Models
DeepSeek V4.1-Flash-1017 and the Search API: same price, better benchmarks, and a free web-retrieval trial
On 17 October 2026 DeepSeek swapped a faster checkpoint in behind the same model name at unchanged prices, and opened two web-retrieval APIs to everyone at $0.75 per 1,000 calls — roughly a thirteenth of what OpenAI and Anthropic charge for hosted search. Calls made through DeepSeek Harness, Claude Code and Codex are free until 31 October 2026 at 16:00 UTC. What changed, what it costs afterwards, and what to verify before you move.
- Published
- 28 Oct 2026
- Reading
- 9 min
- Class
- models
half-life 45dfrom 28 Oct 2026
DeepSeek shipped two things on 17 October 2026, and only one of them made the
headlines. The visible one was a checkpoint: V4.1-Flash-1017, now sitting behind
the unchanged deepseek-flash model name with what the company calls
a performance-and-cost Pareto improvement (API change log, 17 October
2026). The quieter one is the better deal for anyone running agents: DeepSeek's
Search and Fetch APIs opened to everyone at $0.75 per 1,000 successful
requests — and the calls your harnesses already make through them are
free until 31 October 2026 at 16:00 UTC (launch post, 17
October 2026).
If you point Claude Code or Codex at a DeepSeek key, that deadline lands three days after this post publishes. Here is what actually changed, what it costs once the free window shuts, and what you should check before moving a workflow.
What changed in the checkpoint (and what did not)
The API method and the pricing did not change. You still call
deepseek-flash; the model version behind it is now
DeepSeek-V4.1-Flash-1017, and DeepSeek's own pricing page lists that version
against the Flash tier (Models & Pricing,
read 28 October 2026). The old deepseek-v4-flash names still work —
they were retired back on 10 September 2026 and now serve the new checkpoint at
the Flash price (change log, read 28 October 2026).
The benchmark movement is substantial for a same-price update. These are the figures DeepSeek published for the 10 September release and the 17 October update, so the comparison is like for like on labels:
| Benchmark | V4.1-Flash (10 Sep 2026) | V4.1-Flash-1017 (17 Oct 2026) |
|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| ProgramBench | 20.3 | 29.1 |
| Agents' Last Exam | 31.8 | 35.4 |
| HLE (without tools) | 36.8 | 42.6 |
| NL2Repo-Bench | 65.4 | 68.2 |
| DeepSWE v1.1 | 74.2 | 77.0 |
Source: DeepSeek API change log entries for 10 September and 17 October 2026, cross-checked against the Hugging Face model card, both read 28 October 2026. Terminal-Bench was run in Claude Code and DeepSWE in the Mini-SWE-V2 framework, per the model card.
Read those numbers with two caveats attached. First, they are vendor-run: DeepSeek evaluated its own model, in harnesses of its choosing. Second, the model card's comparison table puts 1017 ahead of GPT-5.6 Sol and GLM-5.3 on Terminal-Bench 4.0 (42.3 against 39.9 and 37.9) and above even Opus 5 on DeepSWE v1.1 (77.0 against 74.0) — while Opus 5 still leads Terminal-Bench by nine points (51.8). Numbers like that deserve an independent re-run before anyone re-budgets on them. What the table does establish is direction: a same-name checkpoint moved up sharply on agentic and coding tasks, and DeepSeek says task runtime and output length both fell (change log, 17 October 2026).
Same price, lower cost — where the saving actually comes from
Token rates for deepseek-flash are unchanged since the 10 September
release: $0.15 per million input tokens and $0.60 per million output at off-peak
rates, doubling to $0.30 and $1.20 in peak hours — 01:00–04:00 and 06:00–10:00
UTC, Monday to Friday, excluding Chinese public holidays (pricing page, read
28 October 2026). The saving comes from three places the price list does not
show:
- Shorter outputs. DeepSeek says 1017 finishes tasks faster and writes less doing it, so the same task bills fewer output tokens at the same rate (change log and launch post, 17 October 2026). Output is the expensive side of the bill — at peak it costs four times the input rate.
- Cache hits. Input that hits the prompt cache bills at $0.003 per million off-peak ($0.006 at peak) against $0.15 cache-miss (pricing page, read 28 October 2026). DeepSeek's September announcement noted that cache-hit charges “often account for a large share of agent costs” and that V4.1's compressed KV cache needs a quarter of the previous generation's HBM and an eighth of its SSD storage (V4.1-Flash announcement, 9 September 2026).
