DeepSeek's V4.1-Flash-1017 Hits 77.0 on DeepSWE, Beats Opus 5 in Its Table

By Sophie Lindqvist

DeepSeek has released DeepSeek-V4.1-Flash-1017, and its own benchmark table puts the new checkpoint at 77.0 on DeepSWE v1.1 — ahead of Claude Opus 5 at 74.0. The model went live today on the DeepSeek website, app and API under the deepseek-flash name. Alongside it, DeepSeek opened its Search and Fetch APIs to every developer, giving agent builders a cheap coding model plus a first-party web layer behind one API key.

The stakes here are straightforward. Agent coding is where the serious API money sits right now, and DeepSeek is bundling two things rivals sell separately: a model and a web search layer. At $0.30 per 1M input tokens and $1.20 per 1M output tokens at peak, it undercuts the field while claiming benchmark wins in exactly the agentic categories buyers watch.

What's in the 1017 release?

DeepSeek calls the update a Pareto improvement over the September V4.1-Flash release — stronger scores, faster task completion and shorter outputs. It beats the previous checkpoint on all 11 benchmarks DeepSeek reported, by 2.0 to 11.1 points.

The headline numbers:

• 77.0 on DeepSWE v1.1, up from 74.2, run in Mini-SWE-V2.
• 60.5 on Automation-Bench v1.0.6, the top score in DeepSeek's table.
• 35.4 on Agents' Last Exam, also a table lead.
• 42.6 on HLE without tools — the weak spot, versus 56.3 for Claude Opus 5.

DeepSeek says the checkpoint uses reinforcement learning with self-feedback, tuned for end-to-end project work. Its examples run from game creation to 3D modeling. The model is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later.

In plain terms: this is a post-training refresh aimed squarely at agentic coding harnesses, not a new architecture.

How does the architecture work?

The design comes from the V4.1-Flash technical report on arXiv (2609.19969). It's a multimodal Mixture-of-Experts model with 40 Transformer layers, split into a 20-layer causal encoder and a 20-layer decoder.

The efficiency tricks stack up:

• Causal Encoder-Decoder (CED): decoder global KV is projected from the final encoder hidden states, which the report says nearly halves prefill compute.
• Compressed Sparse Attention 2 (CSA2): layers run in Full, Reindex or Reuse mode, sharing main KV and indexer keys across layers.
• FP4 main KV cache: global KV drops to 890 bytes per token, roughly a quarter of V4-Flash.
• SWA Bounded Replay: persistent KV cache shrinks to about one-eighth of V4-Flash.
• Reasoning effort: a 1-to-100 scalar, with API tiers low, high and max mapping to 50, 75 and 100.

Total size is 552B backbone parameters plus 196B Engram parameters, activating 8B per token in prefill and 16B in decode. Context is 1M tokens with up to 384K output tokens. The mechanism matters because it explains the price: tiny active parameter counts plus aggressive KV compression are what make $0.30 per 1M input tokens economically possible.

How does it score against the field?

All figures are DeepSeek's own. Terminal-Bench 4.0 was run in Claude Code. ZeroBench-main uses Pass@5.

Benchmark	1017	Kimi K3	GLM 5.3	GPT 5.6-Sol	Claude Opus 5
DeepSWE v1.1	77.0	67.5	66.9	73.0	74.0
Terminal-Bench 4.0	42.3	12.6	37.9	39.9	51.8
Automation-Bench v1.0.6	60.5	46.7	48.8	45.8	50.3
Agents' Last Exam	35.4	27.6	28.5	26.7	28.6
Apex Bench v1.1	72.1	47.8	56.6	54.3	65.8
ProgramBench	29.1	17.5	19.0	23.0	37.0
NL2Repo-Bench	68.2	58.0	58.0	56.8	75.3
HLE (no tools)	42.6	43.5	42.0	44.5	56.3
BabyVision (w/ tools)	92.5	85.7	n/a	88.9	94.1

The split is clean. The 1017 checkpoint leads on DeepSWE, Automation-Bench, Agents' Last Exam and Apex Bench — the sustained, real-work agent suites. Claude Opus 5 still leads on Terminal-Bench 4.0, ProgramBench, NL2Repo and HLE. So the picture is a specialist agentic model with a genuine reasoning gap, at least on DeepSeek's own scorecard.

Against open-weight rivals, the efficiency gap is the story. Kimi K3 carries 2.8T total parameters with 104B active; GLM 5.3 sits at 753B. DeepSeek's 8B prefill / 16B decode activation is a different order of magnitude, and DeepSeek is the only one of the three that publishes API pricing.

One caveat for open-source buyers: the September V4.1-Flash weights (763B parameters, FP8) are MIT-licensed on Hugging Face, but the 1017 weights and their hardware requirements are not disclosed. If the checkpoint stays closed, the open-weight comparison only applies to the older release.

What do the Search and Fetch APIs offer?

The Search API returns ranked web results for a query. The Fetch API returns a single page as Markdown. DeepSeek says its in-house engine already serves hundreds of millions of requests daily.

The details:

• Endpoints: POST /v1/tools/search and POST /v1/tools/fetch on api.deepseek.com.
• Search options: 1 to 50 results, optional page content, and operators like site:, lang:, after: and before:.
• Result fields: each result carries a 0-to-3 rerank score and a 0-to-4 authority tier.
• Price: $0.75 per 1,000 successful requests per endpoint; empty results and failures are free.
• Rate limit: 5 QPS per account for each API.
• Free window: server-side calls via DeepSeek Harness, Claude Code or Codex stay free until November 1, 2026.

That free window through November 1, 2026 is a customer-acquisition play aimed directly at Claude Code and Codex users. DeepSeek is telling agent developers they can wire search into existing harnesses at zero cost for months, then pay a metered rate after.

What does it cost to run?

You call the model with deepseek-flash through OpenAI- or Anthropic-format endpoints. Peak pricing is $0.30 per 1M cache-miss input tokens and $1.20 per 1M output tokens. Off-peak rates are half that.

Stack that against the architecture and the positioning reads clearly: DeepSeek wants to be the cheap default for agentic coding workloads, with search bundled in-house rather than bolted on from a third party.

The open question is whether the 1017 weights get released under MIT like their predecessor — and whether independent runs confirm DeepSeek's numbers on the suites where it claims the lead.