DeepSeek V4.1-Flash-1017 Scores 77.0, Opens API [2026]

DeepSeek pushed out a new version of its flagship fast model on October 17, 2026, and paired the release with something that changes how developers can build with it: open access to the company’s own web search infrastructure. DeepSeek-V4.1-Flash-1017 is now live across DeepSeek’s website, mobile app, and API, and it arrives with a jump on an agentic coding benchmark that signals the company isn’t slowing down even as rivals in the United States spend billions chasing the same ground.

The headline number is a DeepSWE v1.1 score of 77.0, up from 74.2 for the prior release. That’s a real gain in a benchmark built to measure how well a model can act like a software engineer rather than just answer questions about code. DeepSeek also opened its Search API and Fetch API to every developer, priced at $0.75 per 1,000 calls, a move that puts the company in direct competition with the retrieval layer that OpenAI, Anthropic, and Google have been building around their own flagship models.

·

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

DeepSeek-V4.1-Flash-1017 Arrives With a Pareto Improvement Claim

DeepSeek announced the update in a post on its official X account and in a release note published on its API documentation site. The company described the release as having “stronger benchmark performance, faster task completion, and shorter outputs than the previous version,” a framing it called a Pareto improvement in performance and cost, meaning the new version doesn’t trade one metric for another. It gets better and cheaper to run at the same time, according to the company’s own description.

That kind of claim is common in AI model releases, but the benchmark numbers DeepSeek published give it some backing. The model is served under the existing deepseek-flash identifier, so anyone already building on DeepSeek’s Flash tier gets the upgrade automatically without changing a line of integration code. The API calling method and pricing structure stayed the same as the prior Flash version, DeepSeek said in its change log.

For teams already running production workloads on DeepSeek’s cheaper tier, that’s the most practical detail in the whole announcement. A drop-in upgrade with no migration work and no new pricing tier to negotiate is rare in a market where every major lab tends to ship new model names, new SDKs, and new pricing pages every time they improve a model. DeepSeek’s approach here looks closer to how a cloud provider patches a managed service than how a frontier lab usually rolls out a new release.

DeepSWE v1.1 Benchmark: 77.0 vs 74.2

DeepSWE is built to test autonomous software engineering: can a model read an unfamiliar codebase, understand a bug report or feature request, make the right edits, and verify its own work without a human walking it through every step. Scoring 77.0 on DeepSWE v1.1 puts DeepSeek-V4.1-Flash-1017 nearly three full points ahead of its own predecessor, a meaningful jump for a model in the “Flash” tier, which is designed to prioritize speed and cost over raw scale.

The company also reported a score of 60.5 on Automation-Bench V1.0.6, up from 54.8. Automation-Bench leans more toward multi-step task completion and tool use, the kind of benchmark that matters most for agentic workflows where a model has to chain together several actions, check its own output, and recover from mistakes along the way. Both benchmark gains point in the same direction: DeepSeek is tuning this release specifically for coding agents and automation pipelines, not general chat.

Benchmark	DeepSeek-V4.1-Flash-1017	Previous Flash Version	Change
DeepSWE v1.1 (agentic coding)	77.0	74.2	+2.8 points
Automation-Bench V1.0.6 (task automation)	60.5	54.8	+5.7 points

DeepSeek’s own framing, that the model completes tasks faster and produces shorter outputs than before, matters for anyone paying per token. A model that reaches the same or better result with fewer output tokens effectively costs less to run even before any price change, since most API billing is tied directly to token volume. That’s a quieter form of a price cut, and it’s one DeepSeek has leaned on before rather than constantly rewriting its public pricing page.

DeepSeek Opens Its Search API and Fetch API to Everyone

The bigger structural change in this release has nothing to do with the model’s weights. DeepSeek opened its Search API and Fetch API to all developers, not just a closed group of partners. Both services are powered by DeepSeek’s in-house search engine, which the company says reliably handles hundreds of millions of search requests from around the world every day. DeepSeek describes its search product as an AI-native general-purpose engine that indexes hundreds of billions of web pages, supports multilingual search, and returns results in real time.

That’s a notable claim coming from a company whose public identity has, until now, been almost entirely about model weights and inference pricing rather than search infrastructure. Building and maintaining an index at that scale is a different kind of engineering problem than training language models, and DeepSeek choosing to surface it as a public API suggests the company has been running this search stack internally for a while before deciding to open it up.