- Off-peak scheduling. Off-peak rates are exactly half of peak, and most hours qualify — all weekend plus the off-peak weekday windows (pricing page, read 28 October 2026). Batch agent work outside those four peak hours and the effective rate halves again.
For calibration, the Coding Agent Index reset piece from 26 October already had Codex running DeepSeek V4.1 Flash on the index's cost frontier: 47.3 points at $0.39 of API cost per task, against $1.50 for GPT-6.1 Sol and $11.42 for Claude Opus 5.5 in the same harness. A checkpoint that does more work in fewer tokens widens that gap rather than creating it.
There is also a self-hosting floor under all of this. The 1017 weights are on Hugging Face under an MIT licence — 763 billion parameters in BF16, with 150 community quantisations listed and 116,770 downloads in the past month (card read 28 October 2026). At that size most teams will rent rather than run it, but the licence means no lock-in on the weights themselves.
The Search and Fetch APIs: what you actually get
The second release is two raw JSON endpoints that share your existing DeepSeek API key (Search & Fetch guide, read 28 October 2026):
- Search (
POST /v1/tools/search) returns ranked, structured results from DeepSeek's in-house engine — the company indexes “hundreds of billions of web pages” and says it filters SEO spam, generated content and manipulation attempts (change log, 17 October 2026). Queries support operators you would recognise from a search box:site:,-exclusions, exact phrases,lang:, andafter:/before:date filters. Each result carries a relevance score, an authority tier, a publication date and, optionally, the full page text. - Fetch (
POST /v1/tools/fetch) pulls a single URL and returns its content as Markdown — the clean hand-off format agent harnesses already prefer.
The billing rules are unusually fair for this market. You pay per successful
request only: a search that returns nothing, a fetch that times out, or a page
the site refuses to serve is not billed (guide, read 28 October 2026). The
count parameter runs to 50 results and does not change the price, so
one $0.00075 search can return fifty candidates. Rate limits are account-level —
100 queries per second on Search, 25 on Fetch, plus a per-domain cap on Fetch —
so a busy shared account, not your script, is the likely ceiling.
What it costs next to the alternatives
Hosted web search has been a $10-per-thousand line item for two years. Prices checked the day before publication:
| Provider | Web retrieval price | What else bills | Checked |
|---|---|---|---|
OpenAI web_search | $10 per 1,000 calls | Search content at the model's input rate | 3 October 2026 |
| Anthropic web search | $10 per 1,000 searches | Results as input tokens, this turn and later | 27 October 2026 |
| Google Gemini grounding | $14 per 1,000 queries after 5,000 free a month | Depends on surface | 27 October 2026 |
| DeepSeek Search or Fetch | $0.75 per 1,000 successful requests, each endpoint | Returned text as model tokens | 28 October 2026 |
Source: DeepSeek pricing page (read 28 October 2026); competitor rates as compiled by Infro (27 October 2026) and confirmed for OpenAI by Querying.ai (3 October 2026).
Thirteenth the call price is the headline, but be precise about what is being compared. DeepSeek is selling a raw API — you wire it into your own loop, or use a client that already speaks it. OpenAI and Anthropic sell a built-in tool that searches for the model mid-turn and injects the results for you. The call fee is cheaper at DeepSeek either way; the tokens the results consume still bill at your model's rate in every column of that table. And one search call is not one result: capped at 50 per request on DeepSeek's side, the per-result arithmetic gets better again, while OpenAI's per-call fee can run several times per response.
The free window, and how to use the last of it
DeepSeek is waiving Search and Fetch charges for server-side tool calls made through DeepSeek Harness, Claude Code and Codex — billing starts at 00:00 Beijing time on 1 November 2026, which is 16:00 UTC on 31 October (pricing page and launch post, both read 28 October 2026; the two statements are the same moment in different timezones).
Those three clients reach DeepSeek directly: the API serves both OpenAI-format
(https://api.deepseek.com) and Anthropic-format
(https://api.deepseek.com/anthropic) endpoints, and DeepSeek
shipped a one-click configuration script for Codex back in August (pricing page
read 28 October 2026; change log, 13 August 2026). Point a Claude Code or Codex
session at a DeepSeek key this week and its web lookups run on Search and Fetch
at no charge. That makes the remaining days useful for exactly one thing:
measuring whether the retrieval quality is good enough to keep paying for on
1 November, when 1,000 searches will cost $0.75 against $10 at the incumbents.
If you skip the trial, the paid maths still stands on its own. A conservative agent doing ten searches a task, one thousand tasks a month: $7.50 at DeepSeek, $100 at OpenAI or Anthropic's call rate — before the content tokens both of them charge on top. Set a budget alert either way; per-request billing on an autonomous loop is the kind of line item that grows quietly.
