A checkpoint, priced like a rounding error

DeepSeek did not launch a new model on October 17. It replaced one in place. V4.1-Flash-1017 is a new post-training checkpoint of the Flash model the company shipped in September, and it arrives under the same API name, deepseek-flash, at the same prices. Anyone already calling that endpoint woke up with a different model. DeepSeek's own docs now list the served version as 1017, and the retired legacy names route to it.

The company's description is "Pareto improvement": better scores, faster completion, shorter outputs. The post-training change it names is reinforcement learning with self-feedback, aimed at building complete projects end to end, the games and 3D scenes that coding agents are now demoed on. Shorter outputs are not a cosmetic detail at this price point. Output tokens are the expensive half of an agent run, and a model that writes less while scoring more is cheaper twice.

Read the whole table

The number travelling around social media is 77.0 on DeepSWE v1.1, above Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. It is a real number, from DeepSeek's own table — the DeepSWE column is not the problem. The problem is treating it as a summary. On HLE without tools, 1017 scores 42.6 against Opus 5's 56.3. On Terminal-Bench 4.0 it scores 42.3 against 51.8. On NL2Repo-Bench, Opus leads 75.3 to 68.2. In DeepSeek's reported DeepSWE evaluation, 1017 leads the named rivals; the rest of the company's table is evidence of a narrower benchmark advantage, not a new overall model ranking. HLE and Terminal-Bench are individual evaluations, not complete measures of reasoning ability or terminal work — but they are the same table's own counterweight, and "roughly on par with Opus" is a selection, not a summary.

There is a second footnote, and DeepSeek published it itself. Its model card for the previous checkpoint scores the same model under different agent harnesses: 74.2 on DeepSWE in the mini-SWE harness, 69.8 in Claude Code, 65.6 in Codex. That 8.6-point spread is substantially larger than the 2.8-point improvement 1017 claims over its predecessor — although the published figures do not isolate how much of the spread comes from the harness rather than other evaluation settings. Cross-model comparisons deserve particular scrutiny whenever harnesses and evaluation settings differ. Independent evaluations have not landed yet.

The part that changes a budget

The cost chart is potentially the more commercially important claim in this release — and it is still DeepSeek's measurement. The company reports a top-scoring DeepSWE run at just under $0.20 per task, the previous checkpoint at about $0.50, and rival frontier models in the dollars-per-task range. At the low reasoning setting, 1017 still clears about 72 on DeepSWE for about $0.10. Those figures are arithmetic on the company's own token counts and prices; they need independent reproduction under comparable settings before they can serve as reliable operating-cost estimates. One caution on the token prices quoted elsewhere: $0.15 per million input tokens is the off-peak rate. Peak pricing — 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays — is $0.30 in and $1.20 out, per DeepSeek's own pricing page. Even at peak, the per-task math stays in cents, on DeepSeek's numbers.

The second release matters to the same buyers. DeepSeek's Search and Fetch APIs, until now internal tools behind its own products, are public at $0.75 per 1,000 successful requests each, each capped at 5 QPS per account and billed only when a request succeeds. Calls made server-side through DeepSeek Harness, Claude Code or Codex are currently free, with per-request billing for those starting at 00:00 Beijing time on November 1, 2026 — and the billing terms differ by integration, so check the terms for the route you actually use rather than assuming direct API calls are free. An agent stack can now take its model and its web layer from one vendor. That bundling, more than any single benchmark, is what competitors will feel.

What is not settled

The open-weights question, at least, is settled — and it is worth saying how, because the early coverage disagreed. The Decoder reported the 1017 weights published to Hugging Face and ModelScope. MarkTechPost reported them as not disclosed, pointing at the September repository, whose model card still carries the previous checkpoint's benchmark table. The disagreement traces to two different pages. DeepSeek published a separate repository under the new name — deepseek-ai/DeepSeek-V4.1-Flash-1017 on Hugging Face — and it holds the full checkpoint: 48 safetensors shards, FP8, under the same MIT licence, with its own card carrying the 1017 table. So self-hosters do get the new checkpoint, not the September one. What they do not get is confirmation: every score in that table is still DeepSeek's own run, and no independent evaluation has reproduced them yet.

Sources

  1. [1] DeepSeek API Docs, Models & Pricing — source for the served model version DeepSeek-V4.1-Flash-1017, the retired legacy names routing to it, peak/off-peak and cache pricing, the peak-hour windows and the Chinese-public-holiday exclusionRead source
  2. [2] DeepSeek API Docs, Search & Fetch guide — source for the endpoints, billing per successful request only, the separate 5 QPS account limits, and the free server-side tool-call window ending at 00:00 Beijing time on November 1, 2026Read source
  3. [3] DeepSeek-V4.1-Flash-1017 repository, Hugging Face (deepseek-ai) — the separately named 1017 repository: downloadable FP8 weights in 48 safetensors shards under the MIT licence, and the company's 1017 benchmark table; file list inspected October 17, 2026Read source
  4. [4] DeepSeek-V4.1-Flash model card, Hugging Face — the September checkpoint's card, source for the harness-by-harness scores of the previous checkpoint (mini-SWE, Claude Code, Codex)Read source
  5. [5] MarkTechPost, "DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API" (October 17, 2026) — first reporting of the 1017 announcement table, and the weights caveat that the separate 1017 repository supersedesRead source
  6. [6] The Decoder, "Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs" (October 17, 2026) — the per-task cost chart and the report of weights on Hugging Face and ModelScopeRead source