improves on every harness DeepSeek tested. The company is explicit that the target was the cost curve, calling it a Pareto improvement rather than a capability leap.

The score table rewards that focus unevenly. Against its own predecessor the 1017 build moves DeepSWE from 74.2 to 77.0, Terminal-Bench 4.0 from 31.2 to 42.3, AutomationBench from 54.8 to 60.5, ProgramBench from 20.3 to 29.1 and Apex Bench from 64.6 to 72.1. Against the frontier it claims wins where automation is measured — DeepSWE, AutomationBench, Agents' Last Exam at 35.4 — and loses clearly where knowledge-heavy reasoning is: Humanity's Last Exam at 42.6 against Claude Opus 5's 56.3, ProgramBench at 29.1 against Opus 5's 37.0. This is a specialist getting better at its specialization, not a small model overtaking the frontier.

The cost story is a verbosity story

Here is the part that does not fit the press release. Artificial Analysis measured the September V4.1-Flash as very verbose — 250 million output tokens to complete the Intelligence Index suite against a median near 130 million, ranking 46th of 46 on the outlet's verbosity measure. A model that talks that much pays for every extra token at $1.20 per million.

So when the 1017 card says outputs are shorter and tasks cost less, the mechanism is not a cheaper chip or a new quantization. It is the model learning to stop pad­ding. DeepSeek's post-training tuned how much the model writes before it answers, and on a per-task billing model, writing half as much is indistinguishable from halving the price. That is a real engineering result — verbosity control is hard, and the predecessor's weakness was public — but it means part of the "20 cents a task" headline is the old task done with less chatter, not a fundamentally smarter model.

It also means the comparison to rivals deserves a squint. The $2-to-$10-plus per-task figures come from DeepSeek's own chart, and frontier labs price their models for a different mix of long reasoning traces and short answers. An Opus 5 run on DeepSWE burns tokens deliberating; whether you call that waste or rigor depends on whether the task ships.

Measure the harness, not just the model

DeepSeek's own card contains the caveat to its own headline. On DeepSWE, the same September model scored 74.2 inside the mini-SWE harness and 69.8 inside Claude Code — a five-point spread from scaffolding alone, on identical weights. The 1017 run used a newer mini-SWE harness version, so the jump from 74.2 to 77.0 is measured across setups that moved as well as the model. That does not make the gain fake; it makes its size uncertain until someone reruns it.

Nobody has. The Decoder noted that independent benchmarks have not appeared for the 1017 build, and Artificial Analysis has not yet scored it — the index still carries the September model at 39 points, ninth overall, at $0.27 per task. For comparison, that same index puts the closed frontier in the high fifties. Until a third party reruns DeepSWE on 1017 with a fixed harness, the honest version of the claim is: DeepSeek reports 77.0, on a harness DeepSeek chose, and the exact delta stays the company's word.

Why the compression is the real story

The technical report DeepSeek files against V4.1-Flash is titled "Pushing the Limits of KV Cache Compression," and that is the thread worth pulling. Long context is expensive because the key-value cache grows with every token; DeepSeek's answer compresses it to 890 bytes per token, roughly a quarter of what the previous V4-Flash needed, using a causal encoder-decoder that projects the decoder's cache from the encoder's final hidden states instead of each layer's own. During prefill the model activates only 8 billion parameters, 16 billion during decode — the design exists to make input-heavy agent runs, where the prompt dwarfs the answer, cheap by construction.

We covered the shadow side of that trick three weeks ago, in DeepSeek's cheap long-context trick leaves periodic blind spots: a ByteDance team showed that chunked KV compression makes retrieval accuracy swing by up to 40 percentage points depending on where a token lands inside its compression window — pad a prompt with equal signs and the answer changes. DeepSeek's mitigation in the V4.1 line was halving the compression stride. The 1017 update inherits that machinery unchanged, which is worth knowing before you trust a one-million-token context for anything where a specific clause must be found in a specific place. Cheaper per task and periodic about what it remembers are properties of the same design.

The lineage also matters commercially. DeepSeek's first V4.1 Flash arrived on September 10, the same week the company cut API prices by as much as 60 percent while preparing a domestic listing — we wrote then that V4.1 Flash ships and retires V4 Pro on Sunday. Two refreshes later the pattern holds: DeepSeek does not move its frontier upward so much as it drags its cost floor downward, which is why the open-weight field measures itself against it. Our deep dive on America's top open model still trailing China's best used DeepSeek's $0.27 per index task as the yardstick for what American open weights have to clear; the 1017 build pushes that yardstick lower.