Giving every developer, not just DeepSeek’s own model customers, programmatic access to a search and page-fetch layer turns the company into a potential infrastructure vendor for retrieval-augmented generation and agentic browsing, a category that includes everything from research assistants to shopping bots to coding agents that need to look up current documentation. Developers already using tools like the kind described in our AI research assistant stack built on other providers now have a lower-cost search backend option to weigh against what they’re currently paying.

Pricing: $0.75 per 1,000 Calls, Free Window Through October 31

DeepSeek set a flat price of $0.75 per 1,000 calls for each Search or Fetch API endpoint. That’s a straightforward, usage-based number with no tiered volume discounts published yet and no separate enterprise pricing page announced alongside the release. For comparison, developers evaluating search-augmented generation pipelines are used to retrieval costs being bundled into opaque per-token charges from model providers, so a transparent flat per-call price is itself a point of differentiation.

DeepSeek is also running a temporary incentive to drive adoption. Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex are currently free for DeepSeek API users. According to DeepSeek’s own announcement, that free window closes on October 31, 2026, at 16:00 UTC, after which standard per-call billing takes effect across all access paths.

Detail	Value
Search API price	$0.75 per 1,000 calls
Fetch API price	$0.75 per 1,000 calls
Free access channels	DeepSeek Harness, Claude Code, Codex (server-side tool calls)
Free window ends	October 31, 2026, 16:00 UTC
Model identifier for V4.1-Flash-1017	deepseek-flash
Weight license	MIT, hosted on Hugging Face and ModelScope

The choice to make the free period apply specifically inside Claude Code and Codex is worth sitting with for a moment. Those are Anthropic’s and OpenAI’s respective coding agent products, not DeepSeek’s own tooling. DeepSeek is effectively subsidizing developers who use its search layer from inside a competitor’s coding environment, a tactic aimed at getting its retrieval infrastructure embedded into workflows that already run on rival models, rather than trying to pull those developers onto DeepSeek’s own chat model first.

How Developers Can Access deepseek-flash Today

Because the upgrade ships under the existing deepseek-flash model name, any application already pointed at that identifier is running the new version without any code change. DeepSeek’s open source tooling repository, DeepSeek Harness on GitHub, also published release notes confirming the new search and fetch tool-calling support the same day. A typical request body for a chat completion style call looks like the structure below, consistent with the OpenAI-compatible format DeepSeek’s API has used for its prior releases.

{
  "model": "deepseek-flash",
  "messages": [
    {"role": "user", "content": "Refactor this function and explain the change."}
  ],
  "tools": [
    {"type": "search"},
    {"type": "fetch"}
  ]
}

Enabling the search and fetch tools inside that same call is what lets an agent look something up mid-task rather than relying only on what it learned during training. That pattern is the backbone of most modern coding assistants, and it’s the same category of capability covered in our Cursor to Claude Code migration guide, where tool-calling access to live information is treated as a baseline requirement rather than a bonus feature.

DeepSeek’s Open Source Strategy and the MIT License

DeepSeek continues to release its weights under the MIT license, publishing them on both Hugging Face and ModelScope. That licensing choice has been the core of DeepSeek’s market position since its earlier releases rattled valuations across the AI sector, and it remains the sharpest contrast with OpenAI, Anthropic, and Google, all of whom keep their flagship weights closed.

An MIT license means any company, researcher, or hobbyist can download the weights, modify them, fine-tune them, and deploy them commercially without paying DeepSeek a licensing fee or asking permission. That’s a meaningfully more permissive stance than licenses like Llama’s, which carry usage restrictions above certain scale thresholds. It also means DeepSeek’s actual revenue model leans more on API convenience, the Search and Fetch services, and hosted inference than on controlling who can run the model itself.

That strategy keeps drawing scrutiny. Executives at rival labs have argued publicly that open-weight releases carry more downstream risk than closed ones, since nobody can control how a freely redistributable model gets fine-tuned or deployed once it’s out in the world. Our coverage of an Anthropic executive warning that open AI is riskier lays out that argument in detail, and it’s a direct rebuttal to the approach DeepSeek just doubled down on with this release.

