The $0.017 Agent: DeepSeek's 1017 Snapshot Cuts Task Cost Without Cutting Prices

Key takeaways

• DeepSeek upgraded the model behind the deepseek-flash API name to V4.1-Flash-1017 on October 17, 2026, with the calling method and every list price unchanged — a same-name snapshot release (DeepSeek's change log). The model under the stable alias moved; the bill's price table did not.
• Terminal-Bench 4.0 jumped from 31.2 to 42.3 (+11.1) in five weeks, and DeepSWE v1.1 reached 77.0, Automation-Bench 60.5, and Agents' Last Exam 35.4 — each above Claude Opus 5 on the Hugging Face comparison card (Hugging Face).
• Opus 5 still leads on Terminal-Bench 4.0 (51.8 vs 42.3), HLE w/o tools (56.3 vs 42.6), NL2Repo-Bench (75.3 vs 68.2), and ProgramBench (37.0 vs 29.1) — the honest split is agentic-execution tier leader, not general-frontier champion.
• A 91.5% cache-hit rate on the official endpoint puts the input price actually paid near $0.017 per 1M tokens (OpenRouter weighted average, aggregator-compiled) — at DeepSeek's rate card, a cache hit costs 2% of a cache miss (DeepSeek pricing; Local AI Zone).
• The same release opened Search & Fetch at $0.75 per 1,000 calls and disclosed that DeepSeek's search engine filters "SEO spam, GEO manipulation, and low-quality AIGC" — a platform-policy claim, published in the product's own change log (DeepSeek's change log).

On October 17, 2026, DeepSeek changed the model behind the name deepseek-flash and left every list price untouched. That combination — new weights, same prices — is quietly becoming the strongest cost lever in agent inference. On the Hugging Face benchmark card, DeepSeek-V4.1-Flash-1017 raises Terminal-Bench 4.0 from 31.2 to 42.3 (+11.1), posts 77.0 on DeepSWE and 60.5 on Automation-Bench — each above Claude Opus 5 — and finishes tasks faster with shorter outputs. DeepSeek's change log calls the result "a new performance–cost Pareto optimum across multiple harnesses." The part with no benchmark table carries the money: the official endpoint reports a 91.5% cache-hit rate, which puts the input price actually paid near $0.017 per million tokens for cache-heavy workloads. This article walks through what shipped, what the cache math does to an always-on agent's bill, and which levers your team controls.

This builds on two earlier reports. Inference Economics: Why Always-On Production Agents Are Now Affordable documented the per-token cost collapse and the 10:1 integration-to-model spend ratio. DeepSeek V4.1-Flash and the Silent Model Swap documented the September auto-routing episode — a forced V4-Pro-to-Flash substitution announced and reversed within roughly 45 hours. The 1017 snapshot is the third act, and it is different in kind: no price cut, no forced routing, just the same endpoint name quietly serving a better, cheaper-to-run checkpoint. Here we focus on the mechanics of that kind of release and the cache economics that decide whether it lowers your bill.

Same name, new model: how a snapshot lands

DeepSeek's October 17 change log is a two-paragraph event. "DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017... The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model." The legacy aliases stay routed, the website and app updated in place, and the four-digit suffix records the date. Nothing in the integration changed; the model under it did.

That is the third release convention in six weeks. September 10 shipped a named new tier with its own prices and an (announced, then reversed) plan to auto-route all deepseek-v4-pro traffic to it — the production risk was consent. The 45-hour reversal followed only user pressure: DeepSeek's own change log states V4-Pro service continues past September 14 "with the billing method remaining unchanged." October 17 is a snapshot instead: no price change, no routing change, an upgrade that lands by default under a stable alias.

Consent is not the issue here — the snapshot is an unambiguous improvement at unchanged prices, and Hugging Face shipped the MIT-licensed weights the same day, a 1M-token context at 2,500 concurrent requests. But behavior drift is still behavior drift: the checkpoint under the name changed. The response is what the auto-routing article prescribed, minus the alarm — log the served model on every response, and regression-test snapshot upgrades against your own eval set before they reach production. An upgrade you asked for and verified is an upgrade; one that arrived under a stable name while you were not looking is a change event, even when it is a Pareto improvement.

