
DeepSeek V4.1 Flash 1017 Jumps 19 Points on FrontierMath Tier 4 — and DeepSeek Never Mentioned It
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 125 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 134 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 127 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 186 tok/s
The most interesting benchmark number attached to DeepSeek V4.1 Flash 1017 is one the vendor did not publish. On 21 October 2026, five days after the checkpoint shipped, the tracker BenchLeader added a FrontierMath Tier 4 result for the model: 46.3% at maximum reasoning effort. That is one problem's worth of daylight away from Qwen3.8-Max, which measured 46.3% on 2 August, and level with Claude Haiku 5.5, which measured 46.3% on 7 October. It is also roughly nineteen points above where every earlier DeepSeek checkpoint sits on the same board. DeepSeek's own changelog for the 17 October update lists nine benchmarks. Not one of them is a math benchmark. It is served on OrcaRouter alongside the September build under deepseek/deepseek-v4.1-flash, so you can run the same prompt set against both checkpoints.
That gap between what DeepSeek claimed and what a third party measured is the story here. The 1017 checkpoint was announced as a cost and speed release — the phrase in the changelog is a "performance-cost Pareto improvement," with shorter outputs and faster task completion. The benchmark table in that same changelog is all agentic work: HLE without tools, Terminal-Bench 4.0, DeepSWE v1.1, ProgramBench, Apex Bench, Agents' Last Exam. Math is not on the list. So either DeepSeek's math improved as an unadvertised side effect of the agentic work, or one tracking site has a row that the others do not. Both readings are worth taking seriously, and they lead to different advice.
What Tier 4 v2 actually asks, and why the problem count matters more than the percentage
FrontierMath Tier 4 v2 is the research-grade tier of Epoch AI's benchmark: a private expansion set of 43 research-level problems, released 12 June 2026, with models allowed to run Python in a stateless sandbox against a fixed library list. The scored denominator appears to be 41 — the published scores divide cleanly by it, since 19 correct out of 41 is 46.34% and 11 out of 41 is 26.83%. Either way, the set is small enough that one problem is worth roughly 2.4 percentage points, and that is the detail that governs everything else on this page. Two scores that look far apart on a leaderboard can be eight questions apart.
Which is exactly what the DeepSeek numbers are. A 46.3% score works out to nineteen correct against that denominator. The 26.8% score the earlier checkpoints carry works out to eleven. The entire "nineteen-point jump" is eight problems. It is a real difference — eight problems out of forty-one is not noise in the ordinary sense — but it is also a sample small enough that a single changed tool, a changed system prompt, or a different effort setting can move a model several places. When four models sit tied at 46.3%, they are not meaningfully separated. They are all within one problem of each other.
The trajectory: three flat checkpoints, then a step
The reason the 46.3 registers as news rather than a rounding artefact is what came before it. Across DeepSeek's last four FrontierMath Tier 4 v2 entries, the line is flat and then vertical:
• DeepSeek V4 Flash 0731 — 24.4%, the July checkpoint
• DeepSeek V4 Pro 0813 — 26.8%, the August flagship
• DeepSeek V4.1 Flash (0910) — 26.8%, the September release that replaced both
• DeepSeek V4.1 Flash 1017 — 46.3%, the October checkpoint, at max effort
The easier tiers moved too, but modestly: BenchLeader lists 71.2% on FrontierMath Tiers 1–3 for the 1017 build against 67.4% for the September one. That is the shape you would expect from a genuine capability gain — a few points where the questions are tractable, a much larger move where they are not. It is also the shape you would expect from a harness change that helps disproportionately on the hardest problems, which is why the next section matters.
Which board you believe depends on which board you read
We checked five trackers. Two of them now carry 46.3% for the 1017 build. Three do not.
• BenchLeader — 46.3% on Tier 4 and 71.2% on Tiers 1–3, at max effort, result added 21 October 2026
• DataLearner, now serving as LLMLearner — 46.34%, tied 21st, evaluated in a "Max (With Tools)" configuration, last result added 26 October 2026, and flagged on the board as the best open-weights score
• benchlm.ai — 26.829%, ranked 30th, "Python-enabled iterative mathematical problem solving," page updated 25 October 2026
• tensor.news — no 1017 row at all; its DeepSeek V4.1 Flash entry reads 26.83%, measured 9 September 2026
• trustthebench — 26%, on a page dated 26 August 2026, which predates the checkpoint entirely

The three dissenters share a provenance problem. The 26.83 figure is, to two decimals, the score the September build earned. A tracker that carries a new model label onto an old measurement produces exactly this pattern, and the August-dated page is unambiguous evidence that at least one of them is doing it. That does not make the 46.3 authoritative — it makes the 26.8 stale rather than contradictory.
What would settle it is a second harness. Neither the 46.3 nor the 71.2 has been reproduced outside the tracking layer, and DeepSeek has published no math evaluation of its own for this checkpoint. Treat the number as a strong third-party signal, not as an established fact about the model. If you need a defensible figure for a procurement document, you do not have one yet.
Why DeepSeek would ship a math gain and not mention it
The changelog's table explains the omission better than any conspiracy would. The 1017 update is presented as a coding and agentic release, and the vendor numbers are entirely in that register: HLE without tools up from 36.8 to 42.6, Terminal-Bench 4.0 from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0, ProgramBench from 20.3 to 29.1, Apex Bench from 64.6 to 72.1. A model tuned on long agentic trajectories — where the reward is finishing a multi-step task rather than emitting a correct final token — will pick up mathematical stamina as a by-product. Hard math and long agent runs reward the same behaviour: keep checking, do not stop early.