Competitive Field: DeepSeek vs OpenAI, Anthropic, and Google

DeepSeek’s move lands in the middle of an unusually aggressive pricing cycle across the whole AI industry. Anthropic cut prices sharply with its Claude Haiku 5.5 release, a move we detailed in our report on the 90 percent price cut that sparked an AI pricing war. OpenAI, Anthropic, and Google have all been racing to ship faster, cheaper tiers of their flagship models to capture developers who don’t need frontier-level reasoning for every single call, a dynamic our side-by-side on GPT-6.1 Sol, Sonnet 5.5, and Gemini 3.8 Flash pricing breaks down in detail.

What DeepSeek is doing differently is bundling a search and retrieval layer directly into that price competition, rather than competing purely on model weights or token price. None of DeepSeek’s big three American rivals currently sell a standalone, openly priced search API at anywhere close to $0.75 per 1,000 calls to developers outside their own walled ecosystems. If that pricing holds and the quality of DeepSeek’s search index is competitive, it gives independent developers a reason to route retrieval traffic through DeepSeek even while running a different company’s model for reasoning and generation.

That unbundling is also visible elsewhere in the market. Meta’s push into coding tools, which we covered when Meta’s Muse Code took on four AI coding rivals, reflects the same broader pattern: big AI players are increasingly competing on specific layers of the stack, search, coding agents, fine-tuning, rather than purely on which company has the single smartest general-purpose chatbot.

Market Impact and Why Wall Street Is Watching

DeepSeek’s past releases have shown a track record of moving markets well beyond the AI sector narrowly defined, and every new release now gets measured against that history. The Straits Times reported that DeepSeek unveiled the update through a series of posts on X, framing it as faster and leaner than its predecessor. A low-cost, high-performing model landing the same week Western labs are fighting over fractions of a cent per token tends to put pressure on the pricing power of every company selling inference as a service, regardless of whether DeepSeek’s specific numbers end up dominant in independent testing.

Yahoo Finance noted that DeepSeek confirmed the release and the new API access through its own official account, making the company’s announcement, not a leak or a third-party report, the primary source for every detail now circulating about the update. A capable model shipped cheaply alongside a public search API the same week reinforces the picture of a competitive AI stack built outside the labs that usually dominate Western headlines.

For enterprise buyers evaluating vendors, the practical takeaway is less about any single benchmark and more about optionality. Every open, MIT-licensed release that keeps pace on agentic coding benchmarks gives procurement teams another lever to negotiate price with the closed-weight incumbents, even in organizations that have no intention of actually deploying DeepSeek’s weights themselves.

Historical Context: From Early Shockwaves to a Steady Release Cadence

DeepSeek built its reputation on releases that disrupted assumptions about what it costs to train and serve a capable model. The V4.1-Flash-1017 update is a smaller, more incremental story by comparison, a dated point release rather than a new model family, but the pattern is consistent with how the company has operated: ship open weights, keep the API pricing stable or falling, and layer in new infrastructure, in this case search, around the core model rather than only chasing a bigger parameter count.

That steadier, infrastructure-first cadence also distinguishes this release from the kind of big-bang model launches that dominate headlines elsewhere in the industry. Where other companies have framed their releases as reaching frontier status with months of buildup, DeepSeek’s October update reads more like a maintenance release with an unusually consequential side effect: a new public API surface that didn’t exist the day before.

Why DeepSeek’s Search API Matters for AI Agents

Agentic coding tools and research assistants live or die on how well they can pull in fresh, accurate information mid-task. A model with strong reasoning but no reliable way to check current documentation, library versions, or breaking changes will still hallucinate or go stale. That’s precisely the failure mode a built-in Search and Fetch layer is meant to close, and it’s why DeepSeek chose to open these APIs at the same time as a benchmark-focused model update rather than as a separate announcement.

Independent studies have already flagged how much friction still exists in real-world AI-assisted development. A widely cited METR study found that developers using AI coding agents were measurably slower on certain tasks, a reminder that benchmark gains on DeepSWE or Automation-Bench don’t automatically translate into faster shipping in production codebases. A better search backend could narrow that gap by cutting the number of times an agent needs a human to correct an outdated assumption, but that’s a claim that will need independent verification over the coming months rather than something DeepSeek’s announcement proves on its own.