The benchmark card, read honestly

The Hugging Face card compares the 1017 build against the September release, Kimi K3, GLM-5.3, GPT-5.6 Sol, and Claude Opus 5. The pattern in the official rows splits cleanly:

Benchmark	V4.1-Flash (Sept 10)	1017	Change	Opus 5	1017 vs Opus 5
DeepSWE v1.1	74.2	77.0	+2.8	74.0	Above
Automation-Bench v1.0.6	54.8	60.5	+5.7	50.3	Above
Agents' Last Exam	31.8	35.4	+3.6	28.6	Above
Apex Bench v1.1	64.6	72.1	+7.5	65.8	Above
Terminal-Bench 4.0	31.2	42.3	+11.1	51.8	Below
HLE w/o tools	36.8	42.6	+5.8	56.3	Below
NL2Repo-Bench	65.4	68.2	+2.8	75.3	Below
ProgramBench	20.3	29.1	+8.8	37.0	Below

The 1017 model tops the card on the agentic-execution benchmarks: DeepSWE at 77.0 vs Opus 5's 74.0 (GPT-5.6 Sol 73.0, K3 67.5, GLM-5.3 66.9), Automation-Bench, Agents' Last Exam, and Apex Bench. It stays behind Opus 5 on Terminal-Bench 4.0, HLE w/o tools, NL2Repo-Bench, and ProgramBench — so the general-frontier story has not moved. Read the card, not the thread titles: community posts framed the release as "now leads open-weight models," which compresses a tiered result into a flat one. The card's footnotes matter too — Terminal-Bench 4.0 ran on the Claude Code framework, DeepSWE on Mini-SWE-V2, so harness choice is part of every number on this page.

The change that matters most for cost is not a benchmark score. DeepSeek's change log states that output length and task runtime were reduced alongside the improvements — "enabling users to complete tasks at lower cost and higher speed." Fewer output tokens per task and fewer retries per task are direct drops on the bill, at prices that did not move. The release post's DeepSWE cost-per-task chart makes the same point visually: the pass@1 score is plotted against cost per task, not against per-token price.

A 91.5% cache hit is the real price list

DeepSeek's rate card for the Flash tier did not change: cache-miss input is $0.15 per 1M tokens off-peak and $0.30 peak; cache-hit input is $0.003 off-peak and $0.006 peak; output is $0.60/$1.20. A cached input token costs 2% of an uncached one — 89.7% below the miss price in both windows. Off-peak is half of peak everywhere, and every weekend hour and Chinese public holiday counts as off-peak — a scheduling lever most teams never pull.

What changed is how often the cache hits. A compilation of OpenRouter data on the official endpoint reports a 91.5% cache-hit rate, putting the weighted input price actually paid near $0.017 per 1M tokens (Local AI Zone, aggregator-compiled — treat the figure as one endpoint's observation, not a guarantee). The arithmetic checks against the rate card: at 91.5%, effective input lands at $0.0155 off-peak and $0.031 peak (0.915 × hit + 0.085 × miss), and the aggregator's all-hours weighted figure sits inside that band. At that effective rate, a production agent streaming one billion input tokens a month — long contexts re-read every turn, which is what agent loops do — pays about $15.49 off-peak where the miss-only price would be $150.00: a 10× reduction that appears nowhere as a price cut.

Three honest caveats bound the claim. The 91.5% rate is specific to the official endpoint and its traffic mix — cache-hit rates vary by provider and by how a workload reuses context. It is workload-dependent: a workflow that rewrites its prompt prefix every step earns few hits, whatever the endpoint averages. And it is an observation, not an SLA — DeepSeek documents the prices, not the hit rate. The implication survives all three: for cache-heavy agent traffic, the effective input price is a function of how you structure context, not only of which model you pick.