What is not settled
- Nobody outside DeepSeek has stress-tested the index yet. The spam and generated-content filtering claim is from the change log (17 October 2026). As of 28 October 2026 no independent quality comparison against Tavily, Exa or incumbent search APIs has crossed this desk. Run your own queries before you trust the ranking.
- Fetch bends to the target site, not to you. The guide is explicit that fetching respects each site's availability and access policy, returning 404 when a page blocks the request or offers no extractable content (read 28 October 2026). Paywalled and bot-protected pages will be a hole in your agent's web access.
- Reseller prices are not DeepSeek's price. The same checkpoint lists across gateways at different rates and peak/off-peak rules (OpenRouter's listing, read 28 October 2026). If you buy through a router, check its table, not this one.
- The numbers are vendor-run and the checkpoint will move again. Every figure here dates from DeepSeek's own pages in September and October 2026. DeepSeek has refreshed Flash checkpoints roughly monthly since July (change log), and the next one will quietly replace these benchmarks under the same model name. Re-read the change log before you cite them.
What to do about it
- Use the free days as a trial, not a treat. Before 31 October 16:00 UTC, run your actual agent tasks through Claude Code or Codex on a DeepSeek key and score the retrieval yourself — was the answer on page one, and did Fetch get through the sites you depend on?
- Re-price any stack with a search line in it. At $0.75 per thousand, retrieval stops being the line that discourages agents from looking things up. That changes how often you should let an agent verify rather than guess — which is the whole argument in harness engineering.
- Separate the model decision from the retrieval decision. DeepSeek's Search API works with any model, and its model works with any harness. You can mix — cheap retrieval behind a frontier model is a legitimate combination, as the price-to-performance guide sets out.
- Update your cost assumptions in writing. If DeepSeek is in your gateway routing or your own board, note the 17 October checkpoint date next to any benchmark you recorded before it. The model name did not change; the model did.
Sources
- DeepSeek API change log — entries dated 10 September and 17 October 2026, read 28 October 2026. Checkpoint swap, benchmark deltas, unchanged API method and pricing, shorter runtime and output, harness optimisation, search engine index claims, retired model-name aliases, 13 August 2026 Codex configuration note.
- DeepSeek-V4.1-Flash-1017 launch post — 17 October 2026. Pareto framing, $0.75 per 1,000 per endpoint, free server-side tool calls in DeepSeek Harness, Claude Code and Codex until 31 October 2026 16:00 UTC, open-weight links.
- DeepSeek Models & Pricing — read 28 October 2026. deepseek-flash = V4.1-Flash-1017, token rates and peak windows, cache-hit rates, 1M context and 384K output, Search and Fetch billing and free-window start (00:00 Beijing, 1 November 2026), OpenAI- and Anthropic-format base URLs.
- DeepSeek Search & Fetch guide — read 28 October 2026. Endpoints, query operators, result fields, per-success billing rules, rate limits (100/25 QPS), fetch site-policy behaviour.
- DeepSeek-V4.1-Flash-1017 model card (Hugging Face) — read 28 October 2026. MIT licence, parameter count, benchmark table including GPT-5.6 Sol, GLM-5.3 and Opus 5 comparisons, harnesses used, downloads and quantisations.
- Introducing DeepSeek-V4.1-Flash — 9 September 2026. Architecture, KV cache compression figures, the cache-hit share of agent costs.
- Server-side tool costs: search, fetch, sandboxes (Infro) — 27 October 2026. OpenAI, Anthropic and Google hosted search rates, DeepSeek's start of billing.
- OpenAI web search pricing, measured (Querying.ai) — 3 October 2026. $10 per 1,000 calls, content tokens billed at model rates, worked per-answer cost.
- DeepSeek V4.1 Flash 1017 on OpenRouter — read 28 October 2026. Reseller listing and provider spread for the checkpoint.
- The Coding Agent Index reset, AI model price-to-performance guide, Harness engineering and OpenRouter as a gateway — fullauto.online, for the method behind the cost comparisons.