Risks and Open Questions

Several details remain genuinely unresolved. DeepSeek’s announcement did not name an individual executive or researcher behind the release, so there is no on-record personal quote to attribute beyond the company’s own official statements. Independent benchmark verification from third parties outside DeepSeek’s own disclosures had not yet circulated widely at the time of this report, which is standard for a release that’s barely a day old, but it means the 77.0 and 60.5 figures should be read as DeepSeek’s own reported numbers until outside labs replicate them.

There’s also an open question about what happens to adoption once the free Search and Fetch window inside DeepSeek Harness, Claude Code, and Codex ends on October 31. Free pilot periods are a common way to seed usage before a paid tier kicks in, but the real test of whether $0.75 per 1,000 calls sticks as a sustainable price point will only become clear once developers are paying for it directly.

Predictions: What Comes Next for DeepSeek

• Expect competing labs to respond with their own retrieval or search-as-a-service offerings within the next one to two quarters, rather than ceding that layer of the stack to DeepSeek by default.
• Watch for third-party benchmark labs to publish independent DeepSWE and Automation-Bench replications within weeks, which will either confirm or complicate DeepSeek’s self-reported 77.0 and 60.5 scores.
• The free access window through DeepSeek Harness, Claude Code, and Codex closing on October 31 will be the first real signal of price sensitivity once standard per-call billing applies.
• Expect continued scrutiny from rival labs over the safety implications of MIT-licensed open weights, especially as DeepSeek pairs open models with more infrastructure, like search, that makes the stack easier to deploy at scale.
• If the Search API gains real developer traction, look for DeepSeek to introduce volume-based pricing tiers beyond the current flat $0.75 per 1,000 calls, following the pattern most infrastructure providers eventually adopt once usage scales.

What This Means for Developers Right Now

For teams currently paying for search or retrieval services bundled into another provider’s agent framework, DeepSeek’s flat per-call pricing is worth benchmarking against existing costs before the free window closes. For teams evaluating which model to put behind a coding agent, the DeepSWE and Automation-Bench gains are real but incremental, not a reason on their own to rip out an existing production stack. And for anyone building retrieval-augmented tools, a new, openly priced search index from a company with a track record of undercutting incumbent pricing is a data point worth tracking even for developers who have no plans to touch DeepSeek’s language models directly.

Frequently Asked Questions

What is DeepSeek-V4.1-Flash-1017?

It’s an updated version of DeepSeek’s Flash model, released October 17, 2026, and live on DeepSeek’s website, app, and API under the existing deepseek-flash identifier.

What benchmark scores did DeepSeek report for the new model?

DeepSeek reported a DeepSWE v1.1 score of 77.0, up from 74.2 for the previous version, and an Automation-Bench V1.0.6 score of 60.5, up from 54.8.

How much does DeepSeek’s Search API cost?

Each Search or Fetch API endpoint costs $0.75 per 1,000 calls, according to DeepSeek’s official announcement.

Is the Search API free right now?

Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex are currently free for DeepSeek API users, but that free access ends October 31, 2026, at 16:00 UTC.

Are the model weights open source?

Yes. DeepSeek released the weights under the MIT license on Hugging Face and ModelScope, the same open approach it has used for prior releases.

Do I need to change my API integration to get the new model?

No. The update ships under the same deepseek-flash model identifier with unchanged API calling methods and pricing, so existing integrations get the upgrade automatically.

What powers DeepSeek’s search results?

DeepSeek says the Search and Fetch APIs run on its in-house search engine, which it describes as indexing hundreds of billions of web pages and handling hundreds of millions of requests daily.

How does this compare to search tools from OpenAI, Anthropic, or Google?

None of DeepSeek’s major Western rivals currently offer a standalone, openly priced search API at a comparable flat rate to developers outside their own ecosystems, which is what makes DeepSeek’s move notable beyond the model update itself.

Related Coverage

• Microsoft Decision-1: New AI Model Scores Choices 35x Faster [2026]
• Anthropic Pitches Congress a $1B AI Safety Bill [2026]
• Veo 3.1 vs Kling 3.0 vs Runway: 1,397 Elo Score Wins [2026]
• Gemini 4 Argon 140 FPS Alan Wake 2 Claim Unverified [2026]
• Apple, Meta Pair Up: Muse AI Needs $21,598 in Macs [2026]