Independent testers found the opposite of what that theory predicts, which is what makes the whole thing interesting. On a hard-math sample, the evaluator sference measured minimum-effort reasoning output falling from 1,974 tokens to 790, with AIME median output down from 1,681 to 1,124. Shorter reasoning at minimum effort is not what you would expect from a model that just gained nineteen points on research mathematics. The likeliest reconciliation is effort-dependent: the 46.3 was recorded at max effort, and what changed is how much thinking the model does when you let it think, not how efficiently it thinks when you do not.
Where 46.3 actually sits
Joint-best among Chinese models is not the same as near the frontier, and the same board makes that clear within a couple of rows. The Tier 4 v2 leaders are a different class of model entirely: GPT-6.1 Sol at 100%, GPT-6 Astra at 97.6%, Claude Opus 5.5 at 95%, Claude Fable 5 at 90.2%, GPT-6 Sol at 90%. The nearest model above the 46.3 cluster is GPT-5.4 at 49%. LLMLearner's leaderboard labels 46.3 as the best open-weights score on the board. A half-century score on the hardest tier is an excellent result for an open-weights model and a poor one for a frontier model, and DeepSeek V4.1 Flash 1017 is firmly in the first category.
• Tier 4 v2 score — DeepSeek V4.1 Flash 1017 46.3% vs Qwen3.8-Max 46.3% vs Claude Haiku 5.5 46.3% (tied, 21st)
• FrontierMath Tiers 1–3 — DeepSeek V4.1 Flash 1017 71.2% vs its own September build at 67.4%
• Weights — MIT-licensed, downloadable, 552B-parameter mixture-of-experts with a separate conditional-memory table; trackers put the full on-disk footprint near 763B parameters
• Price — $0.30 per million input and $1.20 per million output at peak, $0.15 and $0.60 off-peak, unchanged from September
• Context — 1M tokens, with image input
• Vendor math claims — none; FrontierMath appears nowhere in DeepSeek's published tables

The practical read: if your work is competition-style mathematics and your alternative was an open-weights model, 46.3 changes the calculus, because it puts a 552B MIT-licensed model in the same band as two proprietary models that cost more per token. If your alternative was a frontier closed model, nothing here closes the gap — the 95-to-46 spread on Tier 4 is not a rounding error.
Testing the claim on your own problems, without betting a production path on it
Given that the 46.3 rests on 41 private problems and one tracking layer, the sensible move is to measure it against your own workload rather than trust either number. DeepSeek V4.1 Flash 1017 is served on OrcaRouter as deepseek/deepseek-v4.1-flash-1017, alongside the September build under deepseek/deepseek-v4.1-flash — which means you can run the same prompt set against both checkpoints through one API key and see whether the eight-problem gap shows up in your own evals. There is no second contract and no code change; the model string is the only difference.
Two things about the routing layer are relevant to a decision this unsettled. Pricing is passed through at provider list price with 0% markup, so DeepSeek's off-peak rate of $0.15 and $0.60 is live here the same day it is live anywhere, and a vendor price move does not wait for us to sync a catalogue. More useful while a benchmark is disputed: automatic failover and the routing DSL let you send the hard-math slice to the 1017 build at max effort and fall back to a stronger model when it returns an answer you cannot verify — so an unproven nineteen-point gain costs you a routing rule rather than a migration.
What would change this story
Three things are worth watching, and all three are checkable. The first is whether benchlm.ai and tensor.news update their 1017 rows to something other than 26.8; if two more trackers converge on the October measurement, the stale-row reading is confirmed and the number becomes ordinary. The second is whether DeepSeek adds a math row to its own comparison table in a future changelog — a vendor that believed it had gained nineteen points on FrontierMath would probably say so, and silence there is mild evidence against. The third is a no-tools reproduction. The 46.3 was recorded with tools allowed, and a stateless run would tell you whether the gain is mathematical reasoning or better use of a Python sandbox.
Until one of those lands, the honest summary is narrow. A nine-day-old checkpoint, an MIT licence and a $0.30 input price have landed in the same band as Qwen3.8-Max and Claude Haiku 5.5 on the hardest public math tier — on two trackers, at max effort, out of forty-one problems, unclaimed by the lab that built it. That is enough to justify an afternoon of your own evals. It is not enough to justify rewriting a benchmark slide.
Compared in this article2
Detected from this article · Benchmarks: Artificial Analysis · updated daily