That is why this release reads as a cache-economics event rather than a price cut. Claude Haiku 5.5's subagent tier collapsed per-token costs with a $0.10 price (October 7); Sonnet 5.5's release showed a cheap token can still mean an expensive task when token counts balloon 7× (the efficiency trap). DeepSeek's 1017 snapshot attacks the other half: same token prices, shorter outputs, faster completion — and a serving layer that turns agent traffic's natural context reuse into an order-of-magnitude input discount. Teams that only shop the price list will miss the second mechanism.

The one-minute explainer below puts the release on one page:

Search and Fetch: the same release prices the agent's web access

The snapshot's other half is infrastructure. DeepSeek's change log opened its in-house Search and Fetch APIs to everyone at $0.75 per 1,000 successful calls per endpoint — search results with authority and freshness scoring, fetch returning page content as Markdown. Calls through server-side tools in DeepSeek Harness, Claude Code, and Codex are free until billing starts at 16:00 UTC on October 31, 2026 (00:00 Beijing, November 1). At a stated scale of "hundreds of millions of search requests" per day and an index of "hundreds of billions of web pages," this is a real engine behind a commodity price.

Inside that changelog paragraph is a disclosure worth filing for anyone optimizing for AI-citation surfaces: DeepSeek states its search engine "filters SEO spam, GEO manipulation, and low-quality AIGC through strict detection mechanisms, ensuring authoritative and trustworthy information for users and LLMs" (change log, Oct 17). Read it as buyer-side policy, not vendor modesty: a major lab serving its own agents and third-party ones now advertises that generative-engine-optimization manipulation is a filtered defect class. If the layer that feeds model answers treats GEO gaming as spam, the durable citation play stays what it has always been on this site — primary sources, verifiable numbers, prose a generator can quote directly — not tricks aimed at the retrieval layer.

The four levers your team controls

The model side of this release is available to every team; the economics only move if the workload cooperates:

1. Design prompts for prefix stability. Cache hits are earned by reusing byte-identical context: stable instructions, schemas, and retrieved documents ahead of volatile content, and no rewriting of prefixes per step. The 91.5% rate is what a cache-shaped traffic mix can reach; a prefix-churning agent does not get there.
2. Log the served model on every response. The deepseek-flash alias now serves a different checkpoint than last week. Record the actual model version — not the requested name — so the audit trail survives snapshot drift, and keep the fallback routing layer standing (the model swap article).
3. Regression-test snapshots against your own eval set. The MIT weights on Hugging Face make every snapshot reproducible: pin the checkpoint you validated, run your tool-call and output-format tests against the new one, and promote on your calendar — not the vendor's.
4. Route on cost per task, not list price. Measure output tokens per completed task and input price after cache discounts. A low-list model with low hit rates loses to a higher-listed one whose serving layer caches your context (the cost-per-task arithmetic).

A representative build

A mid-market distributor runs a 24/7 quote-preparation agent over 200 weekly RFQs. Its workflow re-reads supplier catalogs and quoting-rules context on every RFQ — the access pattern caches are built for. The build uses an MCP module for the ERP and catalog layer, A2A delegation to supplier quoting agents, and a routing layer that pins deepseek-flash, logs the served checkpoint on every call, and runs the quote-format regression suite before each snapshot promotion. Structured context and stable prefixes push the cache-hit rate near the official endpoint's observed band, which is what makes the ~$0.017/1M effective input real rather than brochure-grade — and the shorter outputs and fewer retries from the 1017 snapshot drop the cost per completed quote again on top of it.

Request a scoped build. One-week discovery. You get a system inventory, workflow map, and fixed scope — whether or not you build with us.

Related reading

• Inference Economics: Why Always-On Production Agents Are Now Affordable — the parent article: the per-token cost collapse and the 10:1 integration-to-model spend ratio this snapshot's cache math sits inside
• DeepSeek V4.1-Flash and the Silent Model Swap — the September consent episode and reversal that set the model-identity discipline this snapshot's alias drift rewards
• The $0.10 Subagent: Claude Haiku 5.5 Collapses the Frontier-vs-Open-Weight Cost Gap — the per-token path to cost collapse, where the 1017 snapshot competes on cache share and runtime instead