The other half of the release

Attached to the model was a product nobody expected from a model company: DeepSeek opened its internal web search engine to any developer, as Search and Fetch APIs at $0.75 per 1,000 successful calls — the service behind DeepSeek's own app already handles hundreds of millions of queries a day, the company says, and server-side use inside DeepSeek Harness, Claude Code and Codex is free until billing starts on November 1, Beijing time, per DeepSeek's pricing page. For anyone building retrieval on top of an agent stack, a first-party search endpoint from a lab that sells the model too is a small rewiring of the vendor map — search, model and harness now come from one bill.

What to watch

First, the independent reruns: Artificial Analysis scoring 1017, and whether third parties reproduce the 77.0 on DeepSWE with the same harness version DeepSeek used. Second, November 1, when Search and Fetch metering starts and the free-through-Oct-31 sweetener on agent tool calls ends — that is the date the "cheap agent stack" arithmetic gets real. Third, the regulatory overhang The Decoder reported the same day: Beijing is reportedly preparing tighter rules on publishing open-weight models, with capability ranking among the top ten as a criterion for restriction. A model whose whole pitch is being the cheapest capable open weights on the market is exactly the artifact those rules would aim at.

And watch the competitor math. Every refresh of this kind compresses the price at which a decent coding agent runs, which is pressure on the per-seat coding subscriptions and on the mid-tier API models in between. DeepSeek is not trying to win the benchmark column. It is trying to make the frontier's price look unjustifiable one decimal at a time — 74.2 at fifty cents, 77.0 at twenty, and a harness table that shows the remaining gap is narrower than the leaderboard admits.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Deep Dive — Gemini 4 Argon: 97.5% or 49.8% on ARC-AGI-3

Deep Dive — Gemini 4 Argon: 97.5% or 49.8% on ARC-AGI-3

Google's Gemini 4 Argon has two official scores on the same benchmark this week, and they are 47 points apart. On ARC-AGI-3, the interactive test widely treated as the hardest measure of whether an agent can learn inside an unfamiliar world, Argon posted 97.5 percent — enough to make Google the second lab after OpenAI to clear 95 percent. Run the same model on the same 55 games under the other official setup, and it posted 49.8 percent.

Deep Dive — America's top open model still trails China's best

Deep Dive — America's top open model still trails China's best

Meta released the weights for Muse Spark Open on Thursday and called it "the most capable US open-weight model to date." Independent measurement agrees with the first half of that sentence and complicates the second: on Artificial Analysis's Intelligence Index the model scores 39, first among American open weights and behind GLM-5.3 at 45, Kimi K3 at 44 and Alibaba's Qwen3.8 at 40. The qualifier in Meta's headline is doing the work.

DeepSeek's cheap long-context trick leaves periodic blind spots

DeepSeek's cheap long-context trick leaves periodic blind spots

Ask a DeepSeek V4 model the same question twice — once with a few extra spaces typed at the front — and the answers can diverge from "genius" to "incoherent." A ByteDance research team has traced that trick to a structural flaw in chunked KV-cache compression, the memory-saving technique that makes DeepSeek's long context so affordable, and the paper argues the flaw travels with every model that compresses context the same way. The bug, in plain terms Long contexts are expensive because…

Deep Dive — The four-token blind spot inside DeepSeek V4

Deep Dive — The four-token blind spot inside DeepSeek V4

ByteDance's Seed research team says it has found the cause of one of the stranger recurring complaints about DeepSeek's models: the same question, asked with nothing changed except a few junk characters bolted onto the front, can flip the model from right to wrong. Their paper, posted to arXiv on September 28, traces the wobble to a memory-saving trick used during long-context inference, and reports that DeepSeek-V4-Flash-Base's retrieval accuracy swings by as much as 40.2 percentage points…

The $1.8B bet that biology's bottleneck is data, not models

The $1.8B bet that biology's bottleneck is data, not models

The biggest AI-for-science announcement in months contains no model release, no benchmark, and no demo — just a very large pile of money pointed squarely at the least glamorous part of the pipeline: the raw measurements a model would train on. What happened On October 7, Biohub — the nonprofit research institute backed by Mark Zuckerberg and Priscilla Chan — announced an expansion of its Virtual Biology Initiative alongside the U.S.

Deep Dive — Inside the sandbox factory that trains DeepSeek's agents

Deep Dive — Inside the sandbox factory that trains DeepSeek's agents

This morning's brief covered the headlines; this is the machinery underneath. On September 19, DeepSeek published a systems paper with an unglamorous title — "DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale" — and an unusually revealing body. DSec is the platform that served every sandbox workload in the company's RL training and evaluation from DeepSeek V3.2 through V4.1.